You are on page 1of 279

Hansen Solubility Parameters in Practice

Complete with eBook, software and data


5th Edition
Steven Abbott, Charles M. Hansen and YAMAMOTO Hiroshi

Published by Hansen-Solubility.com
Book and Software Copyright © 2008-2020 Steven Abbott, Charles Hansen & YAMAMOTO Hiroshi

ISBN 978-0-9551220-2-6

1
The HSPiP team

Dr YAMAMOTO Hiroshi (right) officially joins the HSPiP development team (in his spare time!) as Dr
Charles Hansen (left) handed over a signed copy of the Hansen Handbook. Professor Steven Abbott is holding
the celebratory bottle of champagne.

2
Contents
Introduction & Guarantee by Steven Abbott .......................................................................................................... 5
Chapter 1 The Minimum Possible Theory (Simple Introduction) .................................................................. 11
Chapter 2 The Sphere (The Preferred Method of Visualizing) ....................................................................... 16
Chapter 3 Your first HSP Sphere (Determining the HSP Sphere) .................................................................. 19
Chapter 4 The Grid (A different route to the Sphere) ..................................................................................... 26
Chapter 5 FAQ – The Radius and other questions .......................................................................................... 29
Chapter 6 Coming clean (Finding Good Solvents) ......................................................................................... 35
Chapter 7 Safer, Faster, Cheaper (Optimizing Solvent Formulations) ........................................................... 40
Chapter 8 Coming together (Polymer Compatibility) ..................................................................................... 42
Chapter 9 Sticking, Flowing and Dissolving (HSP and Adhesion, Viscosity and Dissolving) ...................... 45
Chapter 10 Polymer Miscibility ........................................................................................................................ 61
Chapter 11 Shades of Black (Designed Partial Compatibility - Bitumen)........................................................ 68
Chapter 12 Insoluble solubility parameters (HSP for Pigment Surfaces) ......................................................... 72
Chapter 13 Cracks in the system (Environmental Stress Cracking) ................................................................. 76
Chapter 14 Let’s make this perfectly clear … (Formulating clear automotive lacquers) ................................. 82
Chapter 15 That’s swell (HSP and Swelling) ................................................................................................... 87
Chapter 16 Paint failure – the science of blistering .......................................................................................... 90
Chapter 17 Skin deep (HSP and Skin Absorption) ........................................................................................... 94
Chapter 18 HSP and Diffusion ........................................................................................................................ 103
Chapter 19 It’s your call (Rational Selection of Chemical Protective Gloves) .............................................. 122
Chapter 20 Gloves: from minimum data to maximum insight (Exploring Glove Permeation) ...................... 127
Chapter 21 Saving the planet (Finding Improved Environmental Solutions) ................................................. 147
Chapter 22 HSP for ionic liquids (How to Assign HSP to New Materials) ................................................... 156
Chapter 23 Cleaning by numbers (HSP for Surfactants) ................................................................................ 167
Chapter 24 Chromatography – HSP creator and user (Retention Times and HSP) ........................................ 171
Chapter 25 Noses artificial and natural (HSP for Sensors Both Artificial and Live) ..................................... 186
Chapter 26 Attacking DNA (HSP for DNA , Drugs, and Biological Membranes Compared) ....................... 197
Chapter 27 HSP for Pharma and Cosmetic applications ................................................................................. 203
Chapter 28 Exploring with HSP – (Generating and testing research hypotheses) .......................................... 205
Chapter 29 Liquid extraction – a work in progress ......................................................................................... 209
Chapter 30 The HSP of gel formation............................................................................................................. 212
Chapter 31 Going nano (HSP Characterizations of Nanoparticles) ................................................................ 219
Chapter 32 DIY HSP (Methods to Calculate/Estimate Your Own HSP) ....................................................... 229
Chapter 33 Predictions (Many Physical Properties are Correlated with HSP) ............................................... 238
Chapter 34 Improvements? ............................................................................................................................. 243
Chapter 35 Into the 4th Dimension. Donor/Acceptor ..................................................................................... 245
Chapter 36 QSARs .......................................................................................................................................... 250

3
Chapter 37 SFBox-FE ..................................................................................................................................... 262
Chapter 38 A Short History of the Hansen Solubility Parameters .................................................................. 269
Chapter 39 The next steps (What Is Planned and Asked For) ........................................................................ 278

4
Introduction & Guarantee by Steven Abbott
Most practical scientists and technologists who regularly work with solvents and polymers have heard of
Hansen Solubility Parameters (HSP – note that the abbreviation is usually a plural but sometimes can be
singular). They generally know that they are a “good thing” and if pushed they might say that HSP encode the
principle that “like dissolves like”. If you go to the website of any reputable provider of solvents you will find
the HSP of their products as part of their information pack for their customers. So it would be fair to say that
HSP are an accepted part of the technical infrastructure on which we all depend.
Yet if you probe a little deeper, you find two conflicting criticisms of HSP.
The first is that they are too trivial. They can be seen as a cheap trick that’s sometimes useful, but not
something that can take on “real” problems. They can also be seen as “mere correlations” with nothing of
substance to them.
The second criticism is that they are too hard. The poor scientist has to do a lot of work with, say, 48 solvents
then attempt some tricky number crunching in order to determine the HSP of the system in which they are
interested.
The result of these criticisms (sometimes overt, usually un-stated) is that remarkably few of us use HSP as a
routine part of our working lives.
The reason that I’m writing this introduction is that despite having known Charles for many years and reaped
the benefit of his hard work, I’d never really thrown myself the challenge of personally using HSP. This all
changed when I read the 2nd edition of his magisterial Handbook (Hansen, C. M., Hansen Solubility
Parameters: A User’s Handbook, CRC Press, Boca Raton FL, 2007 – referred to as the Handbook in the rest of
this book) where he revealed a few more details of how his Sphere program computed HSP. I realised that I
would be able to write a version for myself. As soon as I got it working, I tried it out on a pressing technical
problem in my business and found that from very little work I got a remarkable amount of insight and, from
that, a new product for the marketplace.
That first version of Sphere now looks rather crude. With Charles’ help and encouragement I added capabilities
that now make it a formidable practical tool that both Charles and I use routinely.
We believe that providing you with a copy of HSPiP, videos (created from within the software) to help you
understand it, a set of the worked examples in this book, and the book itself you will become a convert. You
will find that HSP are neither too trivial nor too hard.
To produce the examples in this book we have had to rely on the generosity of a number of research teams who
have allowed us to use their HSP data. We warmly thank them for their help and acknowledge their individual
contributions in the appropriate chapters.
A word of advice about using HSPiP. It’s really rather powerful. It had to be. Without that power we couldn’t
have done all the things we’ve done for the book. But power tends to come with complexity. I’ve tried to keep
things simple. Which means that some of the power is hidden in neat short-cuts and tricks. To get to know
those, it really will be helpful if you read the Help file. I know that no-one (including myself) reads Help files.
But if you find yourself wishing that HSPiP would do such-and-such, there’s a chance that by looking in the
Help you’ll find that it does.
Because chemical nomenclature is somewhat ambiguous, the chemicals in the main database Sphere Solvent
Data.hsd are provided with CAS Numbers and Smiles nomenclature. Any database is bound to have some
inaccuracies and uncertainties. We’ve done our best to minimize them. We warmly thank our co-author Dr
YAMAMOTO Hiroshi for the huge amount of work he put in to providing the CAS and Smiles data and for his
eagle eyes that detected a number of errors.
Now to the guarantee mentioned at the start of this section. I am taking personal responsibility for some of the
more exploratory ideas in the book and software. I am therefore offering a personal guarantee that when you
show that the ideas are wrong I will (a) upgrade the relevant section(s) of the book/software and (b) make it
clear that I was wrong and (c) acknowledge you (if you wish) as the source of the correction. Science thrives on
its falsifiability and I positively welcome the chance for myself and the HSP community to learn from the
refutation of ideas which seem to me to be reasonable on the basis of the evidence to hand at the time of
writing.
5
Given the litigious environment in which we all live, I have to follow the Guarantee with a disclaimer:
Disclaimer: The theories, examples, formulae, calculations and datasets used in this eBook and software are
based on extensive theoretical research and experimentation over many years by the HSP community. But they
should only be used as a guide to any particular issue. Hansen-Solubility.com cannot be held responsible for
problems resulting from use of the eBook, software and datasets.
Go ahead. Look for the examples that are closest to your technical area, see how valuable the insights from
HSP can be, then use HSPiP to become a convert in the way that I have.
You can email me at steven@stevenabbott.co.uk
Finally, I must thank my company, MacDermid, for giving me the academic freedom to write this book with Dr
Hansen. They are not responsible for any flaws in the book.
Note for the 2nd Edition
HSPiP users have not been shy in offering detailed critiques of the 1st edition, suggestions for improvements,
errors in the data etc. This is exactly what I’d hoped for. In addition to the big changes for the 2nd Edition (e.g.
Y-MB automated HSP and the IGC modeller) there have been many small changes that make it easier for users
to get the most out of HSP. The 2nd Edition Update also added some important extra outputs from Y-MB
including Environmental outputs for intelligent consideration of VOC issues.
Note for the 3rd Edition
I’m now retired from MacDermid and this has given me more time to focus on HSPiP. This was vital because
creating the 3rd Edition has been a major project. We, of course, took on issues raised by the HSPiP community.
But with YAMAMOTO Hiroshi as a key member of the HSPiP team we set ourselves some tough challenges.
As far as we are aware, the severest critics of HSP are ourselves and we spent long hours checking large
datasets for the quality of the predictions by the software. We also found ourselves developing new predictive
techniques for GC, solubility, azeotropes, adhesion etc. The whole HSPiP package now offers a formidable
array of practical predictive tools, backed up by as much validation as we could find. We know it is not perfect.
But, as ever, we hope that you, the user community, will keep us alert to opportunities to improve all aspects of
the package.
We must comment on the Polymer dataset. This has not changed much from the previous editions simply
because little new data has come in. But we’ve added a grading system and a document explaining it, so you
can more readily choose between the different values on offer. At the heart of the problem is the definition of
“polymer” and “soluble”. A “PET” might be one of many things. One data set might be interested in real
solubility of low molecular weight amorphous PET and another might be interested in the swelling of high
molecular weight crystalline PET. Inevitably there will be differences in the results. This isn’t a weakness of
HSP. It’s up to you as a scientist to know what sort of polymer and what sort of solubility is important for you.
If there isn’t good data in the HSPiP set or in the literature then believe me, because I’ve often done it myself,
it’s very easy to measure the value of your polymer for your purposes. We’ve added a whole chapter to explain
in detail how to do this.
The old Sphere Solvent Data.hsd from the previous editions has now been supplanted by an integrated dataset
which contains much more useful data on over 10,000 chemicals spanning a wide range of interests. It’s
divided into two parts. The first ~1200 entries are the “official” set. The data provided with it is, wherever
possible real data. The other ~8,800 entries contain predicted data. But the predictions are based on the real data
available for many of those entries. When you use the program you can include the full dataset by selecting the
“10,000” option. You can still load the old dataset if you wish.
The new version of the eBook comes with an improved reader. A number of users have asked why they can’t
copy or print the text and we always have to give the same reply. We very much want the eBook to remain an
integral part of the whole HSPiP package, and if we allowed it to be copied/printed then pirated editions of the
eBook would quickly appear. We know this is an inconvenience for our honest users, but we really don’t have
much choice in the matter.
Some of the screen shots in the book are from the earlier edition. It didn’t seem necessary to update them all as
you will quickly be able to accommodate any minor differences. For example in the Optimizer the eBook says

6
that the user can click the 2 button to find a good blend of two solvents. It would be obvious to a user of the 3rd
Edition that clicking the 3 button helpfully finds the optimum blend of three solvents.
Looking at the opening of this Introduction I’m pleased with one thing. HSPiP has already started to change
views of HSP. Increasingly they are being seen as having that very rare combination – ease of use with power
of prediction. The feedback from the HSPiP community has confirmed this many times over. The software has
now been cited in prestigious publications and the idea that HSPiP are “mere correlations” now starts to look
rather quaint. The fact that HSP have addressed deep issues in areas as diverse as DNA sequencing and
graphene dispersions indicates that there is a lot more that HSP can do.
In addition to the official Hansen website, www.hansen-solubility.com we encourage you to visit Dr
Yamamoto’s spirited Pirika site, http://www.pirika.com/NewHP/PirikaE/HSP.html . Hiroshi enjoys pushing the
boundaries of HSP with data-driven speculations. If you disagree with his speculations, he will be happy to
respond to your views.
Finally, two Thank You paragraphs.
The first is to the HSPiP community. The interaction with you, the challenges, the queries, the requests, the
feedback have all been much appreciated. With such a large community I can’t always guarantee to give an
instant and satisfactory response, but I can guarantee to try my best.
The second, and he doesn’t know that I’ve added this bit to the text, is to Charles. His constant wisdom and
encouragement, his astonishing collection of historical data and papers (and ability to find just the right bit of
information), his razor-sharp mind for piercing through the inadequacies of my own understanding have always
been much appreciated by me. He’s never doubted, over 40 years, that HSP could continue to provide key
insights into real-world technical problems. HSPiP and the HSPiP community have more than proved him
correct in his views.
Note for the 3.1 release
Another meeting of the HSPiP team led to the usual lively debate about improvements to HSPiP. Because the
HSPiP user community is, we’re delighted to say, very demanding there were many ideas for improvements, as
well as our own roadmap. We reached a few key decisions:
1. To “liberate” the eBook. It’s now included as a straightforward PDF file which you can read, print etc.
as you wish. The Book icon simply opens the file in Acrobat Reader.
2. To update a few HSP values in the dataset. This was prompted by Hiroshi’s careful analysis of
anomalies in values of different series and also some fresh experimental data from an HSP user on
dimethyl and diethyl succinate. These data have caused us to add specific entries for many of these
important esters (glutarate, adipate…) and, most significantly, to change the value for DBE which is a
mixture of such esters. The previous value had been worked out many years ago with a major DBE
manufacturer and seemed to be correct based on the data at the time. But the data on dimethyl succinate
was compelling so we have had no choice but to update the value. The changed molecules are: 1,2,3-
Trichloro Propene; 4-Ethyl-1,3-Dioxolane-2-one; N-Ethyl Formamide; Propionamide; N-
Acetylmorpholine; DBE. We’ve also added some important new solvents: Dimethyl 2-Methyl Glutarate
(a variation on the DBE esters); some potentially interesting bio-derived solvents with interesting
properties: Glycerol carbonate and its acetate and ether; Dimethyl Isosorbide, of great interest to the
cosmetics community.
3. To make sure that users knew what changes have been made to each update (major or minor) so in
future all releases will come with a Version Information document.
4. To add “advanced” options to the sphere fitting. There is now a “data” method where you can enter
actual solubilities, swellabilities etc. And, with many reservations about it, a double sphere method
which acknowledges that some materials may have two domains (e.g. a diblock co-polymer) and
therefore two spheres.
5. To change the standard format of files from .ssd (Sphere Solvent Data) to .hsd (Hansen Solubility Data).
You will still be able to read all your old .ssd files but if you re-save them then they are saved in the .hsd
format. The reason for the change is simple. There is so much possible information to be stored in these
files that we had to “liberate” the format from its rigid structure. Now each file comes with a header row
and the program does its best to identify the key components and create a table with the standard

7
elements in their usual place and any other elements (which could be user-specific if you wish) are
tacked on to the end.
6. To assign Donor/Acceptor values to the δH value to capture the different modes of hydrogen bonding
that are possible. This was our most difficult decision. As you will read in the 4th Dimension chapter this
change hasn’t caused the entire 3-parameter HSP to collapse. Indeed, there are many good reasons why
this change makes very little difference in most cases. Read the new chapter to find out why.
Finally, Hiroshi and I persuaded Charles to write a short history of HSP. This followed his comment during our
meeting that “I never called them Hansen Solubility Parameters until Beerbower started to use the term in his
publications”. As his original version (the Main Track) was a rather too short and modest, we asked him to add
some more personal details (the Side Track). His description of how he devised the first 88 HSP values using
rods, wires and magnets was fascinating and we begged to see a photo. But it seemed that no image existed.
However, after almost giving up the search through his archives, Charles found a picture which we’ve scanned
in as best we can. For those who are used to doing optimization at the push of an Excel button it’s sobering to
imagine how much hard work went in to devising the whole basis of HSP with such apparatus.
Note for the 4th Edition
The world is changing to “apps” that can be used on all platforms from phones through tablets to PCs and
Macs. In the long term, HSPiP will have to be liberated from its reliance on the PC platform. For now that is
impossible – apps and the variety of browsers place crippling limitations on what can be done, making any
attempt at an app-based HSPiP futile. But to help us on the journey we have started the tradition of Power Tools
– extra tools that aren’t absolutely required for day-to-day use in HSPiP but which will appeal to users with
specific needs. The flexibility to create new Power Tools and, importantly, to allow users to upgrade quickly to
the newest versions means that the HSPiP community can adapt quickly. And one of the key Power Tools is a
Sphere Viewer which lets you send .hsd files to colleagues who do not have HSPiP, allowing them to see
Sphere fits for themselves. The limitations on Power Tools are almost all to do with the different philosophies
of different browsers with, inevitably, Internet Explorer being the one with the most problems and the least
functionality for users.
The addition of Power Tools has not distracted us from HSPiP itself. As promised, an improved Y-MB engine
offers better predictions of HSP and other properties. The downside of that good news is that you will find that
some molecules of interest to you will have changed predicted values. As always with predictions, as a scientist
you have to choose which values to prefer. As we’ve often said, precise prediction is impossible – we can only
strive to reduce errors across a broad range of molecules. There are refinements to polymer predictions, the
introduction of a prediction of EACN (Effective Alkane Carbon Number) which are increasingly important in
terms of surfactant theory, especially HLD-NAC for which I’ve written separate (and free) apps. We would
have loved to have had a breakthrough in applying HSP theory to surfactants, but it’s a long road ahead.
There are numerous other enhancements to the 4th Edition, building on the fact that within v3.1 we continued
the tradition of a steady stream of improvements based on user feedback. The three of us continue to greatly
appreciate the ideas, challenges and bug reports from users. Together we have made the whole a much better
package, which is why we continue to offer users free upgrades as a “thank you” for their contributions.
The eBook has had a modest refresh. A key refinement has been within the Diffusion chapter. Charles has been
working hard to put the record straight on how to interpret so-called special cases and has convincingly shown
the errors and inconsistencies in proposed alternatives to the straightforward interpretation that combines
Fickian diffusion with surface resistance, and which is built in to the Diffusion modeller. For those wishing to
know even more via a step-by-step guide to the theory, the Power Tools include a copy of Practical Diffusion
written by myself and Charles.
Version 4.1 of the eBook now includes the FAQ chapter. Although the themes within the chapter are discussed
throughout the book, it was clear that some concise discussions on the most frequently asked questions was
required. If you already know the book fairly well, then all you need to do is go to the FAQ chapter to get some
fresh views of some key issues.
Note for Version 5 and Predictive Power
The Y-MB engine was initially intended to do just one thing – estimate HSP values. But Hiroshi has always had
a deep interest in predicting other properties and the more he tried to improve the HSP predictions, the more he
8
needed to improve other predictions such as density – HSP are Cohesive Energy Density so you’d better get
Density right. And because HSP are related to activity coefficients it’s good to try to get good VLE (Vapour
Liquid Equilibria) predictions, which needs improved Wilson parameters, Antoine Constants and so forth.
With Version 5 it’s now official that HSPiP is also a big properties prediction machine. Hiroshi’s hard work
gives world-class predictions of a large range of properties (more than 50) and so we’ve reconfigured the
database to show all the predicted values (with experimental ones, where available, in the Hansen set). This
means, unfortunately, some adjustment to the interface and a few of the least-used options have had to be
withdrawn because the new Y-MB engine works in a different manner. It occurred to us that we had tended to
downpay the power of HSPiP, so with this version we are proud to use the theme “Predictive Power” to capture
the fact that HSPiP contains a formidable
It also means, once again, that predicted values are different from previous versions. That’s progress. You can’t
have improved predictive capability and keep the same values as last time.
And we’ve added a full QSAR (Quantitative Structure Activity Relationship) engine that uses the Y-MB
predictions as a basis for fitting your chemical datasets both to gain understanding of the phenomena (e.g.
which parameters are important for the fit) and also to provide predictive capacity (more Predictive Power) for
molecules not in the test set.
We thought of charging for the QSAR, but we so enjoy pleasantly surprising the large HSPiP user community
that it comes free.
Because the new VLE engine produces impressively good predictions (even more Predictive Power), including
for azeotropes, we have provided a large set of experimental values that you can load and compare against the
predicted values.
We were delighted to work with the Toxics Use Reduction Institute (TURI) at the University of Massachusetts
Lowell to help them and their associated planners to use HSPiP for reducing their use of toxic chemicals. TURI
then sent us a large list of “potentially safer chemicals” (to be used with caution because of chemical data gaps,
limitations, and uncertainty, as well as lack of resources to fully evaluate all existing chemical data) which we
jointly turned into a .hsd and a .sof data file which you may wish to use in your own explorations for
substituting toxic chemicals with safer ones.
Version 5 is a result of not one but two of our annual HSPiP Developers Conference – a frenetic fortnight of
discussions, arguments and coding, followed 4 months later with a frenetic week of fine-tuning the whole
package. The fine-tuning was necessary because the interface had to be re-written for the new generation of
Windows 10, for 64x machines and for laptops with small screens and high resolutions that have posed many
problems for all who write software for Windows. The new version works much better across a wide range of
screen settings (and with multiple monitors which themselves have different resolutions). All previous versions
have worked on all machines, including ancient Win XP, and also worked fairly well on many video projectors,
but version 5 is for the future so is only suitable for high-res, modern machines and high-quality video
projectors.
Two years on from 5.0.x we came to 5.1.x and finally changed the default file structure from .hsd and .sof files
which were really handy simple text files, to the more powerful .hsdx and sofx files which are in XML format,
more powerful but slightly less convenient for those who need to open them outside HSPiP. The gains are
greater internal coding elegance, with room for expansion of data columns and, at last, the addition of CAS #
and SMILES columns to our standard datasets.
Then came 5.2.x with the powerful new Y-PB method for estimating the properties of polymers. As an extra
bonus, the SMILES is automatically converted to a 3D molecule using the awesome RDKit from Greg
Landrum, for which we express our warmest thanks. As a formal note:
RDKIt is licensed under the Creative Commons Attribution-ShareAlike 4.0 License. To view a copy of this
license, visit http://creativecommons.org/licenses/by-sa/4.0/ or send a letter to Creative Commons, 543 Howard
Street, 5th Floor, San Francisco, California, 94105, USA. The intent of this license is similar to that of the
RDKit itself. In simple words: “Do whatever you want with it, but please give us some credit.”
By the time we reached 5.2.07 we’d used RDKit to do sub-structure searches, a powerful way to find, for
example, work with FindMols to get a list of all the molecules in the dataset with a required functionality.
9
With 5.3.x we introduced the powerful Scheutjend-Fleer (SF) or Self-Consistent Field (SCF) theory that allows
us to link HSP to the behaviour of complex particle-polymer-solvent systems. This coincides with Dr Seishi
Shimizu’s breakthrough paper that proves at a fundamental level that nanoparticles should be addressed via
“solubility” theory not through “dispersion” theory – when there’s no statisfactory definition of what
“dispersion” means. So the whole complex field of particle-based formulation takes on a new level of analysis.
At its core, the new theory confirms the intuitions and observations that Charles made many decades ago when
formulating pigments and dispersants in paints. It simply takes it to the next level of refinement.

10
Chapter 1 The Minimum Possible Theory (Simple Introduction)
Although we want HSP to be practical, we don’t want you to think that they are magic or “just a bunch of
correlations”. At the same time, we don’t want to bog you down with unnecessary theory. So here is the
minimum possible theory necessary for a practical user of HSP.
Kinetics versus Thermodynamics
Thermodynamics tells you if something is possible nor not. You can dissolve sodium chloride in water because
solvated sodium and chloride ions are thermodynamically more stable (energy and entropy) than crystalline
sodium chloride. Barium sulphate crystals are thermodynamically more stable than solvated barium and
sulphate ions, so barium sulphate is essentially insoluble.
Kinetics tells you how fast something will happen if it is thermodynamically possible. So kinetics have nothing
much useful to say about dissolving barium sulphate. But it’s entirely possible to have lots of salt and water in
close proximity without much of the salt dissolving if you don’t get the kinetics right. One large lump of salt
sitting in some very cold water will dissolve far less quickly than a well-stirred fine salt powder in warm water.
Thermodynamics and kinetics are both powerful. But ultimately it’s thermodynamics which is the more
powerful. Kinetics might suggest that you should try harder to dissolve the barium sulphate, but
thermodynamics tells you that you shouldn’t bother. The observation of a slow-dissolving lump of salt might
suggest to you that it’s going to be impossible, but thermodynamics encourages you to try.
So let’s make it clear. The strength of HSP is that they are based on thermodynamics. They are all about
whether something is fundamentally possible or not. We won’t hide from you the fact that kinetics can
sometimes wreck even the best thermodynamic predictions of HSP. But the fact that HSP are essentially a way
for you to reach profound thermodynamic conclusions is their prime strength.
It will become tedious to insert “thermodynamically” into every sentence which says “HSP show that
thermodynamically A will dissolve in B”, so let’s take it that we now understand the difference between
kinetics and thermodynamics.
Note to the sceptics: HSP really do come from deep thermodynamic insights. The fact that most HSP have been
determined by correlation experiments reflects a limitation on our ability to do complex thermodynamic
calculations rather than a limitation of HSP themselves. The recent work of Panayiotou has at last accurately
derived HSP from first principles – with remarkable agreement with the experimentally derived values.
Similarly, the molecular dynamics work of Goddard’s group at CalTech has produced accurate numbers,
showing that it is possible for anyone to obtain HSP from first principles.
Doing it the hard way
If you want to dissolve something in something else then you have to compare two energy losses with one
energy gain. The first loss is the mutual interaction of the solvent with itself. You are effectively making a hole
in the solvent and that takes energy. The more the solvent attracts itself, the more energy it takes. The second
loss is from the mutual interaction of the solute with itself – for the same reason. And the gain is the interaction
of the solvent with the solute. If this interaction is greater than the sum of the losses, then the solute will
dissolve.
So if you want to know if A dissolves in B, “all” you have to do is to calculate the two losses and the one gain.
For simple systems this can be done, but it becomes impossibly hard for more complex systems. And when you
start trying to work out the best mixture of C, D and E in which to dissolve A it’s even more impossible.
The glory of HSP is that in 3 numbers, all those fussy thermodynamic calculations are done for you, with a high
degree of accuracy.
1, 2, 3 (or more?) energies
If we are going to short-cut the hard way, we need to have numbers that characterise the internal energies (the
energy required to create the hole in the solvent and break up the solute) and also the interactive energy.
You could imagine that if the chemicals were all of one general type then one energy value could be sufficient
to enable the calculations. Hildebrand famously tried to do everything via just one energy, but although that one
energy is fundamental, without partitioning it, its predictive value proved to be limited. Indeed we are
11
astonished that Hildebrand parameters still continue to be used. There are many knock-out arguments against
using Hildebrand (see the Hansen-Solubility website for a more detailed review) but one simple example says it
all. Epoxies aren’t generally soluble in nitromethane or in butanol which, as it happens, have the same
Hildebrand parameters. But a 50:50 mix of these two solvents is a good solvent for epoxies. As we will shortly
see, this is easily explained by Hansen parameters and is inexplicable with Hildebrand.
As practical scientists we know that there are at least 4 fairly distinctive forms of energy:
• Dispersion forces (atomic). These are the general van der Waals interactions between just about
everything. Put any molecule a few Angstrom from another molecule and you get a powerful attractive
force between the atoms of the two molecules. Because they are everywhere, and because they are
unglamorous we tend to ignore them, but they are the dominating force in most interactions! The
famous gecko effect that allows a gecko to walk upside down on a ceiling is due almost entirely to the
amazing strength of dispersion forces.
• Polar forces (molecular). These are the familiar “positive attracts negative” electrical attractions arising
from dipole moments. They are important in just about every molecule except some hydrocarbons and
special chemicals consisting of only carbon and fluorine.
• Hydrogen bond forces (molecular) are arguably a type of polar force. But their predictive value in many
different aspects of science goes beyond simply thinking of them as polar forces so it seems worthwhile
to make them distinct. More generally they can be considered as a form of electron exchange so that
CO2 shows strong “hydrogen bonding” forces that make it a good solvent for e.g. caffeine even though
it contains no hydrogen atoms.
• Ionic forces. These are what keep inorganic crystals together.
If you are going to describe molecular interactions in simple numbers it’s clear that you would need at least 2
for every molecule: Dispersion and Polar. By including the third parameter, Hydrogen bonding, everything
except strong ionic interactions became thermodynamically predictable. It turns out that even for organic salts
the polar and hydrogen bonding contributions are sufficient. And as ionic interactions are mostly the domain of
aqueous environments dominated by the extraordinary properties of water, it doesn’t seem to be useful to
include a 4th descriptive parameter when you are trying to understand interactions that don’t involve large
amounts of water. There is a lot of progress being made, but the division of energy types in the aqueous domain
is still not fully understood.
So it seems reasonable that three parameters could be used to describe solvent/solute interactions. But why
should something as simple as 3 numbers be sufficient to describe a process which, by our own admission, is
far too complex for the best computers to calculate?
Do 3 numbers give accurate predictions?
Yes. The data is overwhelming. We’ll come back to that in a moment.
Aren’t 4 numbers even better?
Yes, and no. In principle, dividing the Hydrogen Bonding parameter into Donar/Acceptor terms (as, for
example, in MOSCED) should give even better results. But the practical problems of creating a large, self-
consistent database with 4 parameters, and of visualising issues in 4D space mean that until recently this has not
proven to be a popular way forward. The 4th Dimension chapter provides and update on progress.
Why (in principle) does it work?
The strength of HSP is that they are based on thermodynamics. And the key insight that led to the creation of
thermodynamics is that the law of large numbers lets you calculate things that can’t be done by attending to
individual details. It’s hard to calculate the force on the wall of a container containing 1 trillion gas molecules if
you try to consider what’s happening to each of the trillion molecules, yet it’s easy, and accurate, to calculate
via simple thermodynamic gas laws.
The same applies to HSP. The dispersion, polar and hydrogen bonding forces are impossibly hard to calculate
via the interactions of trillions of individual molecules, yet are easily encoded in the HSP numbers.

12
We have to stress again, that if you can do the calculations (and it is becoming increasingly routine to do them),
then the calculated results confirm the numbers you find listed in the tables of HSP.
Do 3 numbers give accurate predictions?
Let’s think of the most basic thermodynamic situation. We are trying to mix solvent A with solute B. The claim
is that you will have to lose and gain energies. How can we calculate those?
A naïve approach would be to calculate the sum of the (absolute) differences of the three HSP. By definition, if
B is so close to A that it’s the same molecule then these differences will be zero. So the definition of a perfect
solvent is a difference of 0. If A and B are chemically fairly similar then you would expect their HSP to be
similar, and the differences to be small. And if they are utterly different, the difference should be large.
So we might try:
Equ. 1-1 Difference = [DispersionA-DispersionB] + [PolarA-PolarB] + [Hydrogen BondingA – Hydrogen bondingB]
where the [square brackets] imply the absolute value.
As it happens, you can’t add and subtract energies quite like this. If we introduce δD, δP and δH for Dispersion,
Polar and Hydrogen bonding parameters then the true difference is:
Equ. 1-2 Difference2 =4 (δDA-δDB)2 + (δPA-δPB)2 + (δHA-δHB)2
The squared terms mean that we don’t have to worry about absolute values as (δDA-δDB)2 is the same as (δDB-
δDA)2. The units of these solubility parameters are (Joules/cm³)½ or, equivalently, MPa½. In older papers you
will see the units expressed as (cal/cm³) ½. If you ever need to convert between old units, simply multiply by a
factor of 2 (or 2.046 if you want to be precise). Throughout this book and in the software, all quoted values are
in (Joules/cm³)½ or, if you prefer, MPa½. As Molar Volume (MVol) is commonly used throughout the book it’s
worth stating here that its units are cm3/mole. Note, too, that all quoted values are at the standard temperature of
25ºC.
You’ll have noticed that the famous factor of four in front of the δD term has crept into the formula. Many have
questioned the justification for this factor. In the Handbook (pp30-31), Hansen provides some interesting
possibilities based on Prigogine’s Corresponding States Theory. At the heart of the issue is whether the
“geometric mean” is the best way to calculate the differential heat of mixing between components. For non-
polar spherical molecules interacting via Lennard-Jones potential there’s a good case that this is a good
approximation. But there is no reason to believe that the same should apply to polar and hydrogen-bonding
interactions. Furthermore, there is universal agreement amongst diverse luminaries such as Prausnitz, Good,
Beerbower and Gardon that the differential heat of mixing term should be less for polar and hydrogen bonding
than for dispersive forces. How much less is a matter of debate, but values between 1/8 and 1/2 have received
support in a wide range of experiments, with the value of 1/4 providing the best data fit for Hansens’s
polymer/solvent data. So although we regret that we, like everyone else, cannot provide a compelling argument
that the factor should be precisely four, we are confident that it should be at least a factor of 2. Because the
factor of 4 gives spherical plots, fits well with the largest range of practical test correlations and has stood the
test of time in such a wide variety of real-world uses we feel that its continued use is more than justified.
This famous difference equation is the core of HSP. For any problem you just calculate the difference. If it’s
small then the thermodynamic chances are high that the two components will be mutually soluble (or
compatible or, well, “happy” together if you have some interaction such as pigment dispersion where you know
what “happy” means, even if it can’t be defined precisely). If the distance is large then the chances are small.
For those who are interested in the theory, for most cases, a distance greater than 0 means that mixing is
enthalpically unfavourable. But of course mixing tends to increase entropy so the total is energetically
favourable. The smaller the distance, the less you have to rely on entropy to help you. Large polymers have less
entropy gain when they are dissolved so you need a smaller distance from the polymers’ HSP in order to
dissolve them.
That would be fine, but rather limiting. The true power of HSP is that because they are based on the
thermodynamic law of large numbers, a “solvent” can be a mixture of an arbitrary number of components and
the “solvent’s” HSP are simply the average (weighted for % contribution) of the individual components.

13
Here are two examples to show the principle. In both cases there happens to be a 50:50 mixture (so you can
check the answer by inspection). And in both cases you obtain effectively the same solvent, even though they
are created from very different starting solvents.
δD δP δH %
Solvent X 16 8 2 50
Solvent Y 18 10 4 50
Mixture 17 9 3 100

δD δP δH %
Solvent X 14 0 0 50
Solvent Y 21 18 6 50
Mixture 17 9 3 100

Table 1-1 Creating the same solvent properties from very different solvent blends
This is the real power of HSP. A striking example is when you want the HSP of a particular solvent but can’t
use that solvent because it is toxic or too expensive. Simply mix together two (possibly widely different)
safe/cheap solvents in proportions that give you the correct HSP and you have a fully functional solvent,
indistinguishable (as far as the solute is concerned) from the original solvent.
As we will often be referring to sets of δD, δP, δH numbers always in that order we introduce the convention
that [17, 9, 3] means “δD=17, δP=9 and δH=3”. This shorthand makes it easy for us to say that a 50:50 mix of
[16, 8, 2] and [18, 10, 4] gives [17, 9, 3]. When, later on, we introduce a Radius, this will be the fourth element
so [17, 9, 3, 8] means [17, 9, 3] with a Radius of 8.
Let’s go back to the question: Do 3 numbers give accurate predictions? The answer is overwhelming. Not only
is it the case that the 3 numbers do work, but in fact they must work. This is thermodynamics and one of the
great rules of life is never to argue with the laws of thermodynamics.
Do they work all the time? They can’t work outside their own thermodynamic area. So they cannot work for
ionic solids and they are not much of a guide for anything to do with primarily aqueous solutions (though
pioneers are doing good work in this area). And of course there will be times when the HSP will say that a
given solvent blend will dissolve a certain polymer but experiments show that it merely swells. This is because
its excessively high molecular weight means that it will take far too long to dissolve it and so only swells it.
That’s the limit of kinetics versus thermodynamics.
Even here, the HSP can be deeply insightful. If the HSP for the polymer has been determined using a low
molecular weight, then it is entropically probable that solvents which were just good enough for the low
molecular weight version will be inadequate for the high molecular weight. The need for a solvent closer to the
HSP of the polymer, or for a solvent with a lower molar volume, is therefore predictable.
Work by Professor Coleman’s team in Trinity College Dublin has provided an important update to theory. For
many years it was assumed that HSP could be applied to nanoparticles and pigments. There was no explicit
justification for this other than (as we will see) the technique works very well. The paper by J. Marguerite
Hughes, Damian Aherne, Jonathan N. Coleman. Generalizing Solubility Parameter Theory to Apply to One-
and Two-Dimensional Solutes and to Incorporate Dipolar Interactions, J. Appl. Polym. Sci. 2012, DOI:
10.1002/APP.38051 shows that HSP really can apply to nanoparticles such as CNT and graphene.
For some reason, HSP have irritated many people over the years. There have been many attempts to overturn
them, but there is an overwhelming amount of theory and practical success in support. One classic criticism was
around “negative heats of mixing”. Both HSP and the earlier Hildebrand parameter seemingly allowed
“positive” heats of mixing only. This situation was cleared up by some skilful thermodynamic calculations on
solubility parameters which showed that both positive and negative heats of mixing were not only allowed, but
were also required. Experiments by Patterson and Delmas (see, for example, Patterson D., Delmas G., New

14
Aspects of Polymer Solution Thermodynamics, Off. Dig. Fed. Soc. Paint Technol., 34,677,1962) confirmed
these calculations.
So for all practical purposes you can take it that HSP do work and must work.
So all you have to do is to get to grips with the Sphere in the next chapter.
Deeper theory
The Sticking, Flowing, Dissolving chapter contains some deeper theory about aspects of polymer/solvent
solubility. Once you’ve become comfortable with HSP and HSPiP at this basic level of theory, you might want
to dip into sections of that chapter to find out more.

15
Chapter 2 The Sphere (The Preferred Method of Visualizing)
If 2 parameters were good enough to work, then life would have been very easy and HSP would have been used
much more frequently. You could plot everything on a nice X-Y graph.
The problem with HSP is that they need to be plotted in 3D. With modern software this isn’t so hard. But for
papers and books it’s more usual to use 2D plots and this creates an interesting problem.
Most people associate HSP plots with δP v δH graphs, Polar versus Hydrogen bonding. If you have to have just
one plot, then this is the most important. But this tends to diminish the significance of Dispersion, which is
unfortunate. As 3D graphs are hard to understand when shown statically in a book, we will sometimes have to
give you sets of 3 graphs, P v H, H v D, P v D. This is a bit cumbersome, but it’s better than leading you astray
with just the P v H plot. Because you get the software with this book, we generally supply just the 3D plot and
urge you to look at each example live as it is a richer interactive experience.
P v H means H is plotted along the X axis and P along the Y axis and similarly for H v D and P v D..
But first let’s see what the Sphere can teach us.

Figure 2-1 Using file Chapter2


Let’s orient ourselves. The graph shows the P-axis going vertically from 0-25. The H-axis also goes from 0-25
from left-to-middle right. The D-axis goes from 12.5-25 from right-to-middle left. In the program you can
orient the plot any way you like, but in the book we will always keep this orientation.
There is a large green sphere. In the exact centre is a green dot. This shows the position of a particular polymer
with values D=18.5, P=9.9, H=7.9. There is a set of blue dots. These are all inside the sphere (though that is not
obvious from the static view on the page). They are located at the HSP of solvents which dissolve this polymer.
The red cubes show the solvents which do not dissolve this polymer.
The program has taken the will/won’t dissolve data of each solvent and calculated the sphere which includes all
the good solvents (defined in HSPiP as “Inside” the Sphere) and excludes all the bad solvents. If the sphere
were any smaller then it would exclude some of the good solvents and if it were bigger it would include some
of the bad solvents. And of course if the calculated centre of the sphere were in any other position the sphere
would exclude/include good/bad solvents.
In this example, 16 solvents have been used, 9 of them are good and 7 of them are bad. Some of the solvents
used for the test are definitely not ones you would want to use if you wanted to dissolve lots of this polymer.
The list includes chlorobenzene and nitrobenzene. We will see later on why we include unpleasant solvents in
the tests.

16
If you wanted to find a good solvent to dissolve this particular polymer you would find that there isn’t any close
match amongst reasonably safe solvents. However, a 77:23 mix of 1,3-Dioxolane and Ethylene carbonate is a
good match.
Here’s the first reason you need the Sphere. This reason is based on the centre of the sphere, which shows you
the optimum HSP for good solvency. How likely is it that you would have come up with this particular 77:23
mix by chance? It’s highly unlikely. In all probability you’d have very little idea of what an optimum
could/should look like and you would be even more unlikely to come up with that near-ideal mixture.
Incidentally, if you used a classic 2D HSP plot you would conclude that acetone is near-perfect match. In H-P
space it is perfect but the D value is way off.
The second reason is based on the radius of the sphere. There are plenty of reasons why you wouldn’t want to
use that 77:23 mix, which is optimal only for thermodynamic solvency. You might want to reduce cost or meet
some environmental regulations. So you need to find a solvent mix which is “good enough”. What is the
definition of “good enough”? It’s simple. If a solvent blend is inside the sphere then it’s likely to dissolve the
polymer and the closer it is to the centre, the better it is.
So you can now play around with solvent blends in a rational manner. Simply calculate their weighted average
HSP and check how close they are to the centre of the sphere. If you want to be really sophisticated (and in a
later chapter we show how to use the HSPiP Solvent Optimizer to do this), you can do an optimisation using the
HSP distance from the centre as one parameter (bigger is worse, and outside the radius is scored as infinitely
bad) along with, say, cost and environmental impact as the other parameters. How you choose to weight these
three different parameters only you can say. But without knowing both the centre and radius of this polymer
you would not know where to start.
To do that optimisation you need to know “the HSP distance from the centre”. How is this calculated? You
already know the answer:
Equ. 2-1 Distance2 =4 (δDA-δDB)2 + (δPA-δPB)2 + (δHA-δHB)2
And that’s all you really need to know about HSP and the Sphere!
But because the term is widely used and is a useful shortcut, let’s give you one more concept.
The RED number is the Relative Energy Difference and is simply the ratio of the distance of your solvent
(blend) to the radius of the Sphere. A perfect solvent has a RED of 0. A solvent just on the surface of the
Sphere has a RED of 1. It is a useful shorthand that gives quick insights into what’s going on. Relative REDs
are useful. If you have a solvent of RED 0.2 and another of 0.4 you know (a) that neither is perfect and (b) that
the first one is better.
Let’s see how the RED number can help you avoid a simple mistake. We promised that we’d show you some of

the 2D plots. Here they are for this example:

Figure 2-2 Using file Chapter2


In the program, as you move the mouse over each plot you get a read-out of each solvent’s name and HSP.
Note how misleading these plots are. Two “red” solvents are inside the P v H plot which is what most look at. If
17
you relied only on that plot you would be easily confused. For example the red square on top of the blue dot at
7 o’clock in the P v H plot are diethyl ether [14.5, 2.9, 5.1] which is a non-solvent and trichloroethylene [18,
3.1, 5.3] which is a solvent. Their RED numbers are 1.33 and 0.89 respectively which indicates that diethyl
ether is significantly outside the sphere. In this case it is obvious that the low D of ether (14.5) makes it highly
unsuitable for a polymer with a D of 18.5. The trichloroethylene’s D value of 18 explains why its RED is so
much lower.
As we will find in the chapters which follow, the beauty of Sphere is that it captures the essence of a huge
variety of different phenomena. Armed just with the HSP of solvents (or solvent blends) and the HSP and radii
of polymers we can make reliable predictions that work both in the domain of pure science and in the world of
industrial applications. Although HSP have their limitations, it is hard to find an alternative approach that
combines thermodynamic soundness with practical insight.
Let’s first see how HSP can be applied to a basic issue – cleaning up an ink or paint by dissolving the polymers
which hold the ink or paint together.
Note For those who like triangular graphs we’ve added the Teas plot, named after Jean Teas its inventor. It
simply plots δD/(δD+δP+δH) along one axis against δP/(δD+δP+δH) on the next and δH/(δD+δP+δH) on the
third. Although this is a neat way to condense 3D data into 2D, there is no scientific significance to the plotted
values! A perfect Sphere with all “good” inside and all “bad” outside can look a muddle in the Teas plot. And
make sure you test it with diethyl ether and chloroform. You will see that these very different solvents appear
close together in Teas – highlighting how misleading the plot can be. However, some people find it visually
useful. As a visual aid, the computed centre and radius are plotted using their own (δD+δP+δH) value along
with the “bounding circle” and its centre. Again these have no great scientific significance. When you move
your mouse over the Teas plot you get either the % δD, %δP, %δH or if you are near a solvent, the actual δD,
δP, δH and the solvent name.

18
Chapter 3 Your first HSP Sphere (Determining the HSP Sphere)
In earlier editions of HSPiP we took it for granted that the explanation in the previous chapter was sufficient to
help users calculate their first solubility Sphere. But we found that users were very unsure of themselves and we
often had to email advice to them. Thinking back, one of us (Abbott) remembers how nervous he was with his
first Sphere, so the following practical guide reflects his memories of his first Sphere, backed up by the
experience of measuring many different Spheres over 40 years by Hansen.
Let’s start with a typical example. You want to dissolve a polymer for some sort of processing such as coating.
You have severe restrictions on cost, health & safety or environmental impact so finding a good solvent blend
is very hard. If, instead, you are interested in dispersing a pigment or nanoparticle, the discussion below is
identical, just substitute “pigment” or “nanoparticle” for “polymer”.
Our task is to define the HSP of the polymer. A solvent (blend) that is close to that HSP will be a good solvent.
Here’s how to do it.
Get about 20 glass vials. Put a small sample of the polymer into each of the vials. Now add a small amount (say
5ml) of a different solvent to each of the vials. Which solvents should you use? Well, the following list is a
pretty good mixture of relatively common and relatively safe solvents. It’s included as Test Solvents.hsd with
the package and you can easily add/remove solvents to suit your needs.

Solvent D P H
1.4-DIOXANE 17.5 1.8 9
1-BUTANOL 16 5.7 15.8
2-PHENOXY ETHANOL 17.8 5.7 14.3
ACETONE 15.5 10.4 7
ACETONITRILE 15.3 18 6.1
CHLOROFORM 17.8 3.1 5.7
CYCLOHEXANE 16.8 0 0.2
CYCLOHEXANOL 17.4 4.1 13.5
DBE 16.2 4.7 8.4
DIACETONE ALCOHOL 15.8 8.2 10.8
DIETHYLENE GLYCOL 16.6 12 20.7
DIMETHYL FORMAMIDE 17.4 13.7 11.3
DIMETHYL SULFOXIDE 18.4 16.4 10.2
DIPROPYLENE GLYCOL 16.5 10.6 17.7
ETHANOL 99.9% 15.8 8.8 19.4
ETHYL ACETATE 15.8 5.3 7.2
GAMMA BUTYROLACTONE 19 16.6 7.4
HEXANE 14.9 0 0
MEK 16 9 5.1
METHANOL 15.1 12.3 22.3
METHYL ISOBUTYL KETONE 15.3 6.1 4.1
METHYLENE DICHLORIDE 18.2 6.3 6.1
n-BUTYL ACETATE 15.8 3.7 6.3
N-METHYL PYRROLIDONE 18 12.3 7.2
PM 15.6 6.3 11.6

19
PMA 15.6 5.6 9.8
PROPYLENE CARBONATE 20 18 4.1
TCE TETRACHLOROETHYLENE 18 5 0
TETRAHYDROFURAN 16.8 5.7 8
TOLUENE 18 1.4 2

Table 3-1 A reasonable set of test solvents from Test Solvents.hsd


But if you don’t have all of them, no matter. And if you have some others that’s fine. Just don’t have too many
of the same thing. It probably doesn’t help to have methanol, ethanol, propanol and butanol or pentane, hexane,
heptanes and octane. Just ethanol and hexane will be good enough.
Now you need to find out in which solvents the polymer is soluble. Here you have to make a decision. You can,
for example, just hand shake each sample and then see which ones give clear solutions and which ones don’t.
But often you find that such a test is useless as even good solvents might take some time to dissolve your
polymer. So you might decide to put all the vials into an ultrasonic bath for 10 minutes then inspect each vial as
soon as it comes out of the bath. But often you find that everything gets dispersed by the ultrasonics so that
everything looks to be “soluble”. So you might decide to check after the samples have sat for 10 minutes at
room temperature. And maybe the polymer doesn’t “dissolve” in any of the solvents (perhaps you had too much
polymer or too little solvent) but you can still see that the polymer is highly swollen in some solvents and
completely unswollen in others so you can use that as your criterion for “good” and “bad” solvents. For some
polymers you might even have to wait for days to distinguish “good” and “bad”.
Don’t despair if the previous paragraph is too vague for you. You are a scientist and you probably already have
a good intuition about the general solubility properties of your polymer, so you can make a good decision about
what treatment to adopt and whether you judge by dissolution or by swelling.
One word of advice. Because the effects of temperature on solubility can be quite complex, stick with room
temperature tests if possible – at least till you have gained some experience in the whole process.
And if you aren’t interested in polymers but in, say, pigments, the above description applies to you too. Just
find a set of conditions where your pigment is obviously “happy” in some solvents (e.g. nicely and permanently
dispersed) and “unhappy” (e.g. sitting as a lump at the bottom of the vial) in others. Be careful to check if any
pigment is stuck on the side of glass containers.
If you are lucky (or already have a good intuition) after a moderate effort you now have a list of good and bad
solvents. If your system is entirely new then you might spend a day finding the appropriate test conditions,
sometimes getting sidetracked when a “clear” solution is in reality a blob of polymer stuck under the cap of the
vial. But once you have your list of good and bad solvents then the Sphere calculation does the rest.
With all the scores entered, 1 for good and 0 for bad, you click the Calculate button and you get two vital bits of
data, both of which are important. The first is the HSP of the polymer. That’s defined as the centre of the
Sphere. Any solvent close to that centre will be excellent for that polymer. But how close is “close”? That’s
why you need the second bit of data which is the radius of the Sphere. You’ll remember that RED number is the
ratio of the distance of a solvent from the centre of the Sphere, divided by the radius of the Sphere. If your
polymer gives a small radius, say, 4, then a solvent with a distance of 4 is just on the boundary – the RED is
4/4=1. A solvent with a distance of 8 is therefore a bad solvent, with a RED of 8/4=2. But if your polymer is
more forgiving, then the radius might well be 8. So the solvent with distance 4 now has a RED of 4/8=0.5
which means that it’s likely to be pretty good. The solvent with the distance of 8 now has a RED of 8/8=1 so is
borderline.
The previous paragraph is so important that you need to read it again till you’re 100% clear. Both the centre
and the radius of the Sphere are vital for you to know. Later on, when you understand the Polymers form in
HSPiP you’ll find some tricks for working out which solvents (usually from a list that is different from the one
you used to determine the Sphere) are good or bad. By adding your polymer (both HSP and Radius) to the
Polymers form with one click you’ll be able to sort your solvent list in order of RED – with the low REDs
being the good solvents.

20
Here’s an example (included in HSPiP) where the Sphere comes out best matched to chloroform. You can
instantly tell that it’s the best match because the solvents are sorted by their RED numbers and chloroform has
the smallest RED. In later chapters we will see how you can find a solvent blend that would be a close match to
chloroform and the centre of the Sphere, but without the (probably) unacceptable use of chlorinated solvents.

Figure 3-1 A typical first Sphere test using a typical, sensible range of solvents
Bad Spheres
There are two types of bad Spheres that come out of such experiments.
1 The first is a Sphere with (approximately) the same values each time you click the Calculate button, but with
an appallingly bad fit – with many good solvents outside and bad solvents inside. In general there are three
possible reasons for this
• You’ve made some misinterpretations of your data. It’s quite common, for example, to go back to
recheck a “good” result which doesn’t fit the Sphere, only to find that the polymer is stuck underneath
the lid of the test-tube rather than, as we thought, being nicely dissolved. When you correctly class this
as “bad” the Sphere fit greatly improves
• HSP just don’t apply to your system. This is possible, but unusual. Pigments with highly charged
surfaces may fall into this category
• Your test material is impure. In fact we have often proved that materials are impure by finding bad HSP
Spheres and alerting researchers to look out for (and find) impurities
• Your material has a split personality – e.g. a 1:1 block copolymer made from hydrophilic and
hydrophobic parts. For these cases you need to use the Double Sphere method (first introduced in v3.1)
2 The second problem shows up as a Sphere which appears in a different place each time you click the
Calculate button. This is simply a case where you don’t have enough good/bad solvents surrounding the centre
of the Sphere to create an unambiguous best fit. For example, if the real δD is ≥19, the chances are that you’ve
used very few solvents up in this range. There is no way to know if the real δD is 19 or 20 or 21 because there
are no δD=21 solvents giving you a “good” or “bad” result. The only way you can pin down the real Sphere is
to work out (it should be obvious looking at the 3D plot and the 3 2D plots) where you are lacking data. Then
simply find a few relevant solvents and do the tests. We mentioned earlier that sometimes you have to use nasty
solvents in tests. You may well be forced to use such solvents to get data from a relevant region. As long as you
21
can handle the solvent safely for the test, it doesn’t matter that it too nasty to be used in a real application.
Remember that you don’t have to re-do any of your previous tests. You simply add the new data to your list and
click the Calculate button. HSPiP includes the Sphere Radius Check option (discussed below) that provides
some potential alternative solvents for you to try.
This still doesn’t explain why the “best fit” moves around so alarmingly. This is for two reasons – one
mathematical, one philosophical
HSP 3D space is quite complex. If you could test each point in that space to see if it’s the best (or least bad) fit
to the data you would find in the case of really good data sets that there is a clear, deep well down into which
any good fitting algorithm would quickly fall, finding the same result each time it started, from wherever it
started. With under-specified datasets, there are no such deep wells and it’s very easy to fall into a shallow well,
thinking that it’s the best. So the “best” fit depends on your starting assumptions. This is typical mathematical
behaviour.
Philosophically we could make ourselves look good by always leading you to the same false fit, however often
you clicked the Calculate button. But we deliberately want you to see when the fit is poor. The fact that you see
a different “best” fit each time is telling you clearly that the dataset is under-specified and that you need to
gather more data points if you really want to know the centre and radius of the Sphere. Each time you click the
Calculate button, the fitting algorithm starts from a totally different part of HSP space and therefore, in the case
of under-specified data, is likely to end up in a different “best” fit each time.
However, for the 3rd Edition we’ve been able to find a better way to search the whole 3D space and the
“jumping around” problem has been much reduced. In addition, the Core calculation gives an estimate of how
“sharp” the fit is. A narrow set of core values for δD, δP and δH means that small variations in each gives a
large increase in the error function, so the fit is sharp. A broad set means that large variations cause little change
in the fit, so the fit is rather imprecise. Often just one values has a wide core so it’s worth finding a solvent that
could impose a greater penalty for straying one way or another in that direction – again using the Sphere Range
Check as a way to come up with ideas for which extra solvent(s) to use.
Changing the fitting algorithm
The “classic” Hansen fitting algorithm has been used successfully for over 40 years. It systematically explores
the whole of HSP space and weighs the errors of good-out and bad-in depending on how badly out(in) a solvent
actually is. This makes a lot of sense. But one very wrong point can exert a disproportionate effect over the fit.
We’ve therefore provided the option of a totally different way of identifying and finding a best fit. The GA
(Genetic Algorithm) attempts to find the least wrong solvents within the smallest possible radius. The search is
via the genetic approach using “genes” that create the fit and selecting better genes throughout the generations.
Not surprisingly, the two approaches often reach a near-identical result and the classic approach tends to be
faster than the GA algorithm in these cases. But in other situations (e.g. with odd outliers) the GA result seems
more intuitively satisfactory than the classic approach. You, as a scientist, can reach your own judgement for
your own cases. The SphereFit Power Tool gives a bigger range of fitting options (including non-Spherical fits)
for those who want more choices.
Changing the definition of “good”
A lot of people are worried about the flexible definition of “good”. It doesn’t sound precise enough to be good
science. So as an exercise, take the above sample file and deliberately set two of the solvents near the edge of
the Sphere (RED ≥0.98) to “bad”. This could happen, for example, if a colleague looked at your test tubes and
said “I think you are being too generous in your evaluations. I don’t think that dioxane and TCE are good – they
should be classed as bad”. When you calculate the Sphere you get new values [17.10, 3.07, 5.10, 4.10].
The changes in this case are typical of what you find in real-world examples. The centre of the Sphere does not
change all that much, but the radius changes, in this case by quite a lot (from ~8 to ~4). This is an important
result. The key “truth” about this polymer resides in the HSP. These don’t change much when you change your
definition of borderline good/bad. The radius is an important part of the HSP characterisation of a polymer, but
it must be variable – for example a lower molecular weight version of the same polymer will be more soluble
and have a larger radius.
Please do try this playing around with the data. If you eliminate more and more solvents then, of course, the
Sphere will start to move around more. You can’t be too careless about the definition of good. And when you
22
are down to just a few good solvents, the data become statistically less satisfactory. If you find yourself in this
situation you have no other choice than to find a few more solvents in the critical region and testing those for
good/bad. If you find one or two more good solvents then your confidence in the HSP of the polymer will
improve.
Rational improvement
One user of HSPiP asked if we could add a feature to help rationally improve the quality of the fit. What he
wanted was an automatic scan of the 3D fit to identify “diversity voids”, i.e. areas in 3D space where there is no
relevant data. For HSP, “relevant” means ‘close to the edge of the Sphere and in an area where there are no
other solvents’. Why is this important? If you test extra solvents that are either near the centre of the Sphere or
near another solvent or are far outside the Sphere, you get very little extra information. But if you find a solvent
near the edge, and in a direction where there are no other solvents then this one extra data point could be crucial
for defining the edge.
We’ve found a relatively simple way to implement this process. A file called SphereCheckMaster.hsd contains
a list of 80+ solvents that fill a good part of 3D space. These solvents are a subset of the original Hansen 88
solvents so their values are well-validated. When the Sphere fit has taken place, the software scans each of the
solvents to see if it is near the edge of the Sphere (i.e. 0.9<RED<1.1) and if it is whether it is far enough away
from the other solvents used for the fitting. If both those tests are passed then the solvent is shown in the plots
in orange and added to the table of solvents. As an intelligent user you can then make the judgement as to
whether any or all of those solvents would really be worth testing. If you are happy with “good enough” then
you may not even want to do this Sphere Radius Check (SRC). If you do the check you might disagree with the
program’s judgement on some of the solvents but might agree that one or two of the solvents might be worth
doing. You might then find that one of these solvents is unacceptable in your lab. At that point you can
highlight that solvent and go into Solvent Optimizer to find if there is a blend of nicer solvents that could be
used instead.

Figure 3-2 12 solvents (in orange) suggested by the SRC check.


In summary, the SRC is a tool that you are free to use to help you decide on whether it’s worth doing more
experiments. It’s certainly not a perfect tool (there are probably more sophisticated algorithms that could be
used) but we find it helpful so we hope that you will too.
Using numeric data for greater accuracy

23
If you can go to the trouble of measuring actual solubility of the polymer rather than Yes/No solubility then you
have richer data and the chance to get a more accurate fit. The Data fit method in HSPiP will, therefore, provide
a more accurate value, especially if fewer solvents are used. Clearly there is a trade-off: there is more work
required to measure the solubilities, but less solvents (if chosen carefully) are required.
An alternative numeric approach uses measurements of the intrinsic viscosities of the polymer in the solvents.
The idea behind this is that the better the solvent, the higher the viscosity (because the polymer opens out
more). The equations for fitting to intrinsic viscosity data are simple as discussed in, for example, Farhad
Gharagheizi, Mahmood Torabi Angajia, A New Improved Method for Estimating Hansen Solubility Parameters
of Polymers, Journal of Macromolecular Science, Part B: Physics, 45:285–290, 2006. If HSPiP users would like
us to implement those equations we would be happy to do so.
Remember, however, that the use of limited numbers of solvents comes with the danger of not encompassing
sufficient of HSP space. Fitting algorithms can always find an optimum, but if there are no solubility data in
important parts of HSP space then these optima will be false. Many polymers have one or more of the HSP
higher than any potential test solvent. Simple averaging the HSP of the test solvents, weighted in any way,
cannot lead to the correct answer. The sphere fitting process is required. The intrinsic viscosity inherently limits
a study to those solvents that do dissolve the solute. There are in essence no “bad” solvents for measurement of
the intrinsic viscosity, but ranking the intrinsic viscosity into the standard 1-6 ratings, possibly with a 0 for
those solvents known not to dissolve the solute, will allow an analysis in the usual way.
The OB Fit (Optimal Binary)
A team led by Prof Dietmar Lerche from LUM GmbH and Prof Doris Segets working at Friedrich-Alexander-
Universität Erlangen-Nürnberg and now at U Duisberg-Essen came up with an ingenious objective fitting
technique for numeric data.
Their starting point was the LUMiSizer, an analytical centrifuge that allows you to measure the sedimentation
times for (nano)particles dispersed into different solvents. A sensitive light detector follows how the sediment
moves down each of the tubes in the centrifuge. The optical density, OD, starts as equal (orange line). For a
good solvent (green line) it falls slowly near the top of the tube and rises slowly near the bottom. For a bad
solvent, the OD falls quickly near the top and rises rapidly near the bottom.

By defining a cut-off value for the OD in, say, the middle of the tube, it is possible to measure a sedimentation
time, ST, for each sample. When corrected for viscosity and density you get a set of relative sedimentation time
values, RST. More sophisticated analyses can include different accelerations (for a wider dynamic range) so the
RST should also be corrected for those effects.
By looking at the data we could say anything with an RST greater than a given value is a Good solvent and
everything else is Bad. But the team came up with an objective criterion to take away human estimation.
Load the RST data into HSPiP and sort them from high to low. Now assume that the first two solvents are Good
and calculate an HSP by the standard Sphere technique. Next, assume that the first three are good and get a
different HSP. Continue till all but one solvents are classed as Good (even though we know that many of them
will be Bad!).

24
What we’d expect is that there would be relatively large changes (calculated as HSP Distance) between the
estimated HSP as we go from 2 to 3 to … but at some point the difference from adding an extra genuinely good
solvent will be small. If we then start to add a genuinely bad solvent, the difference from the preceding value
should start to be large.
Arguably the point of minimum ΔD, the HSP Distance between consecutive solvents, is the Optimal Binary
point; this marks the end of the Good solvents and all subsequent ones are Bad.
HSPiP (with the authors’ kind permission) has now added the OB Fit algorithm. A graph plots the consecutive
Distance values so you get a good idea of what is going on. A table shows you the calculated HSP values at
each number of “good” solvents along with the ΔD values.
A “Split High/Low” option allows you to make your own choice if you wish to override the automated value.
Although the OB Fit technique was developed for data from the LUMiSizer, the technique works generally for
any non-binary dataset such as solubility or swellability data.

25
Chapter 4 The Grid (A different route to the Sphere)
For more than 40 years the Sphere technique has been used with great success. But sometimes you don’t quite
have the solvent you need so you add one or two pseudo-solvents created via the mixing rule. So if you don’t
want to use dimethyl formamide in a test you can create something close (using the Solvent Optimizer) with,
say, a 60:40 mix of DMSO and THFA. This technique has been used numerous times.
But why stop there? Why not make a whole Sphere test from mixed solvents? If, for example, 3 pairs of
solvents covered the range of interest then instead of finding 20-30 different solvents, you can stock just 6 of
them, and if you have a robot it’s rather efficient to make up the set of possible blends from the 3 pairs of those
6 solvents.
Although others may have done this before, the Brabec group at U. Erlangen have shown how powerful the
technique can be in a paper that tried to narrow down the HSPs of materials used in organic photovoltaic
systems: Determination of the P3HT:PCBM solubility parameters via a binary solvent gradient method:
Impact of solubility on the photovoltaic performance, Florian Machui, Stefan Langner, Xiangdong Zhu, Steven
Abbott, Christoph J.Brabec, Solar Energy Materials & Solar Cells 100 (2012) 138–146.
The technique seemed so powerful that for HSPiP it was termed the Grid method (because it assembles a grid
of points throughout the 3D space). The initial version of the Grid was highly experimental but rapidly proved
so popular that it has been expanded, with the ability to create a Grid automatically from a list of chosen
solvents.
An example using the Default set is shown:

Figure 4-1 A Grid based on the Default set


In this example 4 pairs of solvents are chosen: Acetonitrile:Cyclohexanol; Propylene Carbonate:Ethanol;
DMSO:Toluene; Tetrachloroethylene:Butanol. Abbreviations are used because, as can be seen from the main
form, the individual points are made up of the step-wise variation of the pairs and their full names would over-
fill the space for the names.
This particular grid is chosen because it rather effectively fills a large part of HSP space. But of course it is not
perfect. There are some significant holes in that space. But the point of the Grid is that it is entirely flexible.
You can fill those holes by changing the pairs or you can add a few specific single solvents. It is your choice.

26
This sort of Grid is general purpose. An alternative, which is equally powerful (and is the choice in the
Erlangen paper) is to start with a single good solvent and explore the Sphere in 3 directions.

Figure 4-2 A Grid using a single solvent (hopefully) at the centre of the Sphere
Here the idea is that the material of interest is soluble in DBE and by scanning out in 3 different directions the
radius and (probably) the centre can be defined more precisely than tends to happen with the usual assortment
of solvents.
One user loved the idea but wanted to be able to create Grids automatically from a chosen set of solvents which
happen to be the only ones usable (for cost, tox, volatility reasons) in their area. So the AutoGrid method takes
a selected list of solvents (from the Solvent Optimizer) and a Target Zone of interest and attempts to create a
suitable Grid. Finding a perfect algorithm for this has proven difficult, but it seems that the choice of solvents is
even more important than the choice of algorithm. If, for example, Hexane:MEK turns out to be a good scan
line within the Grid, then Heptane:MEK would also be a good scan line. If the solvent list contained both
Hexane and Heptane then the chances are that both lines would be included in the Grid (because they are both
“good”), adding nothing whatsoever to the quality of the resulting data. Although in principle a smart algorithm
can remove such pseudo-duplicates, given that humans are smarter than computers it’s a good idea just to
include heptane as it’s usually rated less toxic than hexane.

27
Figure 4-3 An AutoGrid aimed at the [17, 12, 6] zone
This particular AutoGrid is limited to 32 experiments, so at 5 steps/scan-line, that means 6 pairs of solvents.
The “Centered” option was chosen to force the lines to cross the sphere as much as possible. No doubt a human
could come up with a better Grid from the given set of solvents, but it would take much longer and at least the
automated version offers a starting point for human intervention.
A question that is often asked is: “How good are mixed solvents compared to the same HSP from a single
solvent?” The answer from experiments carried out over 40 years is “In HSP terms they work very well, but in
kinetic terms, and in situations where entropic effects are important they act as if they were a solvent with a
larger MVol”. So if in the example at the start DMF was a fast solvent for polymers, the DMSO:THFA mix
would be rather slower. And because entropic effects are important in polymer solubility, the mix would not
give as high a solubility as the single solvent.
This means that the Sphere radius from mixed solvents will be rather smaller than if it had been measured with
single solvents. For most users, the ease and convenience of the Grid and the chance of a more accurate Sphere
centre seem to outweigh the issue of a potentially smaller radius.
One criticism of HSP is that it does not provide a reliable tool for calculating solvent miscibility. It is therefore
possible to generate grids that are impossible because one or more chosen pairs are immiscible. Although
HSPiP now includes a miscibility predictor, users have to apply their own knowledge and intuitions when
setting up grids because miscibility predictions defy even the best available solvent tools.

28
Chapter 5 FAQ – The Radius and other questions
1 The Radius
The most common question asked over many years of use of HSPiP is something like:

“How can I calculate the Radius of my solubility sphere?”

There are many variations on the question. For example


• “The RED of a theta solvent must be 1, but for my polymer the RED of a known theta solvent is 0.9, so
what is wrong?”
• “I see that the radius for my polymer is 6 but I have a solvent whose distance is 4 and it doesn’t dissolve
the polymer – what’s happening?”
• “Why don’t you quote the radii of solvents?”
• “How is it possible for CNT to have a radius when they aren’t even soluble?”

So let’s make the point very clearly: HSP values are thermodynamics, Radius is the scientist’s judgement.

This means that for the same polymer (but see below), everyone will agree on the HSP values, but there will be
no agreement on the Radius, because it depends on the choice of the individual scientist.

At first, this sounds very unscientific. Values aren’t supposed to depend on the scientist, they are supposed to
be universal. Yet the fact that you get to choose the Radius in your application is good news, not bad news.

Imagine that your application requires 10% solutions of a polymer of 1,000,000 MWt. As you do the HSP
Sphere measurements you will probably find at most 1 solvent. So your Radius will be approximately 0.

Or imagine that your application requires 0.001% solutions of the same type of polymer but with a MWt of
1,000. You might find the Radius to be something like 20 – in fact, finding a bad solvent will be quite difficult.

Both extremes are probably stupid – but the point is that you as the scientist get to make the choice. You know
why you need the HSP values and so the definition of the Radius is meaningful to your application. Your
definition may be completely irrelevant to someone who wants 1% solutions of 100,000 MWt of the same
polymer, or 1% solutions of 100,000 MWt of a high crystallinity, or 1% solutions of 100,000 MWt of a lightly
cross-linked version of the polymer.

The mention of cross-linking reminds us that doing “solubility” tests is meaningless if things aren’t soluble. So
the Radius of a rubber depends on your judgement of swelling – and that depends on the application. If you
really want to process a rubber in a solvent then your definition of “good” will be a large swelling, so the
Radius will be comparatively small. But if, for exactly the same rubber, you are worried about long-term
damage in a joint, then small amounts of swelling will be seen as “good” (in terms of solubility, though bad in
terms of your application) so the Radius will be large.

And although one can argue that nanoparticles can really be seen as soluble, many people will choose to
measure the HSP by looking at the “happiness” of CNT, quantum dots, nanoclays etc. in the different solvents.
“Happiness” may be visual (“It looks nicely dispersed and isn’t settling out too quickly”) or it may be measured
(“Resists centrifugation for 10s at 10,000 rpm”), but it is your scientific decision and therefore your Radius.

But let’s not get too unscientific. Under good, controlled conditions the HSP distance for a theta solvent is
precisely defined. In the Polymer Theory form that you select from Polymers you can see how the free energy
of solution changes with mole fraction of polymer and how it can flip from soluble to insoluble with just a
small change in distance or, its equivalent, small change in Flory-Huggins Χ parameter. Ultimately things come
down to entropic effects and the effects of the relative MWt of the polymer and MVol of the solvent can be
calculated precisely. It is from these sorts of calculations that the idea of Radius=8 being at the border of good
and bad solvents for “typical” polymer MWts and “typical” solvents. This is a useful rule of thumb, but it

29
applies only to high-quality polymers in careful scientific environments – which is not what concerns most of
us most of the time in our formulation work.

Going back to the issue of the “same” polymer. Yes, if it is a relatively simple polymer, relatively pure,
everyone will agree on its HSP values and disagree only on the radius. But there are many ways that the “same”
polymer can be different. For example:
• The ratio of head-to-head monomers to head-to-tail might be different. In PVdF, for example, –
CH2CF2–CF2CH2– units have very different dipole moments (and therefore different δP values) from
the normal –CH2CF2–CH2CF2– so they show very different solubilities. So the behaviour of different
PVdFs will be very different.
• If you look at swelling of polymers, a side-chain (e.g. a hydrocarbon) might allow absorption of small
amounts of a solvent so give the impression of one type of HSP at low % swelling, but the main chain
will not be affected by such solvents and as you go to measurements with higher % swelling, the centre
of the Sphere will change to represent the true value of the polymer main chain.
• Polymers such as CAB (Cellulose Acetate Butyrate) represent “typical” levels of substitution of the
cellulose hydroxyl groups by acetates and butyrates. But celluloses are so complex that it is highly
unlikely that one CAB will give the same HSP as another CAB.

The conclusion of all this is that for your own real work on your own real polymer system, there is no substitute
for doing your own measurements and interpreting the results using your own scientific insights. Although
HSPiP cannot magically solve the complexities of real polymers, it at least provides a rational way to think
through the issues and provide relevant measurements for you to use.

2 The Fit
The second most common question is “I’ve just done a Sphere fit – how accurate are the values?”

The answer to the question is similar to the answer to the Radius question – although the software can give you
some “fit” numbers, only you as a scientist can form a judgement as to how good the fit really is.

The first number supplied by the software is the Fit. If all good solvents are inside and all bad are outside, the
Fit=1. If you have lots of wrong values then the Fit is small. But it’s still possible (though unlikely) that the
second is better than the first. Suppose, for example, that you did a fit using hexane, heptane, octane, decane
and ethanol, propanol, butanol (we’ll talk about methanol in a moment). You will probably get a Fit=1, but the
value will be useless as you’ve not explored much of HSP space. In the CNT example later in this book, the Fit
is not very good – which isn’t surprising given the complexity of CNT, but the insights provided by the fitted
value are very powerful.

The second numbers from the software is the Core values for δD, δP and δH. These are calculated by
deliberately moving the centre and seeing how far you can move it before the Fit value decreases rapidly. If you
have a very good fit then the Core values will be small (e.g. ±0.25-0.5) but if you have a very bad fit (e.g. from
the example above) then the Core values might change by ±1. Sometimes you find a tight Core for two of the
parameters and a loose Core for the third – meaning that you don’t have enough solvents in that dimension to
make much difference across the fit.

So much for numbers. Now onto judgement. We provide three general fitting options: Classic (the default), GA
and Double Sphere. If you are fitting a block co-polymer then the fits with Classic and GA will be poor, but a
Double Sphere might be excellent. But if we focus just on Classic and GA what do we find? With very good
data sets, the results from Classic and GA are the same. So one way to check quality is to compare the two fits.

But what happens if Classic gives a different result from GA – as can often happen? There’s a very good reason
that we plot the Sphere in 3D. You, as the intelligent scientist, can see what the different Sphere fits are doing.
Usually it’s quite obvious that one method is producing a result that doesn’t make sense to you. Like all fitting
routines, the methods have to impose a penalty for “good” being out and “bad” being in. The two algorithms
impose different penalties. Arguably, for good data the Classic method is the better one. But the GA method is

30
often more “sensible” in its choices. Classic tries much harder to bring in “good out” solvents than GA. GA is
often more sensible in forgetting about some wrong datapoint. Which brings us to the subject of methanol.

It is universally noted that methanol can give higher solubilities/permeabilities than would otherwise be
expected (see Prof Darren Williams’ wonderful paint-stripping video based on HSPiP
http://www.youtube.com/watch?v=2lgRZOQmZKU). Whatever the explanation for this (perhaps it forms tight
clusters that present a hydrophobic exterior) the Classic method might try too hard to include an anomalous
methanol result and the GA will more easily ignore it.

You might find it unsettling that different fitting algorithms give different results. But to repeat – with good,
simple systems the two methods give comparable results, it’s when there are complications outside simple
theory that the assumptions behind the fitting algorithms give different results. This is good news – it allows
you as the scientist to form your own judgement about which is the more realistic fit for your specific system.

And why rely on algorithms? If you think that methanol, or some other solvent, is giving an anomalous result
for some special reason, you can set its value to “-“ in the table and try the fit without it. If the fit doesn’t
change much then you know that the anomaly is not important. If the fit changes a lot, and if the quality (Fit
and Core) look much better, then it is up to you as a scientist to decide whether ignoring that data point is
justified or not.

Most of us have found examples where the scientific judgement has told us that a specific point in the Sphere fit
must be wrong and when we go back to the test samples we find that for some reason we’d mis-recorded the
value. The classic error is a wrong “good” value, only to find the blob of polymer stuck under the cap of the
tube and not, as we thought, beautifully dissolved in the solvent.

There’s another area of judgement. If you score the tubes on the 1-6 scale (remember, this is an historical
scheme where 1 is good and 6 is bad) then you can choose to find a fit based on just the 1 solvents, or the 1 and
2 solvents and so forth. Where do you stop? If you have a really good system then the only thing that changes is
the Radius – reminding us again that Radius is based on scientific judgement. But often you find that the HSP
values are similar for the 1’s and 2’s but very different for the 3’s. Again, it is for you to judge which values (of
HSP and Radius) to use for your specific purposes.

We provide another option to help you refine your fit. Testing with another 10 solvents that are near the centre
and another 10 solvents that are far from the centre adds little extra information. But testing with 1 or 2 solvents
that are near the border can make a very big difference in the quality of the fit. So the trick with Sphere testing
is to only use solvents near the border. When you start your measurements you don’t know the centre and the
radius, so you can’t use just border solvents as you have no idea where the border is. But if you’ve got a Sphere
and you are unhappy/uncertain about the fit, all you have to do is test with a few solvents that are near the edge.
You can do this using your judgement. Or you can select the SRC (Solvent Range Check) option and click the
fit button. This looks through a list of ~80 solvents and finds ones that aren’t in your current solvents and which
are near the edge of your fit. By testing with one or two of those solvents (depending on availability, safety,
aroma etc.) you will greatly increase the quality of your fit.

A variation on the above is to produce a very poor quality fit very quickly, then to refine it by checking for
solvents near the radius. The Grid method allows you to do this easily. If you know one really good solvent for
your system (and most of us do) and if you know a couple of bad ones, then a quick Grid scan is guaranteed to
give you a centre and a radius. They won’t be great values, but very quickly you have an idea of where the
radius might be. It’s then easy to do more tests either with a new Grid scan or with some specific solvents, or
both. This “hybrid” method is becoming increasingly popular in the HSPiP world.

In summary, scientists are smarter than computer programs. Nothing we can do will provide absolute proof that
your fit is perfect. Only you can use the available tools (Classic v GA, 1 v 2, include/exclude solvents, SRC…)
to reach a judgement on the quality of the fitted values.

3 Why don’t you let us calculate Y-MB on very large molecules?


31
In the eBook we haven’t reached the topic of automatically estimating HSP values via the Y-MB method. But
most readers will know about it and have probably tried it out already. And this is such a popular FAQ that it
makes sense to answer it now.

The current limit on Y-MB is 120 “heavy” atoms, i.e. atoms other than H. Why don’t we increase it so, for
example, a user can calculate the HSP of a large polymer with lots of side chains via an enormous SMILES
string?

Part of the answer is practical – the more atoms the slower it becomes to work out what fragments are in the
molecule. But mostly the answer is philosophical: it’s probably not scientifically worthwhile to estimate Y-MB
on such large molecules. At first this seems a shocking statement. We seem to be saying that Y-MB is no good
for large molecules. There is a good reason for this. Imagine, as a simple example, a di-block molecule, one
half of which is mostly hydrocarbon and the other half mostly oxygenated – such as a PE:PEO di-block. Y-MB
will give a value that is approximately half of PE and PEO values. This is “correct” but useless. PE:PEO does
not behave as an average molecules – it behaves as two very different molecules joined together. This is an
extreme example. But imagine a complex pharmaceutical such as a steroid. The large steroid part might be
largely hydrophobic and there might be a large complex side chain with lots of sugars on it which are very
hydrophilic. The Y-MB will give an average which certainly does not capture the science of that molecule.

So what is the answer? In principle Y-MB could automatically split a molecule into two regions, calculate the
two values and provide them plus the average. However, although to a human it might be obvious where to split
a molecule into two different regions, this is a near-impossible task for a computer. So the answer is the same as
the other answers in this section: scientists are smarter than the computer so should do the split themselves.
Take your molecule into your favourite chemical structure software, break it into two (finding an appropriate
functionality at the splitting point), ask the software to produce the SMILES for each structure, calculate the Y-
MB of each structure and then reach your own conclusion about what will happen in the combined molecule.

The problem is finding an appropriate functionality at the splitting point. Suppose the two halves are joined via
an ester link. You could split them so that one half ends with an –OH and the other half ends with –CO2H. But
this would introduce two –OH functionalities and raise the δH value unrealistically. In this case it’s probably
better to end one half with –Ome and the other half with –CO2Me. This is now introducing one ether and one
ester function to replace one ester function. It’s not perfect, but it’s better than splitting into the alcohol and
acid. For small molecules the error of introducing the ether is significant, but for small molecules Y-MB is OK.
For large molecules that require this sort of splitting, the ether error is probably insignificant. Or, if you as the
scientist think that it is significant, then find a different way to create the split.

4 Salts, transition metal complexes etc.


Users would love it if they could enter the SMILES for salts or for transition metal complexes or ionic liquids
etc. and get good Y-MB values. This is impossible.
There is a whole chapter in this book on ionic liquids so they won’t be discussed here. But the general point is
that although it is perfectly sensible to talk about the HSP values of salts or transition metal complexes, the
reason we can’t predict them is that there is no data set containing the HSP values of thousands of such
molecules. So Y-MB cannot be “trained” on those data, so it is not possible to find values for the, say, Na+
functional group.
So if you are working on, say, organic electronics (OPV, OLEDs etc.) it is frustrating that you cannot rely on
Y-MB estimates for your molecules. But remember that the “truth” in HSP comes from experimental data. If
you set up your systems correctly, it is very easy to measure a large number of HSP values, especially if they all
happen to be in one part of HSP space so your solvent choice and/or Grid choice gives you the most data for the
fewest experiment. If you are not set up to do such measurements routinely, then at least in Europe there are
two companies who offer high-throughput measurement services for HSP values, and hopefully other service
providers will appear in other parts of the world.
While we are talking about impossible SMILES, another question is about “dotted SMILES”. As a simplistic
example, ammonium acetate might be shown as N.CC(=O)O – with the all-important “dot” between the
32
SMILES for ammonia and for acetic acid. Although in principle we could provide the individual HSP
estimates, we have chosen to not allow dotted structures because they raise so many questions. What use is
there, for example, in providing you with the HSP of ammonia and acetic acid if the molecule is really
ammonium acetate?
5 The problems of water (and methanol)
Everyone agrees that water is a strange solvent. It has an exceptionally low MVol, so its kinetics and its
thermodynamics (entropic effects) are both exceptional. And its hydrogen bonding capabilities are also
exceptional. So most solvent theories have problems with water. For HSP it’s even worse – the HSP of water is
very far from everything else (δH=42) so HSP distances from water are usually very large – “proving” that
water will be a bad solvent for most things. But water is often an excellent solvent, so HSP must, therefore, be
wrong.
Our view is that in many cases (especially salts!) you should forget about applying HSP to purely aqueous
systems. But this doesn’t mean that HSP forgets about water.
First, adding a few % water in a solvent blend can often produce dramatic changes in solubility. HSP
sometimes does a very good job in explaining this within the Solvent Optimizer.
Second, as discussed in the DNA-related chapters in this eBook, there are cases where you can do very smart
things in largely aqueous systems if you are careful.
Third, you don’t always have to use the extreme value of water. In the database there are two other values for
water. These are both experimental and are therefore both “true”, just as the standard value is clearly true
thermodynamically. So how can three different sets of values all be true? It’s because when water is a relatively
small fraction of a solvent blend it no longer behaves like bulk water. Presumably it self-associates more and
masks a lot of its extreme δH characteristics. When it is in a minority, it behaves with a δH ~ 16 and some of
the high δH has been changed into a higher δP. This makes intuitive sense.
But when should you use which values? The answer is the same as the rest of this FAQ – only the scientist can
judge under the given circumstances.
As mentioned above, methanol seems often to be highly anomalous. It can readily, for example, diffuse through
a highly-fluorinated membrane when ethanol is blocked. We have therefore added a representative value for
“clustered” methanol which has been tested against a variety of datasets (including 2 “anomalous” fits in the
Polymer88 files) with good results. Why the word “clustered”? The intuition is that the methanol molecules
self-associate and therefore hide much of the δH within these self-associated groups. We have no evidence for
this but simply propose it as a working hypothesis. The value for this form of methanol is [14.7, 5, 10]. Because
the clustering (or whatever the effect is) itself will depend strongly on the system, you should use your own
wisdom as to how seriously to take these values. For example, data on the solubility of long-chain carboxylic
acids is best fit to [14.5, 3.1, 5.6] which we think is the low extreme with the standard value of [14.5, 12.3,
22.3] as the other extreme. Note that the lowest value happens to divide the δP and δH by 4. Does this indicate
a tetramer? Others have certainly spoken of methanol tetramers. What MVol should be used for “clustered
methanol”? The tetramer would be 160 but the fitting data for the carboxylic acids suggests 120. In the database
we have used a value of 100 just to signal that this is an open issue. We will be interested in any comments on
these issues.
6 Miscibility
It frustrates HSPiP users when the SO suggests and interesting solvent blend – which then turns out to be
immiscuble. The obvious question is:
“Why can’t HSP distance be used to calculate miscibility?”
The short answer is that miscibility is astonishingly difficult to calculate. It seems that very subtle interactions
can flip a blend from miscible to immiscible if it is in some borderline region. The classic Wilson Parameter
approach to activity coefficients is famously unable to predict immiscibility – no matter how large the activity
coefficients (which imply that the solvents really dislike each other) there is no thermodynamic way for Wilson
Parameters to suggest immiscibility.

33
But for v4.1 we decided to do our best to provide at least a reasonable estimate of miscibility. This turned out to
be a difficult task. Assembling a dataset of ~900 solvent pairs (based on 70+ solvents) took considerable effort
because much of the original data behind the miscibility data (e.g. the famous Godfrey dataset) seems to be
unavailable and tables of implausible certainty have been put in place with no way to check them from the
original data. Furthermore, what were we supposed to do with “borderline” cases. Should these be classed as
miscible or immiscible or should we aim to include “borderline” in our classification. In the end we decided to
define “borderline” as immiscible because experience shows that it doesn’t take much (like the presence of a
small contamination or a solute) to flip “borderline” into immiscible.
Our best efforts came up with a correlation showing a 90% success rate at predicting the results in the dataset.
The default HSP “prediction” that everything is miscible gives a 70% success rate – meaning that the dataset
was imbalanced in favour of miscible pairs. A sophisticated thermodynamics package predicted ~80%, though
with some refinements it could probably have done better. We could push the 90% to 92% by adding extra
fitting parameters, but this gives small gains for probably spurious reasons. To get 90% we included a
donor/acceptor cross term for δH. If we omit that term we get an 88% fit. In the software the results of both fits,
plus (if available) the result from the dataset is provided in the Miscibility form in DIY and the 90% fit is used
in the quick Miscibility Check within SO. What is interesting is that the best fit required the classic 4:1:1
Distance² term plus a MVol1 term (the sum of the two MVols). So the essentials of HSP (MVol is always
related to Distance² in HSP) are retained even though the fitting algorithm was given freedom to simply get the
best fit. So although the correlation is far from perfect, it is not entirely arbitrary.
If users provide us with more data on miscible and immiscible solvent pairs then we can refine the algorithm
for future releases.
FAQ in summary
HSP is a pragmatic, practical tool with a strong theoretical underpinning but with clear and obvious limitations.
Given those limitations you can decide to use alternative solubility prediction tools (and we sometimes use
those ourselves) or you can use your scientific judgement to work within those limitations and reach valid
formulation decisions. To an astonishing degree, blending HSP (HSPiP) with a good scientific brain produces
powerful formulation insights. Understanding the limitations of HSP is the beginning of using its many
strengths.

34
Chapter 6 Coming clean (Finding Good Solvents)
It’s a common problem. You have used some ink or paint and you need to clean it up. In Denmark, as in most
countries, painters usually used white spirit. It was cheap and effective. But when Denmark (a leading country
in health & environment issues) discouraged the use of white spirit because it was harmful to the painters it
became necessary to replace it with something safer.
That was the start of a process that saw many formulas having to be replaced by something even safer (to
humans or to the environment). What was the best way to do this?
One of us (Hansen) became deeply involved in this technical challenge and looks with some pride at the fact
that the formulae developed many years ago using HSP principles are the backbone of many of the
“environmentally friendlier” solvents in use around the world today.
Trial and error can be made to work, but it is slow and inefficient. It is also highly unlikely to come up with an
optimal solution. It is hard to imagine that anyone would have found the 74:26 mixture in the earlier chapter
through trial and error.
With HSP the process is clear. Find the HSP of the main polymer in the ink/paint. Find the radius of the sphere
around the central point. Then choose a solvent blend that is well inside the sphere. Because the whole process
is based on numbers it is a lot more scientific and, more importantly, a lot quicker to come up with an optimum
solution.
It is simplistic to define “optimal” simply to mean “best dissolving” but of course it should include other
factors such as cost, evaporation rate, odour etc. as discussed in the previous chapter. The point is that if you
can’t put a number to “best dissolving” you can’t include it in any other numerical optimisation.
So, let’s take our ink and test it with a bunch of solvents. As we can do this in the fume cupboard in the lab and
can wear appropriate personal protective equipment we can use a range of unusual solvents to give us the best
possible information.
How many should we use? The original Hansen work used 88 solvents because it was trying to work out a self-
consistent set of values for those solvents. What a set of ~30-40 achieves is probably all you need if you really
want an authoritative answer and if the ink is genuinely unknown. If you have a reasonable idea of the ink’s
properties you might be able to skip hydrocarbons if the ink is obviously highly hydrophilic or alcohols if it is
obviously highly hydrophobic. And there is lots of good data from people using 16-24 solvents.
The point here is to be flexible. If you are really interested in the deep science of a polymer, then ~40 is a good
idea. If you know that you won’t do a test with 40 solvents (life’s too short) then persuade yourself that 16 isn’t
too hard and then throw in another few just to make sure. Or use a Grid scan.
The really important thing is to make sure that your chosen solvents span a large range in D, P, H space. Let’s
look at an absurd example.

35
Figure 6-1 Using file Absurd
The program gives a good fit to the data. But there’s obviously something wrong with the chosen solvents. For
this absurd example we simply took the lowest 16 D values from a large list of solvents and assigned the
solvents randomly as Inside or Outside. It simply makes no sense to use this range of solvents, unless you are
certain that your polymer has a very low D value.
Now look at this plot, from Hansen’s Polymer B (a polymethylmethacrylate) using the 88 original solvents:

Figure 6-2 Using file Polymer88B


Solvents span a wide part of D, P, H space but you can see that there are clusters which over-represent some
areas and voids where there are too few. The voids, unfortunately, are nature’s fault – there just are some areas
where there are few solvents. The clusters show that 88 is too many solvents for practical work – you get little
extra information for a lot of extra work.
You see the same thing in the 3 2D plots:

36
Figure 6-3 Using file Polymer88B
So now let’s go to the “life’s too short” approach and just use 16 solvents.

Figure 6-4 Using file Chapter 3


This is a repeat of the second figure in the previous chapter. You can see that in the P v H plot the solvents
cover a good span, though a few more high-P solvents might be helpful to be sure about the boundary in P-
space and we are definitely short of high D solvents – notice all the empty space in the H v D and P v D plots.
So, is 16 enough? The good news is that you can make your own judgement after you’ve tested the 16. If it’s
really important for you to characterise this polymer then choose, say, 3 extra solvents (perhaps via the Sphere
Range Check) and do the tests. You don’t have to re-do the original 16 tests, just 3 more targeted tests.
Here are the data with the 3 extra solvents:

37
Figure 6-5 Using file Chapter4
We can now compare the HSP and Radii of the two data sets
δD δP δH Radius
Original 16 18.5 9.9 8.0 8.4
Revised 19 17.5 9.1 6.9 7.5

Table 6-1 Comparison of radii of two datasets

38
The centre has shifted slightly and the Radius is smaller. That difference of ~0.9 in the Radius could make all
the difference if you were considering using a marginal solvent.
Now that you know what the HSP of your polymer are, you will want to find a solvent with which to dissolve it
for efficient cleaning.
The program lets you sort the solvents by their RED number – with low numbers being a smaller distance from
the polymer.
Reminder: RED is Relative Energy Difference and is the ratio of the distance from the centre of the sphere for a
given solvent to the radius of the sphere.
Here are 15 of the 19 solvents used in the test:

Figure 6-6 “Inside”=1 and “Outside”=0 solvents


The first 4 solvents are quite good matches (low RED number) and considered generally not safe to use, the
fifth, acetone, is controversial: it’s “good” because in the USA it is “low VOC” but in most respects it’s “bad”
because most of it evaporates before you can use it.
But remember, for defining the Sphere for this polymer we used solvents that were technically useful (such as
the tetrabromoethane) but which we would never want to use in practice. It’s perfectly possible that none of
your test solvents would be of interest for the real world. The vital thing is to get a good idea of the polymer’s
HSP and radius. After that you need a rational way to find the best solvent for your application.
That is the topic of the next chapter.

39
Chapter 7 Safer, Faster, Cheaper (Optimizing Solvent Formulations)
With the Optimizer you can quickly home in on the HSP of your target. Let’s assume, as in Coming Clean, that
we want to dissolve a polymer in an ink. If we know that the polymer has HSP of [18.6, 10.1, 7.8] then we can
use HSP tables to find the best match. From a large list you can find that N-Acetyl Piperidine and
Hexamethylphosphoramide are excellent matches, but you are not likely to want to use either of those!
Out of your 19 solvents, the RED number shows you that N-Methyl-2-Pyrrolidone is not a bad match, but it’s
expensive, slow to evaporate and has some health & safety issues.
It looks as though we’re stuck. But remember that a solvent blend with the same parameters as the polymer is
thermodynamically identical to a pure solvent. So if we can’t find the perfect single solvent, let’s find a blend.
If you’ve ever tried doing this manually from a list of solvent HSP you will know that it’s a bit of a slog. So
let’s get the computer to do it.
The Optimizer comes with a list of “friendly” solvents – ones that you might find in general use and which
aren’t too toxic or expensive. Everyone’s definition of “friendly” differs so you should feel free to modify the
list for your own purposes.
When you enter the Target HSP (in this case [18.6, 10.1, 7.8]) you can then select one or more solvents and
their % and click the Calculate button to compute the HSP of the blend – which is simply the weighted sum of
the individual components. The Distance is also automatically calculated – the smaller it is, the better.
This is helpful for scouting purposes but it is hard to find an optimum this way.
A quick short-cut is to click the 2 button which does an exhaustive search of all possible combinations of 2
solvents to find the closest match (smallest Distance). With a little tweaking after finding the best automatic
suggestion, the following blend is a good match.

Figure 7-1 Solvent Optimizer trying to match the polymer’s [17.4, 9.0, 6.8]. The blend of THF and Propylene Carbonate is a
good match.
This will undoubtedly be a fast dissolving blend. Both molecules have a small Molar Volume (MVol) and small
means fast diffusing (kinetics) and large entropy change (thermodynamics) for good dissolution.
So we have Faster. But what about Safer? THF is generally frowned upon because of its peroxide problem so
you might not like to include it. It is often said that cyclopentyl methyl ether is an adequate, safer (no peroxide
problem) substitute. What % should be used? The simplest way to find out is to click the Optimize button. It
turns out that a 68:32 mixture is optimal. The Distance is a bit larger, but it should still be OK. The MVol is
also OK. So now we have Faster and Safer. If either of these turns out to be a bit expensive we would have to
work a bit harder to get Cheaper. By exploiting some Advanced Optimization tricks within the Optimizer it’s
possible to find that a good combination. For example, enter 50% for the ether and 25% for the carbonate. Now
click the 1 button to find the one solvent which, at 25%, will give the best match to the target. In this case it is
glycerol diacetate:

40
Figure 7-2 A blend optimized by clicking the 1 button with a 50:25 mix of the ether and the carbonate.
Of course there are always trade-offs. You may not like these solvents for various reasons. But you can start to
see how the process might work.
Let’s remind ourselves what we’ve done in such a short time. With 16 simple tests of whether a solvent
dissolved or didn’t dissolve our ink we found the HSP of the ink. We did 3 more tests just to refine the value.
Then after about 30 minutes on the computer we found a 50:25:25 mix of Cyclopentyl Methyl Ether:Propylene
Carbonate:Glycerol Carbonate as a Faster, Safer, Cheaper blend. Can you imagine how long it would have
taken you to find such a blend without HSP?
When in doubt go higher
If you had a choice of two solvents, the same distance from the target, and one is of low δTot and the other is of
high δTot, which one should you choose? Hansen’s view is that higher is better. Why? The analogy is with the
Kauri Butanol number. A “poor” solvent causes the kauri to crash out after relatively little dilution, a “good”
solvent is tolerated to a much greater extent. Because (by definition) butanol is used in the test, high δTot
solvents are likely to be more compatible with the butanol and therefore limit the crashing out of the kauri.
Armed with HSPiP one could probably find plenty of exceptions to this rule of thumb, but to the extent that the
Kauri Butanol number is of any value (and that’s debatable) the “higher is better” rule is a reasonable guide.
In the next chapter we show how HSP can make the incompatible, compatible.

41
Chapter 8 Coming together (Polymer Compatibility)
How do you get immiscible polymers working together? A good solvent for one will naturally be a bad solvent
for the other so there is a genuine difficulty in finding a way of bringing them together. For polymers which
aren’t too far apart, it’s still fairly easy to make a good guess and get a result. But what happens if they are
seriously immiscible?
A good example is when you want to combine the properties of a silicone with those of another polymer. In
order for the final system to be stable, there must be some reactive groups on both so that mutual cross-linking
ensures that the phases can’t separate. But because it needs only a relatively small amount of cross linking we
can assume that the polymer HSP are the standard, unmodified ones even though the real polymers would have
a few percent of some reactive group such as an alcohol, an amine or a methacrylate.
A typical HSP test for a real-world silicone shows values around [17.0, 1.9 , 3.2,5.8]

Figure 8-1 A real world silicone


The challenge is to compatibilise it with an epoxy [17.4, 10.5, 9, 7.9]:

The distance between these two is 10.4 so their centres are each outside each other’s sphere. These are
genuinely incompatible.
When the two polymers are entered into the Polymer table and selected, then the Solvent Match button clicked,
the following appears:

42
Figure 8-2 Silicone and Epoxy compatibilisation
You immediately find some helpful solvent suggestions. Something volatile such as Ethyl Acetate will be ideal
for generally allowing the polymers to mix, but you also tend to need some “tail” solvents to hang around
whilst the reaction is finishing so something like MIBK would seem a good idea. Because the optimal point is
shown in the output as being [17.2,5.5, 5.7], a quick visit to the Solvent Optimizer shows that cyclopentyl
methyl ether might be a good starting point.
Readers might be a bit disappointed at how easy this seems. But the authors’ experience is that without the
convenience of HSP thinking (and HSPiP) this sort of problem has involved many months of avoidable work
by major corporations.
Self organisation
You can use the same sort of ideas to do some clever self-organising coatings. Suppose you want an acrylate
polymer as the top surface of a dual-layer coating with an epoxy at the bottom. Of course you could make this a
two-pass coating. But in some applications a one-pass coating, if feasible, would save lots of time and money.
So what we would like is a spontaneous separation of the components when the solvent evaporates. But for this
to happen the solvent must be rather poor for both of them. Hence the RED number for each polymer with the
solvent should be in the 0.8-1.0 range. The polymer with the lower surface tension is expected to be at the air
surface then becoming the topcoat if motion within the film allows this. Its further accumulation at the air
surface to achieve a significant thickness is then enhanced by a reduced affinity for the other polymer, which
then forms a primer. Clearly other surface active components in the coating can interfere with this.
The Polymer form provides an Acrylate [20.7, 4.1, 10.7, 11.5] and an Epoxy [18.5, 9, 8, 9.8]. If you select
Friendly Solvents in the main form then when you select the Acrylate and click Solvent Match you find that the
RED number for Xylene is 0.87. Selecting the Epoxy and clicking Solvent Match gives a RED of 0.97 for
Xylene. Selecting both of them gives you:

43
Figure 8-3 Mutually incompatible polymers designed to phase separate during evaporation
This example is a simplified version of a real dual-layer coating. As the Xylene evaporates, the mutual
incompatibility of the two polymers becomes evident and phase separation begins. The lower surface energy of
the acrylate polymer brings it to the surface. The phase separation continues till you have an almost perfect
dual-layer. The “almost” is important. In order to preserve adhesion, there must be some intermingling of the
polymer chains at the interface. The kinetics of the system ensure that the phases don’t separate totally.
Again, this seems easy in retrospect, but if you tried to do this without the aid of HSP thinking it would take a
very long time to get even the basic functionality working correctly.
Non-solvents coming together
One of the striking and unexpected predictions from HSP is that mixtures of non-solvents are perfectly capable
of being excellent solvents. Hansen showed this back in 1967 when working on the polymer series for which
we’ve provided up-to-date correlation data in files Polymer88xx. By choosing Polymer88E you find its HSP are
[19.3 ,6.0, 10.4, 10.5]. If you check the data table you find that Diethyl ether [14.5, 2.9, 5.1] and Propylene
carbonate [20.0, 18.0, 4.1] are both non-solvents. But a 50:50 mixture [17.3, 10.5, 4.6] is inside E’s sphere and
is calculated to be, and was shown in practice by Hansen to be, a solvent. There are numerous examples of such
mixtures of non-solvents being solvents and it is important for formulators to think outside the box (or, rather,
outside the sphere) and start to get different solvency characteristics by mixing solvents they would otherwise
have totally ignored. It’s worth saying once more that the Hildebrand solubility parameter simply cannot do this
sort of thing because there is no coherent way of dealing with the issue that very different solvents can have
very similar Hildebrand parameters.

44
Chapter 9 Sticking, Flowing and Dissolving (HSP and Adhesion, Viscosity and
Dissolving)
In earlier editions, this was a short chapter. It was about adhesion and its main aim was to counter naive ideas
that adhesion was “all about surface energy”. The chapter has now expanded to bring in some serious polymer
physics of adhesion and has been extended to include polymer solution viscosity and a return to basic polymer
solubility. The logic of combining them in one chapter will hopefully become clear.
On first reading you may well want to abandon this chapter once the polymer physics starts to get difficult. But
if you come back to the chapter from time to time you’ll find that the concepts in the later sections really throw
a lot of light onto many of the issues you face when using polymers and solvents. As is so common in science,
insights from the discussion of one problem (e.g. polymer adhesion) can be helpful in an apparently unrelated
area (e.g. the disturbing ability of polymer solutions to become very viscous with small changes in conditions).
First let’s turn to the adhesion between two polymer formulations. Whether you are interested in classical
polymer-to-polymer adhesion or getting an ink to stick to a polymer surface, the considerations are the same.
Sticking together
It turns out that HSP can provide insights that help us distinguish between three types of adhesion. The first
type of adhesion, based on surface energy, is insignificant. The second, delightfully named “nail adhesion” is
quite common and (as those who prefer to screw things together already know) quite unreliable. The third is the
real adhesion for those situations where things must stay stuck together – the wing on your aeroplane or the legs
on your chair.
Real adhesion is not based on surface energies
There is a common misconception that adhesion is all about surface energies. It is well known that a polymer of
low surface energy such as polyethylene has low adhesion, but once corona treated has (a) a high surface
energy which allows an applied coating to flow out rather than bead up, and (b) good adhesion. From these
facts the conclusion is that the increased adhesion to the polyethylene is due to the high surface energy.
There is a very simple disproof of this view of adhesion. The surface energies of PVC film and PET film are
identical. Yet it is very easy to get printing inks to stick to PVC and very hard to get them to stick to PET.
There is another common half-truth about adhesion. It is well-known that roughened surfaces tend to have
stronger adhesion than smooth surfaces. The common explanation for this is that there is more surface area for
adhesion on the roughened surface. This is true, but the effect is small. A typical “rough” surface like this
probably has only 10% extra surface area:

Figure 9-1 A normal rough surface has a small increase in surface area
To double the surface area you need a surface of a dramatic roughness (a high 1:1 aspect ratio):

Figure 9-2 It takes extreme roughness like this to double the surface area and double the adhesion due to surface energy
The main reason why roughened surfaces increase adhesion has to do with crack propagation. If there is a
smooth interface and a crack appears, all its energy can be transmitted straight along the interface and the crack

45
moves along swiftly, destroying adhesion. If there is a rough interface, much of the crack energy is sent into
directions away from the crack so the crack does not propagate. For pressure sensitive adhesives, roughness
generally decreases adhesion – as those familiar with the Dahlquist criterion will know.
The myth of surface energy being important for adhesion comes from some good, simple science. The work of
adhesion Wa of two surfaces can be simply calculated:
Equ. 9-1 Wa= γ * (1 + Cos)

Where γ is the surface energy and  is the contact angle of the second material with the first. If the surface goes
from wetting ( = 0º) to non-wetting ( = 90º) the work of adhesion changes only by a factor of 2. This is a
small effect when we’re talking about real-world adhesion where differences are factors of 100’s.
The real explanation for the dramatic increase of adhesion to polyethylene after corona treatment is the fact that
the polyethylene crystals are broken up at the surface and an open polymer surface is available to interact with
whatever you are trying to stick to the surface. The meaning of “interact” is explained below. The
functionalisation of the surface by the treatment is a surprisingly small effect – often only 1 in 100 surface
carbon atoms has been oxidised. This is enough to allow good wetting of an applied coating (an important
effect) but largely insignificant in providing real adhesion
Before we come on to the key science of real adhesion, it’s worth mentioning the Gecko effect. There have
been many bad explanations of how a gecko can climb up a smooth surface. It turns out to be very simple. Van
der Waals forces of two surfaces in intimate contact are more than strong enough to support the weight of a
gecko. So all the gecko needs to do is make sure there is plenty of surface contact. The feet are designed with
multiple levels of compliance, ensuring that there is every chance for bits of the feet to be in contact with as
much of the surface as possible. In simulations, it hardly makes much difference if the feet are high or low
surface energy. Going back to the issue of crack propagation and surface roughness, it’s clear that cracks fail to
propagate along the many interruptions in the surface contact of the gecko.
But the key thing about gecko adhesion is that it is perfectly useless for “real” adhesion. The gecko can lose
adhesion with a mere flick of its foot. And of course this is vital because the gecko has to be able to walk. If
adhesion were too good it would be stuck forever. So gecko adhesion is closer to “scientific adhesion”. It’s
elegant, functional, explainable with simple surface physics, but ultimately useless for most areas of adhesion
that are of practical interest to humans. It’s common, for example, to have layers of ink with good “gecko-like”
adhesion but which peel off all too easily when put under real-world test conditions.
We shall soon put some numbers to all this which will show that surface-energy adhesions of ~0.05J/m2 are a
factor of 100-1000 too small to explain real adhesion.
In what follows we largely ignore adhesion of polymers to non-polymers such as metals. However, a later
discussion about chemical bonding between surfaces (of particular relevance to metal/polymer bonding) shows
that entanglement issues play an important and little-known role.
HSP to the rescue
When dealing with polymers, real-world adhesion comes mostly from intermingling of polymer chains. The
more the chains are intermingled, the harder it is to pull them apart when stress is applied at real-world speeds.
It’s like untangling a ball of string. If you pull the string very slowly and carefully it’s possible for it to slide out
from the tangles. But if you pull it quickly the tangles tighten upon themselves making it impossible to move.
In scientific terms if the polymers are above a “critical entanglement limit” then adhesion is strong.
To get polymer chains to entangle (and stay entangled – i.e. we’re ignoring temporary tricks such as solvents
compatibilizers) they have to be thermodynamically more stable when extended as opposed to closed in upon
themselves. If you try to mix two polymers with very similar HSP (think of the extreme case where the HSP are
identical because the polymers are identical) then they will entangle readily (provided there is some kinetic
boost such as heat or solvent). Two dissimilar polymers will simply not mix.

46
Figure 9-3 No intermingling (surface energy only, poor adhesion ~0.1J/m 2), straight intermingling (“nail” adhesion ~1J/m 2)
contrasted with entanglement (real adhesion 100-1000J/m2)
It’s obvious from the picture that the polymer system on the left will have poor adhesion. The one in the middle
has good interpenetration of the polymers so you might think that the adhesion will be OK – but it won’t be
much good. We’ll see that it takes the situation on the right with strong entanglement before decent adhesion is
obtained.
The measure of HSP similarity between two polymers is the HSP distance between them (taking into account
their respective radii). A short HSP distance is a strong indicator of the potential quality of adhesion, providing
they have a kinetic chance to mix via solvent or thermal treatment. The effect of chain length is more difficult
to predict. Short chains will not have strong entanglement, but they are easy to intermingle. Long chains will
have very strong entanglement, provided they’ve had sufficient exposure to solvent or heat. Similarly, linear
chains are easier to entangle but easier to disentangle than branched chains. And in a complex polymer, it may
happen that the average HSP is not suitable for compatibility but certain segments (side-chains, blocks) are
more compatible, provided they are given the chance to come to the surface. Sometimes it requires a
compatible solvent to get the correct segments to flip to the surface in order to bond with the other surface. But
even though there are many subtleties, the HSP distance is a crucial component of adhesive thinking.
(We note in passing that these entanglement issues are highly relevant to Environmental Stress Cracking, where
higher molecular weight variants/mixes of a polymer are often more resistant to ESC).
The HSP radii of the two polymers encapsulate one other element of the problem. Polymers with small HSP
radii are, for various reasons such as high molecular weight or high crystallinity, harder to access, so require a
closer HSP match.
The polyethylene example now makes more sense. Wetting and film retention alone does not assure acceptable
adhesion, but by breaking up the crystal structure through the corona treatment, interaction with whatever you
are trying to stick becomes possible. By adding some modest (oxygenated) functionality, the HSP compatibility
with many adherents is increased. There’s a final sting in the tail about corona treatment for increasing
adhesion. The destruction of the surface which is a desirable feature can be taken too far. You then have
polymer rubble on the surface. Adhesion of your ink to this rubble might be quite good. But it is building a
house on sand. It’s well-known, for example, that PET corona treatment can create rubble and considerably
reduce adhesion in many systems, making it a difficult judgement call if you find that for your system the
adhesion is enhanced (once more, by breaking up the crystalline domain of bi-axially oriented PET). How can
you tell if that enhancement is permanent or something that’s going to fail because the pile of rubble falls apart?
The polymer physics of adhesion
The description above is OK as far as it goes, but it’s a bit on the vague side. Let’s get deeper into the polymer
physics. For those who want a masterful modern summary of the physics, try the review by Richard P.Wool,
Adhesion at polymer–polymer interfaces: a rigidity percolation approach, Comptes Rendus Chimie 9 (2006)
25–44. What follows is most relevant to strong adhesion. For pressure sensitive adhesives controlled low
adhesion is required – in which case the physics can be used to avoid situations with too strong a bond (though
there is much more to PSAs than that!)
A useful starting point suggested by Professor Wool is “nails” adhesion. His paper is a classic and well worth
reading: Richard P. Wool, David M. Bailey, Anthony D. Friend, The nail solution: adhesion at interfaces, J.
Adhesion Sci. Technol., 10, 1996, 305-325. Take two blocks of wood and nail them together. The more nails N

47
you have, the longer, L, they are and the higher the friction µ between the nail and the wood, the greater the
adhesion. In fact the Fracture Energy =0.5µNL2. If you do the calculation for typical short polymer chains
across an adhesive bond (as shown in the middle image of the diagram above), the fracture energy turns out to
be something like 1J/m2. This is higher than any reasonable surface energy bonding energy of ~0.1J/m2. But
“real” adhesion is in the 100’s to 1000’s of J/m2, so “nails” adhesion is rather close to useless. If you want two
blocks of wood to really stay together you use screws instead of nails. For polymers you can’t have screws, but
you can have entanglement.
There’s a precise description of “entanglement” and it is therefore possible to calculate some “critical
entanglement parameters”.

Figure 9-4 The classic “Wool” diagram (with permission) showing the three crossings necessary for entanglement.
A polymer chain can be said to be entangled across an interface if it crosses it three times. One or two crossings
are equivalent to one or two nails, only three (or more) crossings ensure entanglement.
It turns out there is a critical length required for a polymer chain to be able to wander back and forth that many
times. This can be expressed as a critical molecular weight Mc or a critical length Lc or critical “radius of
gyration” Rgc (they are, of course, inter-related). The formula for Mc and Lc are:
Equ. 9-2 Mc=31(bz/C)2jM0C∞
Equ. 9-3 Lc=bs(Mc/6M0)½
Unfortunately you need to know quite a lot about your polymer before you can do the calculation. b is the
average bond length, bs (often confusingly shown as b) is the statistical or “Kuhn” length, z is the number of
monomers per unit cell of size C, j is the number of bonds along a monomer (e.g. 2 for PS, PMMA, PE… and
12 for PC), M0 is the monomer molecular weight. C∞ is the “characteristic ratio” for your polymer, a number
that encapsulates the stiffness of the chain and which varies between about 2-18.
A useful collection of these parameters can be found in L. J. Fetters, D. J. Lohsey and R. H. Colby, Chain
Dimensions and Entanglement Spacings, Physical Properties of Polymers Handbook 2nd Edition, Chap. 25,
445-452, 2006.
Here are some typical values deduced from that and other references:

48
Polymer j C∞ M0 z C-Axis Bond Length Kuhn Length Critical MWt Critical Length

nm b in nm bs in nm Mc Lc in nm

PE 2 6.7 28 1 0.255 0.154 1.50 4,379 7.7

PP 2 5.8 42 2 0.65 0.154 1.10 3,501 4.1

PS 2 10 104 3 0.65 0.154 2.00 33,626 14.7

PVOH 2 8.3 44 2 0.55 0.154 0.62 7,330 3.3

PVA 2 9.4 86 3 0.65 0.154 0.65 26,137 4.6

PMMA 2 8.2 100 4 1.04 0.154 1.53 18,411 8.5

PC 12 2.4 254 2 2.08 0.143 1.60 4,426 2.7

PVC 2 7.6 63 2 0.51 0.154 1.00 11,088 5.4

PEO 3 4.2 44 7 1.95 0.151 0.57 5,213 2.6

Figure 9-5 Typical examples of the key parameters for entanglement calculations.
When it comes to adhesion across a polymer boundary these critical values are important. The two polymers
form an interlayer width, d. If d<Lc, the critical length, then you only have nails adhesion. So the question
arises, what is the typical value of d?
There is a well-known formula (after Helfand) for this:
Equ. 9-4 d=2b/(6χ)½
where b is the Kuhn statistical bond length and χ is the interaction parameter. It’s interesting to note that χ is
sometimes assumed to be a specific Flory-Huggins term, but in fact Hildebrand derived it from pure solubility
parameter considerations and it was later shown (especially by Patterson) that (a) the Hildebrand and the Flory-
Huggins χ’s are essentially the same and (b) they both contain approximations that happily tend to cancel out
and give good results. Hansen took the Hildebrand definition and put it into its now-familiar form:
Equ. 9-5 χ=MVol/RT*((δD1-δD2)2 + 0.25*(δP1-δP2)2 + 0.25*(δH1-δH2)2)
Now we have our χ it seems as though we can easily predict adhesion from our HSP! Unfortunately it’s not that
simple. There is significant disagreement in the literature about the use of χ in these calculations, not least
because a standard way of measuring χ using Small Angle Neutron Scattering (SANS) is itself open to much
dispute. As discussed in the chapter on Polymer Miscibility it is arguable that Donor/Acceptor effects should be
included in the calculation of χ so a Donor/Acceptor option is included for those who think it is helpful.
Note that the MVol in the formula is the molar volume of the polymer (not of the monomer unit). If you
quadruple the molecular weight then the MVol is quadrupled and the interlayer width is halved (a 1/MVol½
relationship). This means that the laws of physics are against you. As you increase the molecular weight to
approach or exceed Lc, the interlayer width decreases making it harder for d>Lc. That’s one of the reasons it’s
so hard to get really good adhesion between polymers.
For those who want a more refined version of the Helfand formula that takes into account the MWt then an
approximation based on NA and NB the number of Kuhn monomers per polymer molecule gives:
Equ. 9-6 d=2b/(6χ)½(1+ln2/χ(1/NA+1/NB))
Getting d≫Lc is important and the Number of Entanglements, Ne=d/Lc is a very significant number. There are
good experimental data showing that bond strength is small when Ne<1 (i.e. below the critical entanglement
limit) and then has a linear dependence on Ne>1. So if d=2Lc the bond strength is double that of d=Lc.
The Adhesion calculator in HSPiP can only be seen as indicative. The χ calculation itself is uncertain, and you
have to enter estimates for many parameters to calculate d. Furthermore you are calculating the interface for
two polymers with only one bs and MVol. And you are likely to be dealing with polymers of broad molecular
49
weight distributions rather than the idealised single molecular weight implicit in the theory and used in the
experiments that have proven the theory. But as a guide to the sorts of issue you face when designing an
adhesive interface it seems to be much better than having no tool at all. What is certain is that a smaller HSP
distance will give better adhesion if the following three conditions are met:
• The molecular weights of both polymers exceed Mc for critical entanglement
• The calculated d>Lc
• There has been sufficient bonding time and mobility for the entanglement to take place
The last factor is the great unknown. If you are melt-bonding two polymers then the time for them to diffuse to
distance d depends on t¾ MWt-7/2. The problem is that it’s relatively easy for the polymers to diffuse by
“Rouseian” dynamics but these, by definition, do not involve entanglement, they just give you nail adhesion. As
you become entangled you are in the domain of Reptation dynamics (“a snake sliding down a tube”) which are
slower with a large dependence on MWt. The “obvious” way to increase entanglement is to lower the viscosity
via a good solvent. Unfortunately, the more solvent that’s present, the further apart the polymer chains are so
the less likely they are to entangle. This is not to say that solvent-based entanglement is bad, it’s simply to point
out that polymer physics are against you if you’re not careful.
A logical way to attain a good interface between two polymers is found frequently in the world of real
adhesives. A di-block co-polymer made from the two polymers will obviously be compatible on each side of
the interface. The trick is to make sure that each block is long enough to be firmly entangled within its own
polymer, otherwise you have di-block nail adhesion which is not as effective. The classic study on this (Creton
et al, Macromolecules, 25, 3075, 1992), using PS-PVP di-blocks showed a fracture energy of ~5J/m2 when the
PS was above its entanglement length and the PVP was below it, with all the fracture failure on the PVP side.
Once the PVP exceeded its entanglement length the fracture energy rose to ~100J/m2.
Polymer physics gets you again
In a paper Polymer-Solid Interface Connectivity and Adhesion: Design of a Bio-based Pressure Sensitive
Adhesive, most conveniently found online at http://www.che.udel.edu/pdf/WoolPaper.pdf Wool and Bunker
provide a startling demonstration of the importance of entanglement. Suppose you have a surface (e.g. a metal)
with functional groups –Y and you want to bond a polymer with functional group –X which forms a strong –X-
Y- bond.

Figure 9-6 What are the optimum φ(X), φ(Y) for good adhesion? Less than you would think!
It seems obvious that the stronger the bond and the larger the fraction of –X and –Y bonds (φ(X), φ(Y)), the
stronger the bond will be. “More” is obviously “better”. Yet Wool (and others) have shown that a φ(X)>~1%
leads to a rapid decrease in bond strength. How can this be?
The key to this puzzle is the mode of failure. Below this critical φ(X) the failure mode, not surprisingly, is
adhesive: the polymer pulls away from the surface. Above the critical value the failure mode is cohesive: a thin
layer of polymer remains on the surface. The reason for this is entanglement, or, rather, loss of it. As you get
more and stronger –X-Y- links, the more the surface layer of polymer gets flattened onto the surface and
therefore the less it can entangle with the rest of the polymer partly because of reduced mobility and partly
because the chains do not extend far enough into the bulk in order to entangle. This is not a small effect. For
their particular adhesive the bond strength goes from 250 J/m2 at 1% to 50J/m2 at 1.2%, a catastrophic reduction
in practical adhesion for just a small increase in φ(X).

50
There is no HSP message in this small section – we simply find this to be a fascinating bit of science which is
surprisingly little-known.
A quick way to test for contamination
All the adhesion science is worth nothing if your surfaces are contaminated. So a quick way of testing for
contamination can be very handy. Although this chapter stresses that the common myth about surface energy
being important for adhesion is out by a factor of >1000, surface energy tests can be an excellent way to
identify contaminants. Although one can use surface-energy pens (“Dynes pens”) there is an even easier way. It
was developed by the same Hansen who developed HSP and is an official Nordtest Poly 176:
(www.nordicinnovation.net/nordtestfiler/poly176.pdf ). A set of vials contain ethanol/water mixes varying from
pure ethanol (surface tension=22 dynes/cm) up to pure water (72 dynes/cm). Starting from the ethanol end of
the range, drops from each are placed on the test surface till one is found that does not spread. That, then, is a
guide to the surface energy of that surface. This is not a way to measure the surface energy, it is simply a guide.
The point is that if you know that you get a good bond with a surface that spreads drops up to 46 dynes/cm
(10.9:89.1 ethanol:water) and your test piece spreads only up to 36 dynes/cm (23.4:76.6) then there’s something
wrong with your surface.
Hansen has used the technique to identify contaminants on many items including wind-turbine blades and the
cement lining of a major tunnel. In both cases, just a few drops of ethanol/water were sufficient to identify
problems that could have had major long-term consequences. Ethanol/water is not an aggressive solvent mix for
most surfaces so the results are seldom confused by surface damage from the solvent.
The test isn’t perfect. If your 46 dynes/cm is covered by a 46 dynes/cm contaminant then you’ll never know.
But very often the contaminant will be some sort of lower surface energy oil and the test will reliably detect it.
Polymer viscosity
This section doesn’t have much to do with adhesion. But because so many of the considerations are similar, and
because the calculations in the software are on the same modeller, we’ve included this summary here. You may
find the following a little frustrating. There seem to be so many parameters and assumptions to hold on to and
there seem to be many uncertainties in the estimates. That seems to be the nature of this literature. It is very
difficult to find papers that pin down all the key variables in one go and then link them to experimental values.
We’ve done our best. As ever, if the HSPiP community can provide better ways of doing this, we’ll be happy to
update this text and the program.
First we need to define Rg, the Radius of Gyration of the polymer. It is given by a formula that is similar to the
Mc calculation above. The C∞, j and M0 values are given for a range of polymers in the software and b can be
assumed to be 0.15nm if you don’t have a better value to hand.
Equ. 9-7 Rg0=b [C∞Mj/(6M0)]½
Unfortunately, Rg depends on how “happy” the polymer is within the solvent. This formula gives the Rg for a
polymer in a neutral solvent – i.e. one with neither positive nor negative polymer/solvent interactions. Such a
solvent is termed a theta solvent and so we call this value Rg0. As the solvent gets better, the polymer tends to
expand so Rg is larger. The traditional way is to say that Rg=Rg0 α2 where α is an expansion factor.
Unfortunately there is little agreement on how to know α. In HSP terms, polymer “happiness” is defined by the
RED number. When RED=1 the polymer is on the border between happy and unhappy – i.e. at or near the theta
point and the Rgo from the above formula is used. As the RED reduces, the Rg increases (the polymer likes to be
in the solvent). Note that RED is in solubility parameter units. Chi and the interaction coefficient are in energy
units. A RED of 0.9 converts to a chi effect of 0.81 by the squaring process. In other words the reduction in
RED is more important than a linear reduction. We’ve chosen a simple estimate of this effect that works
adequately with data from the classic known polymers. It’s saying that for a good solvent, Rg goes as Mw1.6, for
a theta solvent it goes as Mw1.5 – but with a higher constant of proportionality. Mw is the weight averaged
molecular weight. For these sorts of calculations, Mn, the number averaged molecular weight is not so useful.
Equ. 9-8 RgRED=Rg Mw(1.6 - 0.1 * RED) / Mw1.5 (1 + 1.5 * RED) / 2.5)
Because the viscosity effects are highly dependent on Rg, we decided to let you decide which value to input. So
we output Rg0 and RgRED to give you an indication of likely Rg values; you then use your scientific judgement
in providing an input value.

51
An alternative way to estimate Rg0 is to use handbook tables of Rg0/Mw½. We’ve provided a selection of these
values in the program. Obviously you need to multiply them by Mw½ to get the Rg0 which you can then
compare to the value calculated in the program.

Figure 9-7 The complexities of polymer adhesion and viscosity made as simple as we possibly can
We must at this point take a side-track into the connection between the definition of a theta solvent and RED
number. At first we hit an intellectual problem. A theta solvent corresponds to a RED=1 (Chi=0.5) where, by
definition, we have no solubility. Yet polymer physicists happily make solution measurements in theta solvents
so there has to be some solubility. The fix for this problem arises because the definition of a theta solvent
applies to a polymer of infinite molecular weight. As the molecular weight gets smaller the polymer starts to
have some solubility. In other words, RED=1 for an infinite polymer means RED<1 for a finite polymer. The
theory of all this is discussed in the next section, but in short, a RED of 0.9 for a polymer of MVol=10,000 in a
solvent of MVol=100 corresponds to a theta solvent. For a polymer of 100,000 the RED is 0.97. Given the
many assumptions behind these corrections, for the purposes of the viscosity calculations, a RED of 1 is
assumed to be equivalent to a theta solvent, with the understanding that the finite polymer is inside the border
of the sphere and can therefore be dissolved in sufficient amounts to make theta-solvent measurements. Errors
in other estimates far exceed the errors in this RED=1 approximation.
The viscosity of a dilute polymer solution (we will define “dilute” in a moment) depends on the intrinsic
viscosity [η]0 and the concentration. At very dilute solutions the specific viscosity is given by [η]0C (Zimm),
where specific viscosity is (ηp- ηs)/ ηs and C is the concentration (in the software C is in g/l). In other words, the
specific viscosity is the viscosity of the solution containing the polymer ηp corrected for the viscosity of the
solvent ηs. By extrapolating specific viscosity to zero concentration you obtain [η]0. We can calculate [η]0 in
units of l/g via:
Equ. 9-9 [η]0=5 AvRg3/Mw
An alternative way to estimate [η]0 is to use handbook tables of K0 where [η]0 =K0Mw½, where the “0”
indicates that these are values for a theta solvent. Handbooks may also provide K and a values for calculating
52
[η] in good solvents where [η] = KMwa and a is higher than the 0.5 for theta solvents (typically 0.6-0.7). It’s
impractical to include these in HSPiP so use the given K0 values to cross-check (multiplying by Mw½) the
intrinsic viscosity calculation using the Rg for a theta solvent, then change to a good solvent and assume that the
enhanced Rg translates to a correctly enhanced [η].
Polymer chains start to interact even in dilute solutions. Because of this the general “dilute solution” formula
for the traditional way of reporting polymer viscosities, ηp/ ηs is given by the following formula. For simplicity
the ratio ηp/ ηs will be called η in the following;
Equ. 9-10 ηdilute=1+[η]C+0.4([η]C)2
Above a certain critical concentration, C*, the solution ceases to be “dilute” and Rousean dynamics take over
and the (hypothetical) Rousean viscosity is given by:
Equ. 9-11 ηRouse= ([η]C)1.3
C* is the point at which the polymer “blobs” start bumping into each other. It is calculated as
Equ. 9-12 C*=Mw/(AvRg3)
Where Av is Avogadro’s number (6.022E23).
The Rousean domain doesn’t last long. As the concentration increases, the polymers start to get entangled and
viscosity starts to rise dramatically. This critical concentration depends on Ne (the number of monomer units in
a critical entanglement molecular weight Me). Ne= Me/M0, where Me is 1/3 the Mc we’ve calculated above. So:
Equ. 9-13 Ne=(Mc/3)/M0
Knowing Ne we can then calculate the critical entanglement concentration, Ce. We called the Rousean domain
hypothetical because the entanglement effects start to kick in fairly quickly (at C>0.5Ce) and it seems hard to
spot true Rousean viscosity because at 0.5Ce you can be close to the dilute limit.
Equ. 9-14 Ce=Ne(3v-1)/(6.6/C*)
The 3v-1 term would be a simple 0.5 in a theta solvent where the polymer/solvent interactions are neutral. But
as you go into good solvents it gets easier to entangle. For classic polymers in good solvents, the exponent term,
v ~0.585, making 3v-1=0.755, a big increase on 0.5. As a practical heuristic we’ve proposed the following for
v:
Equ. 9-15 v=0.585-(0.085*RED)
With all these equations we can now predict viscosity over the entire range of polymer concentrations. If you
play with the Viscosity modeller you will find that small changes in v or in Rg can make large changes to the
calculated viscosity. This makes it very hard to be sure that you’ve got good predictions. You will also notice
that the three graphs plotted (dilute, Rousean, entangled) don’t overlap. This is because no entirely self-
consistent description of all three domains seems to exist and in any case the Rousean curve is hypothetical.
The point of the graph is to show the general form of typical polymer viscosity behaviour.
If you find this disconcerting, the reality of polymer solutions is that relatively small changes (molecular weight
and/or its distribution, concentration, temperature, solvent blend) can produce relatively large changes in
viscosity, so the errors apparent in the lack of overlaps are likely to be smaller than the effects arising from, say,
molecular weight distribution.
If you start going to high RED numbers (>1) and large concentrations you will find that the calculations
become decidedly odd. But remember, at RED>1 you cannot get high concentrations, so the calculations are
irrelevant!
There’s one more key fact. The viscosity we’ve calculated after all this hard work is only the zero-shear
viscosity. Predicting the non-Newtonian changes in viscosity at higher shear is, perhaps, a challenge for future
editions of HSPiP.
But the point of this section is not to provide a 100% reliable viscosity predictor (as far as we know such a thing
doesn’t exist). Instead we’re giving you a modeller where you can see the general effects of molecular weight,
solvent compatibility (RED number), concentration etc. and therefore better understand what’s going on in your
own system. If you are already an expert in the field, then this section won’t be necessary. If you are not an

53
expert then hopefully the calculations and graphs will provide you with some insights that will help you better
understand the literature of polymer viscosity and apply it to your own polymer/solvent system.
Because we’ve taken a giant side-track into polymer physics, we might as well finish off with some
thermodynamics that tidy up some loose ends in the explanation of the polymer sphere radius.
Really understanding polymer solubility
In the “minimum theory” chapter we described how solubility was about thermodynamics. But we didn’t give a
thermodynamic formula for this. It’s now time that we did. We use the convention that “1” refers to the solvent
and “2” the polymer and we use V1 and V2 for the molar volumes and ϕ1 and ϕ2 for the volume fractions. It’s
more normal in thermodynamics to use mole fractions, but because of the huge disparity in molar volumes, the
numbers don’t look very helpful. Because of this disparity the “x factor” is introduced. This is simply x= V2/V1.
So now we can introduce the formula for the partial Gibbs free energy of the solvent on dissolving a polymer:
Equ. 9-16 ΔḠ 1=RT [ln(ϕ1) + ϕ2(1-1/x) + Distance2 ϕ22 V1/(4RT)]
The first two terms are the entropy of mixing, the third term is the enthalpy of mixing. The factor of 4 is there
because of the factor of 4 in the (δD term of the) HSP Distance calculation. The combination of the first two
terms is usually called the Flory-Huggins entropy.
Because negative free energies are good for solubility, it’s important to note that the 2nd and 3rd terms are both
positive. So solution is only possible because of the ln(ϕ1) term. As you dissolve more polymer (ϕ2 increasing),
ϕ1 decreases so ln(ϕ1) gets more negative – the classic entropic driving force. But of course the ϕ2 term gets
ever larger, fighting against solution. Only if x≤1 does the ϕ2 term cease to be a problem. But x is always large
so the 1-1/x term is always >0. For an infinite molecular weight polymer, 1/x=0 so the ϕ2 effect is maximum.
As the polymer gets smaller the ϕ2 effect reduces (though not by a lot) and the chance of being soluble
increases. Non-intuitively, a larger solvent means a reduced ϕ2 effect, so in this respect small solvents are bad.
The Distance2 factor in the third term reminds us that the smaller the HSP distance, the higher the chance of
dissolving. The sphere radius then comes from when the Distance is large enough for the 3rd term to produce a
positive ΔḠ. As we will shortly see, this distance is between 7 and 8, which is what is found experimentally for
a large number of polymers. The reasons for variations around this number will shortly emerge. And because
the 3rd term also includes V1, we will soon see that small solvents are nearly always a good thing, as intuition
and practical experience shows. In other words, the 3rd term solvent size effect (small is good) usually
outweighs the ϕ2 effect (small is bad).
It’s not obvious how all these terms work out in practice. So we’ve made a Polymer Solution Theory modeller
that lets you input the key properties: V1, V2 and Distance. The ΔḠ is then plotted over a ϕ2 range of 0-0.3. The
scale for ΔḠ is deliberately made small so that the details can be seen around the important 0 line. In particular,
the classic case of a 10,000:100 polymer:solvent gives an inflexion point at a critical value which translates to a
Chi value of 0.605. For an infinite polymer, this point of inflexion moves to 0.5, which is why a Chi of 0.5 is
seen as the definition of borderline solubility (theta solvent). Because of the relationship between Chi and
Distance for an infinite polymer a Distance of 7.05 provides the boundary of solubility. This can be calculated
from first principles, but you can show this for yourself using the modeller.

54
Figure 9-8 For an “infinite” MVol polymer the thermodynamics of solubility go critical at Chi=0.5 or a Distance of 7.06.

For the 10,000 polymer (as shown in the figure below) it’s a distance of 7.77. The ratio 7.06/7.77 ~ 0.9, which
is why we stated above that a RED of 0.9 suffices for a real polymer compared to an infinite polymer.

Figure 9-9 For a 10,000 MVol polymer in a 100 MVol solvent, the thermodynamics of solubility go critical at Chi=0.605 or a
Distance of 7.77. That’s why we often say that a typical polymer HSP radius is 7-8.

55
Remember, this is all theory based on many assumptions. It is well-known, for example, that the Flory-Huggins
entropy is not a perfect approximation. Real solvents and polymers can’t be calculated with this precision and
some spheres have a radius of 10 and some of 5.
Let’s first see why we might have a small radius. There’s an important term missing from the ΔḠ equation. If
the polymer is crystalline then (thermodynamically speaking) before it can dissolve it first has to melt. This
“crystalline” energy (sometimes confusingly called “ideal solubility”) is discussed in the Predictions chapter
but the formula is simple:
Equ. 9-17 C = ΔF/R*(1/Tm – 1/T)
where ΔF is the enthalpy of fusion and Tm is the melting point. For normal crystalline solids these terms are
readily known. For polymers things are seldom so crystal clear. We are therefore making no attempt to predict
it. However, what is obvious is that the larger the value of C the smaller the Distance needed to ensure
solubility. That is why highly crystalline polymers have such small radii. Polyethylene is a classic example. Its
HSP mean that finding a compatible solvent is not at all hard. But solution experiments using polyethylene
normally take place at ≫100ºC so that the (1/Tm – 1/T) term becomes small enough to allow the solvent to
dissolve the polymer. The reason that you find solvents such as Decalin being used for polyethylene is not
because they have special solubility parameters, but because they have a high enough boiling point to be able to
do experiments at 130ºC.
Rather satisfyingly we can come back to the corona treatment of polyethylene. The disruption to the
crystallinity at the surface means that the C term becomes small so that ordinary solvents at low temperature
have a chance of interacting strongly with the top few nm of the polymer film, provided their Distance from
(perhaps slightly functionalised) polyethylene is not too large.
One possible reason for a large radius comes from the following inverted reasoning.
Backward thinking
It’s obvious that if the polymer isn’t soluble in the solvent then the solvent isn’t soluble in the polymer. This
“obvious” fact is completely wrong. The “x-factor” above explains why. If the 1 and 2 of solvent and polymer
are swapped then the thermodynamics stay the same but the important ϕ2 (1-1/x) term is transformed. For a
10,000 polymer and the same solvent, x is now 0.01 so the ϕ2 term is multiplied by 1-100 which is decidedly
negative, so this term strongly aids solution.
In other words, even though the polymer may be hardly soluble in the solvent, the solvent may be highly
soluble in the polymer. This very important point is often missed.
For example, it illuminates the discussions on diffusion into polymers. Even though the polymer may never
dissolve in the solvent, there is still a chance that the solvent will dissolve in the surface of the polymer, starting
the diffusion process. The HSP distance is still very important – the smaller the distance the larger the amount
of solvent that will dissolve in the polymer – even though classic HSP solution experiments might show that the
polymer is effectively insoluble in the solvent. You can (as discussed in later chapters) determine HSP via
diffusion/permeation experiments. Solvents that are no good for dissolving the polymer can still permeate
reasonably, so the sphere based on these experiments can be large.
This effect also explains why the HSP spheres of solvent swelling experiments can be larger than spheres of
true polymer solution.
This doesn’t mean that the diffusion/permeation or swelling spheres are “wrong”. You as the scientist choose
your definition of the sphere according to your purpose. If you are interested in polymer diffusion/permeation
or polymer swelling then such a large sphere is correct. You only go scientifically wrong if you try to use the
numbers from one domain (swelling) in another (polymer dissolving in the solvent).
This backward thinking isn’t just for polymers. A classic example is that the solubility of water in hydrocarbons
is hardly affected by the MVol of the hydrocarbons. Yet the solubility of hydrocarbons in water is highly MVol
dependent. The best way to understand this is via the classic solubility equation. At first there seems some
confusion as the Distance term now contains V2, the solute, whereas in the polymer case it was V1. The reason
is that in the polymer case we were calculating the partial Gibbs free energy of solution, with the focus on V1.
In the classic solubility case we are calculating the activity of the solute, so the focus is on V2.
Equ. 9-18 ln(a2)=ln(mf2) + Distance2 ϕ12 V2/(4RT)]
56
Here we have the activity of solute 2 in solvent 1. We use mf2 for mole fraction rather than the conventional “x”
to avoid confusion with the use of x above.
When we are trying to dissolve water in hydrocarbons, V2 is that of the water and because Distance2 doesn’t
vary all that much between hydrocarbons, solubility does not change much.
When we try to dissolve hydrocarbons in water, V2 is that of the hydrocarbons and the activity increases
(solubility decreases) as they get larger.
Finally, although in the case of water dissolving in hydrocarbons MVol makes little difference, we need to
return to the general rule that smaller solvents are better for polymers.
This can easily be checked from the Polymer Solubility Theory modeller. Just change MVol Solvent from 100
to 90 (whilst keeping the HSP Distance the same) and a situation of critical solubility becomes one of good
solubility:

Figure 9-10 For the same HSP Distance, if the solvent MVol decreases, the free energy curve ceases to be critical and the
polymer is likely to dissolve.
The Solvent Optimizer contains a simplified version of the solubility/distance equation. If you enter the Radius
of your polymer, the “Relative Solubility” is calculated as exp(-Distance2 V2/Rterm) where we’ve assumed a
constant ϕ12 and use an “Rterm” which is proportional to the chosen Radius (and contains the RT term). A
larger Radius means a smaller negative exponential and therefore a larger Relative Solubility. With the mouse
you can read off the various solvent values from the graph.

57
Figure 9-11 Solubility graph in the Solvent Optimizer.

Flory-Rehner theory for rubber swelling


It is convenient (because the theory is similar and the calculation is done on the same form) to show how the
swelling of a rubber/gel depends on Distance. The theory used is the well-known Flory-Rehner theory.
Although other such theories can be used, this is adequate for our purposes. The classic form of Flory-Rehner
shows how the number of chains between cross-links, n, depends on the volume fraction of polymer in the
swollen rubber, φ 2, the MVol of the solvent, V1, and the χ parameter, expressed in terms of Distance.
Equ. 9-19 –[ln(1- ϕ2)+ ϕ2+Distance2 ϕ22 V1/(4RT)]= V1n(ϕ20.33- ϕ2/2)
As n is not such a familiar term, it’s better expressed in terms of the MWt of the polymer, M2, its density ρ and
Mc which is the MWt of the chains between cross-links
Equ. 9-20 n=ρ/Mc(1-2Mc/M2) or, more usually when M2 is large, n=ρ/Mc
If, for simplicity of inputs, we sweep up the polymer density and MWt into the MVol term, V2 which is an
input for the solubility graph and note that the measured swelling, V=1/φ2, then it is possible to create a graph
of Mc v Swelling for a given solvent, polymer and HSP Distance. As Distance gets large the equation starts to
break down so the plot is terminated if the calculated Mc is ever less than the previous value.

58
Figure 9-12 Flory-Rehner swellability

If you have some swelling data then you can move your mouse over the graph to your equilibrium swelling
value (φ2 is also shown) and read out Mc. If you know the actual value of Mc (e.g. from RPA tests) then you can
adjust Distance till the value you read out is correct. Of course one can use Flory-Rehner and swelling data in
various solvents to fit the HSP of the polymer. Although such fitting is not included in HSPiP it is readily done
within something like Excel.
Pressure-sensitive adhesives
The above discussion of strong adhesion doesn’t really apply to PSAs where controlled, low adhesion is
required. A lot of the magic of PSAs comes from controlling their modulus so that they meet the so-called
Dahlquist criterion or fit into Chang windows (terms familiar to those in the business!). Even more, formulators
need to understand time-temperature equivalence (the WLF equation!) and be at ease with G’, G’’ and Tan δ
for their systems. None of this is directly related to HSP. However, top formulators agree that you can’t design
a good PSA without knowledge of HSP. This is because the tackifiers that are usually required for a viable PSA
must be completely compatible with the main resin – as judged by a short HSP distance. And for those systems
that rely on “physical cross-linking” via phase separation of end-group blocks (such as polystyrene) it is vital
that the tackifiers are not compatible with these end blocks unless you really know what you are doing through
the use of end-block tackifiers. This brief section is meant to pique the interest of those who weren’t expecting
to find anything on PSAs in the eBook.
Summary
Congratulations if you’ve reached this far. Our view of this chapter is that it poses three further challenges.
1. The “easy” challenge lies in understanding the formulae and getting used to playing with the modellers in
order to work out what’s going on. We’ve peppered the modellers with the formulae so that when you get a
surprise (as we often do) you can work out why changing parameter A affects parameter B which in turn affects
the output C in a way you didn’t expect.
2. The “hard” challenge is to find ways to map the theory into one’s scientific and technical work. Even though
the theory is complex enough, it still contains lots of simplifying assumptions and, as we admit, only gives
“indicative” numbers. Yet we’ve found in our own real-world work that the modellers and the ideas behind
59
them have proved crucial in disentangling (pun intended) complex issues. They are particularly good in
providing a common numerical language that a cross-functional team with varying expertise in physics,
chemistry and analytical science can use to build a fuller picture of what’s going on in their system.
3. The final challenge is for ourselves, with the help of the HSP community, to build further on these
foundations and increase the range and depth of the predictive power derived from the theory.

60
Chapter 10 Polymer Miscibility
The ideas on polymer solubility in the previous chapter allow prepare us for a look at polymer/polymer
miscibility.
To put this topic into context, it comes as a bit of a surprise to most of us that high MWt deutero-PE and normal
PE are immiscible, or, to go to more familiar polymers, PMMA and PEMA (methyl methacrylate and ethyl
methacrylate) are immisicible.
By normal HSP standards the immisicibility of PMMA and PEMA makes no sense. Their HSP Distance is
something like 2.4 which most of us would accept as “close”. Yet PMMA through to PHMA
(hexylmethacrylate) are all miscible with PVC even though their HSP Distances span a range larger than 2.4.
The explanations for all these effects can be summarised in two words: Coleman-Painter (or, equivalently,
Painter-Coleman. From now on we will use them abbreviate in alphabetical format as C-P). Professors Michael
Coleman and Paul Painter, both at Penn State University, have effectively sorted out the whole issue and their
approach is immensely powerful. Although they use solubility parameters, they show convincingly that in their
standard format(s) they can predict only that most polymers are immiscible. Although they use Hildebrand
parameters (with their many limitations), even if HSP are used the same problem is encountered – that most
polymers are predicted to be immiscible.
An early guide to the general issue of polymer miscibility is found in Michael M. Coleman, Carl J. Serman,
Dorab E. Bhagwagar and Paul C. Painter, A practical guide to polymer miscibility, Polymer, 1990, 11-87-1203.
This was written before they had properly developed the C-P methodology but sets the scene for all their later
work. The definitive guide to their full theory is contained in the book/CD, Michael M. Coleman and Paul C.
Painter, Miscible Polymer Blends, 2006, DesTech Publications. Their academic publications on the topic are
readily found and provide deep insights into these complex issues.
The key to C-P is that hydrogen bonding between a donor polymer and an acceptor polymer provides a negative
free energy change that can drive mutual miscibility. Standard HSP can only give positive free energy, hence
cannot predict miscibility when the HSP Distance is larger than a minimum value discussed below.
C-P recognise that there aren’t many pure H-bond donors (though there are plenty of pure H-bond acceptors).
When it comes to strong H-bond donors such as alcohols they are also strong H-bond acceptors. So there is
competition between the donor/acceptor pairing of the alcohol polymer itself and the alcohol donor and the
acceptor on the other polymer.
The competing inter/intra effects can be studied scientifically via IR spectroscopy. A series of elegant studies
have allowed C-P to parameterise these effects and therefore build a powerful predictive model not just for
homopolymer/homopolymer interactions but also homopolymer/copolymer and copolymer/copolymer
interactions. The book/CD package mentioned above allows you to specify the (co)polymers and will calculate
miscibilities at any given temperature as well as phase diagrams and miscibility maps.
The C-P approach includes adjustments for the fact that polymers are not able to fold fully onto themselves so
the number of donor/acceptor bonds in a homopolymer, for example, will be less than expected.
The intention in this chapter is to explain a simple tool that has been added to HSPiP that uses the C-P insights
but takes none of their specific H-bonding science. Instead it uses the HSP Donor/Acceptor idea to calculate a
new HSP Distance which, given the right polymer pairs can be reduced to below the critical distance for
miscibility. No attempt is made to calculate C-P energetics via specific H-bonding calculations, nor are there
corrections for the folding limitations. Those who need full C-P capabilities should use the C-P methodology
which we regard as a major addition to the science of polymers. The approach here is more about building
intutions and playing what-ifs than trying to make serious predictions – though our own experience in using it is
that it is surprisingly good in replicating the trends shown in the literature and highlighted in the many excellent
C-P papers and their book/CD.
The basic law of polymer mixing
Assuming you have read the previous chapter on the Flory-Huggins theory of polymer solubility, the key
formula for polymer/polymer miscibility will look very familiar. Where we have RT as gas constant times

61
Temperature (°K), φ1 and φ2 as the volume fractions and M1 and M2 are the number of monomer units
(explained below) of Polymers 1 and 2 then:
Equ. 10-1 ΔḠ mix= RT[(φ1/M1)ln φ1 + (φ2/M2)lnφ2) + Distance² φ1φ2 100/RT]
For the moment, Distance can be thought of as the normal HSP distance. The first term is the entropic
contribution – mixing is always a good thing in terms of free energy – so is negative. The term containing
Distance is the enthalpic term and is always going to be positive, which means that at best it can be neutral
(Distance=0) but in general fights against miscibility (Distance >0)
The factor of 100 is a “typical” MVol that seems to work adequately for all C-P work.
The “monomer units” are simply Polymer_MWt/Monomer_MWt
If we take a 50:50 blend of two polymers of 100K MWt and monomer wt of 100, then the entropic term is -
1.71. For a Distance of 0.2 the Distance term is 1.69 so the net free energy<0 so the polymers are miscible at
50:50. But if the Distance is 0.3 the second term is 2.25, so the free energy of mixing is positive and the
polymers are no longer miscible. This starts to explain why PMMA and PEMA are immiscible. As C-P point
out, the chances of HSP being close enough to be in the <0.5 range are quite small and knowing with sufficient
accuracy that the Distance is 0.2 or 0.3 is essentially impossible.
If the MWts are both reduced to 10K then (as is obvious by inspection) the entropic term becomes more
negative by a factor of 10 and you need a Distance of 0.83 before the polymers are immiscible. This is an
improvement, but 10K polymers are getting quite close to oligomers and even here miscibility is difficult to
achieve via HSP matching.
So how is it that PVC and PVA which are clearly different in HSP terms are known to be generally miscible?
The only way to improve on miscibility is to add a negative component to the HSP Distance calculation, and
the only way to get this to happen is via Donor/Acceptor interactions. These are discussed in the chapter near
the end of this eBook, but briefly we can split δH into δHdon and δHacc and use the fact that donor/acceptor
interactions are going to be enthalpically favourable. How is that split done? In the Miscibility calculator you
are the one who decides where in the range between all donor and all acceptor each of the polymers lie. Given
that choice, the balance of δHdon and δHacc must follow the rule that:
Equ. 10-2 δH² = δHdon² + δHacc²
This is done for you automatically. As a guide, something like PVC has a small but significant δHdon and very
little δHacc, so the slider deciding the donor/acceptor split should be over to the left. Like most other polar
polymers, PVA is pure δHacc, so the slider should be over to the right. For something like polyvinylphenol,
PVPh, it’s safe to put the slider in the middle for approximately equal donor/acceptor, though arguably it should
be a bit to the left as alcohols are somewhat better donors than acceptors.
It is generally unknown what the formula is for the contribution of donor/acceptor to the Distance, but
experiments with a number of options showed that the Beerbower formula was the most reliable in predicting
the trends of known polymer pairs. So the HSP Distance is given by the standard first two terms followed by
the Beerbower variant for δH:
Equ. 10-3 Distance² = 4(δD1-δD2)² + (δP1-δP2)² + 2(δHDon1-δHDon2)(δHAcc1-δHAcc2)
Crucially, the Donor/Acceptor term can be negative thereby reducing the overall free energy. What this means
in terms of Distance (which might be the square root of a negative number) is not important because the
miscibility formula uses Distance².
Second Derivatives
The discussion above has all been about free energy. Clearly if the free energy is not negative then there can be
no mixing. But even if it is negative for some values of φ1 and φ2 the system might not be miscible. This is
because as φ1 and φ2 change the slope of the free energy curve might be such that the system will go into
“spinodal decomposition” which means that phases separate. To ensure good mixing it is important that the
second derivative of the free energy curve with respect to the phase volumes remains positive. For those who
want to understand this fully, the C-P book is a good guide. The point of this section is that it is possible to
create a phase diagram of the whole system (an option in the calculator) by systematically varying the
temperature, scanning the phase volumes and seeing if the second derivative becomes negative, which means
62
phase separation. For those interested in this aspect of mixing, the standard plot in the Miscibility calculator
shows both the free energy and the second derivative.
Miscibility calculator
On the Polymers section clicking the Miscibility button brings up the calculator. The user puts the HSP of the
polymers into the respective boxes. If they have been copied from, say, the Polymers database this can be done
via Ctrl-D, Ctrl-P as happens throughout HSPiP. Without selecting the Donor/Acceptor option it quickly
becomes clear that even relatively close polymers show no miscibility when their MWts are in the range of
100K.

Figure 10-1 For two polymers with a Distance of 0.22 the free energy (blue) is negative and the second derivative (red) is
positive, so miscibility (at 100°C) is OK.

63
Figure 10-2 For only a slight increase in HSP Distance, the situation changes dramatically. The free energy goes positive and
the second derivative goes negative. The polymers will not be miscible.
The situation changes dramatically when the donor/acceptor option is applied and the two polymers are
complementary donor/acceptor.

64
Figure 10-3 This extreme donor/acceptor pair show a negative distance and despite significant differences in δD and δP the
polymers are highly miscible.
It should now be obvious how vital it is to have donor/acceptor interactions for polymer miscibility.
Things become interesting when the donor/acceptor interactions are not as perfect as in the artificial example
above. Swapping to phase diagram mode and tweaking some of the settings you find regions of immiscibility
(shown in blue):

65
Figure 10-4 In phase diagram mode it is possible to see regions of immiscibility (blue) in a wider area of miscibility .
As soon as the polymers differ in terms of MWt, the symmetry is lost and a more realistic phase diagram is
obtained:

Figure 10-5 The same set-up but with low and high MWts leading to a pronounced asymmetry.
You now have enough information to explore the issues of polymer miscibility in your own system. Remember,
this HSP-based approach is not intended as a substitute for the C-P method, but rather it is a way of building
intuitions of what are the important issues for miscibility.

66
Polymer miscibility is a sensitive test for free energies because the entropy/enthalpy balance is delicate and
easily swung via some donor/acceptor interactions. Hopefully with more data on more polymer mixes it will be
possible to further refine our thoughts on donor/acceptor.

67
Chapter 11 Shades of Black (Designed Partial Compatibility - Bitumen)
This chapter is based on the work by Per Redelius of Nynas Bitumen which appeared in Chapter 9 of the
second edition of the Handbook. His data are reproduced by permission.
It is very easy to take bitumen (usually called asphalt in the US) for granted. It’s just black sticky stuff that
comes from oil.
That’s true till your life depends on it. Road safety depends strongly on the qualities of the road surface which
in turn depends on the subtle details of the bitumen used to make it. More sophisticated road surfaces use
bitumen that has been modified with polymers. It is important to know the compatibility between the bitumen
and the polymer.
Bitumen is produced by distillation from crude oils and its properties can vary widely depending on the crude
oil used. Unfortunately it is exceedingly difficult to analyse. And because it is so black and viscous, it is hard to
know whether, for example, you have a solution or a dispersion when you add a polymer to the bitumen. And
of course whether it is dissolved or dispersed makes a big difference to the ultimate properties of the bitumen
and the road surface.
We know that a thermodynamic definition of “compatibility” is provided by HSP. So if we know the spheres of
the bitumen and the polymer we can calculate how much they overlap. If the overlap is strong we can be
confident of compatibility. If the overlap is small then compatibility will be unlikely.
First we need to characterise our bitumen. From the solubility in 42 solvents we can calculate the properties:
Well, actually, it’s not quite that simple. It’s very hard to know if you’ve got a perfect solution or if there are
some bits or if it’s merely swollen or... A Yes/No solubility test is rather hard to apply. So you can score the
solubility from 1-6 where 1 means that there’s not the slightest residue if you filter the solution to 6 which
means there’s not a hint of colour in the liquid. You then need to define what you mean by “good” or, in the
language of HSPiP “Inside” the Sphere. If you’re really careful, you’ll define “Inside” as “1”. Or you might be
interested in “good enough to be inside” so define “good enough” as “3”.
This brings in the crucial question of subjectivity in HSP calculations. How can I tell the difference between a
score of 2 and a score of 3? How worried should I be about a mis-scoring?
The good news is that the software lets you try out the effects of allowing different scores to be classed as
“Inside”. With this particular bitumen, the variation in parameters between 1 & 2 is not all that great. When you
admit 3’s as “Inside” then the radius (not surprisingly) increases.
Inside score δD δP δH R
1 18.4 5.8 2.9 6.2
2 17.9 5.1 3.1 5.8
3 18.2 3.8 4.6 9

Table 11-1 How the Sphere parameters depend on the definition of “Inside”
The plot below (based on a generic bitumen) uses a 0-15 scale for P and H to provide a more detailed view for
this and subsequent plots.

68
Figure 11-1 Using file Bitumen
So now we can start looking at different polymers. If we try polyether sulfone we find that the spheres are
highly distinct – so it is unlikely that the polymer will be anything other than a dispersed solid when mixed with
the bitumen. To get this plot, open the Polymer form and enter the Bitumen data as the last item. Select
Bitumen and also select Polyethersulfone. Then click Solvents and the two polymers and the solvents are
plotted for you.

Figure 11-2 Comparison of Bitumen and Polyethersulfone


A plot comparing bitumen with Styrene-Butadiene block co-polymer (SBS) is more interesting because SBS is
the key polymer used with bitumen in road construction.

69
Figure 11-3 Bitumen and an SBS polymer #271
It shows significant overlap but a significant part of the SBS is outside the bitumen. Given that SBS has two
very distinct polymeric components it is interesting to note that if you compare polystyrene with polybutadiene
you see that they are not very mutually compatible – which is why SBS is an “interesting” polymer. And if you
calculate the overlaps between bitumen, polystyrene and polybutadiene you find that they have similar degrees
of overlap.

Figure 11-4 Polystyrene and Polybutadiene spheres – the radii are made smaller to help visualise the plot for this figure
The SBS must neither be fully compatible nor fully non-compatible. The HSP of the bitumen must be such that
the appropriate marginal compatibility is achieved for optimum performance at both high and low temperatures.
With an extensive table of polymer HSP it should be easy to find polymers with a strong match of bitumen.
This is easy to do within the program. If you double-click on bitumen, the polymers in the table are sorted with
respect to their double-overlap with bitumen. Double-overlap? Yes, you have to report two figures – how much
the bitumen sphere overlaps with the polymer sphere, and vice versa. For example, if Polymer A is a small
sphere that is completely inside Polymer B, its overlap score is 100%. But because Polymer B’s sphere is so
70
much larger, it is overlapped by only a small % by Polymer A, so the score is, say, 20%. Thus the combined
overlap is shown as 120 100/20. Without this double-overlap score it is easy to jump to the wrong conclusion
about the mutual overlap of spheres. Here is what you see on the screen.

Figure 11-5 Polymers sorted by mutual overlap with Bitumen


It is interesting to note that near the top of the list are two styrene polymers – just as you would expect from the
previous discussion. Whether you would want to use Hypalon 30 in your bitumen, only you can tell – but you
could be confident that it would be highly compatible.
In summary, HSP have brought light to a very black subject. With just a few experiments you can know where
to look for interesting compatibilities or incompatibilities with polymers and have a high degree of confidence
that when you mix them you will have results along the lines you expect. This is a far more elegant approach
than experimenting in the dark.
Another form of black material is carbon. Like all pigments/fibres it is insoluble so the “S” in HSP certainly
doesn’t apply – or does it?

71
Chapter 12 Insoluble solubility parameters (HSP for Pigment Surfaces)
A lot of the colours we see around us come from pigments. By definition these are insoluble, so it seems to
make no sense to worry about their “solubility parameters”. Yet the HSP approach has proved immensely
valuable – giving lots of practical insight for comparatively little work. In this chapter we’ll stay mostly in
black, with the various forms of insoluble carbon. Yet the principles apply to pigments of any hue.
It seems an admission of defeat to introduce the concept of a pigment being “happy” in a solvent. How can such
a term apply to something as scientific as HSP? Let’s turn the question around. As a scientist, you can shake up
a sample of, say, carbon black in a solvent and know that the pigment is happy or unhappy in that solvent. For
example, a happy carbon black will go into dispersion in a solvent with a mere shake of the test tube. An
unhappy carbon black will simply sit as a lump in the bottom of the test tube no matter how much you attack it
with high energy ultrasound. If you use a range of solvents covering HSP space you can form a judgement of
happy/unhappy and put those data into HSPiP and calculate the HSP of the carbon.
For those who want to be more sophisticated, you can score the happiness in more objective ways. For example
you can measure the sedimentation rate and assign numbers on the basis that faster sedimenting pigments have
poorer solvent interaction than slower ones. However, note that Ch.7 of the Handbook introduced the concept
of RST – Relative Sedimentation Time – to help correct for differences in sedimentation due to
density/viscosity:
RST=ts(ρp- ρs)/ η
where ts is the actual sedimentation time, ρp and ρs are the densities of the particles and solvent and η is the
viscosity. The RST values, rather than the raw ts values should then be used to decide between “good” and
“bad” solvents.

Either way, you will find yourself with a plot such as the following:

Figure 12-1 Using file CarbonBlackLow


If you try a different type of carbon black you find a very different result:

72
Figure 12-2 Using file CarbonBlackHi
These two simple experiments reveal a profound difference between two pigments both labelled “carbon
black”.
δD δP δH R
CarbonLow 16.5 9.1 6.8 6.9
CarbonHi 20.4 10.9 13.0 11.5

Table 12-1 Comparison of parameters for CarbonLow and CarbonHi


Indeed, there are many different types of carbon black with very different surfaces and therefore very different
abilities to interact with solvent or polymer binder. If you don’t have the HSP, how can you rationally optimise
your carbon black formulation?
If your binder and pigment have identical HSP then you have perfect compatibility. But what do you do when
you have a coating containing pigment, binder and solvent? It seems obvious that your solvent should also have
the same HSP. But this would mean that the binder/solvent interactions were so strong that the binder/pigment
interactions could be overwhelmed. If the binder has HSP somewhere between the solvent and the pigment, and
if the solvent is on the boundary of the binder then parts of the binder will tend to associate strongly with the
pigment, probably leaving its solvent-compatible parts on the outside and thereby giving very good solvent
compatibility for the whole system, whilst ensuring that the binder is nicely locked on to the pigment when the
solvent evaporates.
Just pause to think on that paragraph. All you need in order to come up with a good starting point for a practical
pigment dispersion are the HSP of pigment and binder. With help from the program you can rapidly identify a
solvent that is on the outer rim of the binder sphere, with the pigment still further away.
Let’s try it with PMMA and the CarbonBlackHi. Load a typical list of solvents (such as FriendlySolvents),
select PMMA in the Polymers form, make sure you’ve selected PolymerR so that RED numbers are calculated
on the basis of PMMA’s radius, and click the Solvent Match button. When you look for solvents with a RED
number ~1 (i.e. on the border), Propylene Carbonate looks a good fit.
δD δP δH R
PMMA 18.6 10.5 7.5 8.0
CarbonHi 20.5 10.9 13.0 11.1
Propylene Carbonate 20 18 4.2 RED=1.03

73
Table 12-2 Finding a borderline solvent
If you wanted a good place to start to generate a good formulation using PMMA and this carbon black, then
MIBK would be a good place to start.
If you were using the CarbonBlackLow you would, of course, choose a very different solvent.
As this is a chapter about the truly insoluble, we introduce a pleasing digression. We came across a wonderful
YouTube video http://www.youtube.com/watch?v=jQdCRARzOv8
which shows how to determine the solubility parameter of glass. You simply find which liquids completely wet
the glass (“Good”) and those which don’t (“Bad”) and run the Sphere correlation. We are grateful to Prof
Darren L Williams of Sam Houston State University, Texas for permission to reproduce his data here.

Figure 12-3 The HSP of glass


As you can see, glass is estimated to be [13.3, 3, 12.8]. It will be very interesting to see if Dr Williams’
technique can be extended to other surfaces and add insights beyond the traditional surface energy calculations.
As surface energies are often broken down into sub-components such as Dipolar, Polar, Lewis Acid/Base it
would seem an interesting research project to see if the HSP breakdown into δD, δP, δH proves to be fruitful in
understanding surfaces. It is worth noting how unusual this HSP set is. By using the entire Master Data set and
putting the glass values into the Polymer table and clicking the Solvents button, the glass sphere is outside the
entire solvent range. It will be interesting to know if this is an artefact of the fit or a real insight into the glass
surface:

74
Figure 12-4 The glass sphere plotted in the context of the entire Sphere Solvent Data set, showing how unusual it is
This is our first venture out of the “solubility” comfort zone of HSP. Let’s carry on to see another area where
the last thing you are interested in is polymer solubility.

75
Chapter 13 Cracks in the system (Environmental Stress Cracking)
It’s not every day that you are asked to solve a very expensive problem for a large aquarium. A fire had
damaged the large PMMA front of the shark tank for a famous aquarium. Replacing such a large piece of
PMMA would have been very expensive. But it was unthinkable to risk having it fail whilst full of water and
sharks. With the aid of a few drops of liquid and a deep understanding of HSP and environmental stress
cracking (ESC), one of us (Hansen) was able to say authoritatively that the tank was entirely fit for purpose – a
decision vindicated by many years of safe use since it was refilled.
We’ll discuss a rather simpler example in this chapter, but it’s good to know that the principles can be applied
to many different situations, including large shark tanks.
ESC is a huge practical problem. It is said that at least 25% of failures in plastics are caused by it. Almost by
definition it is a difficult problem. Yet HSP can very quickly tell you if you are likely to be in an ESC danger
zone.
Let’s remind ourselves of the problem. You have a polymer part that you don’t want to be damaged by contact
with solvent. So you do some tests with a range of solvents. If you try a good solvent (i.e. inside the sphere of
the polymer) then, yes, you have a problem in that the polymer dissolves. But really this isn’t a serious
problem. The test takes a few moments and you then know that you must keep that solvent away from that
polymer part. If the part has to handle that solvent then you have to change the polymer. With a solvent you
know to be outside the sphere you know the answer in advance but you do the test anyway and, of course, the
solvent doesn’t do anything to the plastic. It will not be a problem, so you say that this solvent is safe to use.
And yet, some days, months, years later after exposure to this “safe” solvent the plastic suddenly cracks and
you have an expensive repair bill or an angry customer. Where had you gone wrong? The answer is in the
thermodynamics. You’d rightly concluded that there was a net increase in energy if the solvent infiltrated the
polymer and therefore it would not interact. What you’d forgotten was the “S” in ESC. The stress is another
factor that can tip the thermodynamic balance. If you have a large concentration of stress, and if the solvent and
polymer are near the thermodynamic tipping point then the solvent can enter the plastic and, snap, you have
your cracked or broken part.
When the problem is expressed in these terms, the solution is obvious. You will only get ESC from solvents
that are near the border of the sphere. Solvents inside give dissolution and you easily identify them. Solvents far
outside will never be close to the thermodynamic limit and will therefore never cause ESC. So for those of us
who have to worry about ESC the rule is simple: Beware of solvents at the boundary – and therefore make sure
you know where your boundary is.
Here is a typical example based on the Topas 6013 ESC example in the Handbook.

For this plot, only true solvents were used to define the HSP

Figure 13-1 Using file ESC with “Inside” set to 1


This doesn’t tell you much about ESC.

76
For the next plot, solvents that tended to show ESC are included in “Inside”. The calculated HSP and radius has
changed.

Figure 13-2 Using file ESC with “Inside” set to 2


The point of this plot is that the solvents outside this radius are almost certain to be safe from ESC.
The “almost” is there for two good reasons. First, the radius defining ESC depends strongly on the Stress. If the
stress is low, then the radius is close to the first plot. If the stress is high then the ESC radius will expand. Only
you can judge the level of stress and therefore the margin of safety for solvents. Second, the molar volume
plays a role. Smaller molecules are able to diffuse more easily into the polymer and contribute to the weakening
of the polymer and potential cracking. So when you form your judgement on whether a solvent outside the
radius is safe, err on the side of caution if that solvent is small.
Stressing “stress”
It’s important to stress the “stress” part of ESC. At a high enough stress a polymer will crack even with a
poorly-compatible solvent. An excellent paper to illustrate this point is C. H. M. Jacques and M. G. Wyzgoski,
Prediction of Environmental Stress Cracking of Polycarbonate from Solubility Considerations, Journal of
Applied Polymer Science, 23, 1153-1166, 1979. PC parts were stressed to different extents and placed in test
solvents. In each case there was a critical stress above which the part would crack. For obviously good solvents
such as toluene, the critical stress was low (<0.3%), for obviously bad solvents such as ethylene glycol the
stress was high (>1.9%).
One way of looking at these data come from Hansen’s review of HSP and ESC, Charles M. Hansen, On
predicting environmental stress cracking in polymers, Polymer Degradation and Stability, 77, 2002, 43–53.
First the HSP sphere is calculated not on the basis of solubility but on how much strain is needed before the PC
cracks in solvent. In the re-work of that data (including some corrections to the data from other papers) the
Sphere comes out at a surprising [21.9, 10.2, 5.2, 13.8]. But remember that this is a Sphere based on ESC not
on solubility for which the radius, and presumably the HSP will be rather different. For reasons that will
become clear in a moment this is forced to be [21, 7.6, 4.4, 10.2] for further data analysis. Hansen then plotted
the RED number v Molar Volume for these solvents, dividing the plot into classes depending on whether they
required low, medium or high strains before cracking. Here is a simplified version of the original graph, using
the revised data:

77
Figure 13-3 RED v Molar Volume for the PC ESC data file
Not surprisingly, there is a general correlation between low RED number and low strains before cracking. But
what is also clear is that low molar volume solvents must be further away from the PC (i.e. higher RED
number) – in other words, for a given RED number, small molecules give more stress cracking.
In a graph (not shown) of RED v strain required to crack, there is a general trend (as expected) that high RED
requires high strain. But the fit is very poor. Instead, a fitted trend was created using a combination of RED and
Molar Volume. It turns out that the best fit comes from RED * MVol0.333. And the best fit was found with the
HSP values mentioned above.

Figure 13-4 PC Critical Strains % correlated with RED and MVol0.333


Clearly the correlation isn’t perfect. Nor should it be. For example a detailed analysis in the Jaques’ paper
shows (as we would expect) that branched hydrocarbons require a higher critical strain than unbranched
equivalents because (as discussed in the Diffusion chapter) branched solvents diffuse more slowly.
Similar plots can be produced from the % strain data of other polymers. For Polysulfone (not shown) there is a
similar good fit to a 0.333 dependency on MVol. However, for PMMA and PPO there are big differences:

78
Figure 13-5 PMMA Critical Strains % correlated with RED and MVol0.8

Figure 13-6 PPO Critical Strains % correlated with RED and MVol0
On the basis of these four polymers a tentative conclusion can be reached about the MVol effect. The Sphere
radii for the four fits are: PMMA 8, PC/PSF 10.5, PPO 13.5. It seems reasonable (though of course it is
unproven without considerably more experimental data) that for polymers with large radii (PPO), the ease of
molecular access is high so the MVol effects are small. Conversely, for polymers with smaller radii (PMMA)
there is a far greater specificity and therefore a larger dependence on MVol. For the ESC predictor described
below we have set up a MVol dependency varying from a minimum of 0 at 13 and above linearly up to a (fitted
and meaningless) value of 2.2 at a radius of 0. This means that a Sphere of radius 7 has a linear dependency on
MVol.
Readers will be unhappy with the tentative nature of the above. But those who wish to criticise should first
consider the amount of work required to get good data. First there is the basic experimental data. It is thankless
and dull work to get data over a sufficiently large range of solvents that span both the HSP space and also
MVol space. Only two of the correlations feature molecules with MVol > 150 so it is hard to get statistically
meaningful fits. Second there is comment from these authors that there is no substitute for the original data.
Some of the values for some of the polymers quoted in some secondary literatures are just plain wrong. It was a
sobering experience to spend the time entering data from the secondary literature, getting absurd plots, then
discovering that the data had been misquoted.
For old polymers there may be no reason to re-do the ESC data as the user community more or less know which
solvents to avoid for long-term use. But with so many new bio-polymers coming on the market it would seem a
79
good idea to invest the time and energy producing critical strain data. Although it is a lot of effort, compared to
the consequences of an unexpected failure out in the real world it would seem to be a good investment of
scientific time.
Because users of HSPiP wanted an ESC predictor we have provided what we call an ESC guide. Because ESC
depends on HSP, on MVol, on branched/unbranched solvent shapes, on stress and (sometimes) on surface stress
it’s simply not possible to provide a complete ESC predictor. Instead we’ve provided a colour-coded guide that
follows the rainbow. The colours are based on the combination of RED and MVolx, where the power x varies
from 2.2 to 0 between a radius of 0 and 13. The larger the Sphere the more solvents will be caught in the ESC
trap, so it’s important that you know what your sphere is based on. If it’s based on solubility then it will tend to
be on the small side. If it’s based on a critical stress value then it’s more likely to be realistic. Red and yellow
are for molecules usually well inside the Sphere. These will probably not cause ESC simply because they will
obviously damage the polymer in early tests. Such solvents have been known to give ESC, however, when they
have been found in aqueous mixtures even at very small concentrations. If the water can evaporate, the solvent
concentration can become very high, and ESC has been found. Greens will tend to be around the border so
should be prone to ESC. Blues will tend to be far-enough away to provide low ESC probabilities. But
remember that if the stress is large enough almost any solvent can cause ESC.

Figure 13-7 ESC colour-coding taking into account MVol as well as RED. Blue is typically safer than green. The data shown
are in the middle of the range. A GA fit was used as it was more compact.
It’s therefore up to you to use the colour guide as a guide.
To summarize: to avoid ESC make sure that the solvents used are well outside the Sphere and have a large
molar volume and, if possible, branched, low diffusion shape. If your polymer will encounter larger stresses
then the solvents will have to be further outside the sphere with an even larger molar volume. On the polymer
side it is known that incorporation of a smaller amount of higher molecular weight polymer (bimodal) vastly
improves stress cracking resistance, while still allowing for acceptable processing conditions. This is
reminiscent of the phenomena discussed early on adhesion and polymer chain entanglement. The intimate
mixing of polymers having widely different molecular weights is possible by simultaneous use of two different
catalysts for polyethylene, for example.
These theoretical considerations can be illuminated by some practical examples encountered by Hansen. ESC
with PC has been found on evaporating insulin solutions stabilized with 0.15% m-Cresol. There are also
examples of surfactants in automotive windscreen cleaners that have cracked PC parts on the car. Then there is
the classic of the stick-on label on a PC helmet giving ESC. An example of ESC with serious environmental

80
consequences arose several years after the gluing of a PVC joint. There was a little piece of a rag in the
threading that over time gave enough stress to initiate the crack. The PVC near the rag failed in a mode that
started with stringing polymer, which meant solvent in some quantity was present, and this then extended
through the whole joint and gave an environmental catastrophe since strong base in a large amount got into a
nearby stream. This was a large media event at the time, and all because a solvent wiping rag was able to
initiate ESC.

81
Chapter 14 Let’s make this perfectly clear … (Formulating clear
automotive lacquers)
If you are trying to make a hard, clear coating a good starting point would be an isocyanate cross-linked polyol
system. There are many isocyanates, many polyols and therefore it is a straightforward process to create a
formulation with the right balance of pot-life, cure speed, cross-link density, hardness, scratch resistance etc.
The final desirable feature, clarity/gloss, requires the whole cross-linking and drying process to take place
without any phase-separation of any of the components. This isn’t as easy as it sounds. It’s likely that a solvent
blend is used, for the right balance of solubility, drying rate, health and safety, as discussed previously. The
problem is that a solvent blend which might have been good for the starting isocyanate and polyol may not be
good for the polymer as it forms. Or even if that blend were OK, the fact that the different components
evaporate at different speeds means that the solubility parameters are changing, perhaps for the better or
perhaps for the worse.
Formulating such a coating by trial and error is therefore very hard work. As the authors know from personal
experience, formulating with the help of HSP makes the task much easier and, ultimately, more profitable.
In this book we use real-world examples as much as possible, but in this case we will merely be illustrative.
Readers should be able to extend the principle to their own situation.
For an isocyanate we’ll use TDI, Tolylene diisocyanate. For a polyol we will use an oligomeric hydroxyl-
propyl methacrylate. The HSP for TDI are known, [19.3, 7.9, 6.1] Using S-P group contributions (see the DIY
chapter for what S-P means) the oligomer can be estimated as [16, 10, 12]. The polymer can similarly be
estimated as [18, 11, 7]. {Although HSPiP has advanced since these crude estimates, the values are preserved
simply to keep the example constant after many years}.
In this simple analysis we can choose to ignore the issue of dissolving the TDI as it is a liquid at the presumed
25ºC for making our basic formulation. So our problem is to have a solvent which is excellent for the oligomer
and which becomes excellent for the polymer.
If we go to the solvent optimizer our first task is to find a “friendly” solvent with a sufficiently high δH without
it containing any –OH functionality which would interfere with the cross-linking reaction. We also need this to
be relatively volatile as we want the δH component to drop significantly during the curing/drying process.
By Ctrl-clicking on the solvents containing alcohols (or creating a “friendly solvent” list without alcohols) it is
relatively easy to find that a 69:31 mixture of PMA (Propyleneglycol Monomethylether Acetate) and GBL
(Gamma ButyroLactone) is an adequate fit, though not surprisingly the δH is rather lower than desirable. But as
it’s only an oligomer we are trying to dissolve, its radius will be rather large so this mix should be OK.
The HSP of that mixture are [16.3, 9.0, 9.1].
For our polymer we clearly need a significant increase in δD. The GBL is less volatile than PMA and has a high
δD so we’re off to a good start, but the δP is very high. So we need something to bring that down. DBE
(DiBasic Ester) is low boiling and as a 57:43 mix with GBL is quite close to our polymer at [17.2, 11.5, 7.8].
So now we need to work out a blend of PMA with DBE and GBL to get the best possible compromise. A
75:14:11 mix of PMA, GBL and DBE gives [16.0, 7.0, 9.3], a distance of 3.9 from the oligomer. This seems
adequate.
Now we need to check how the solvent blend changes with time. The relative evaporation rates are
PMA:GBL:DBE = 30:3:1. Within the Solvent Optimizer there’s the option to model the evaporation. Here’s the
output from this particular simulation:

82
Figure 14-1 Change in composition in time as the solvents evaporate. The setup can be loaded as PerfectlyClear.sof
At the end of this simulation, the total solvent (Blue data/graph) has decreased to 12% of its original amount.
The PMA (Red data/graph) decreased rapidly from 75 to <5% after a short time. Because cross-linking
reactions tend to slow down with time, we can guess that a good deal of the cross-linking has taken place in that
time, so the removal of the PMA by that time is about right. The PMA:GBL (Green):DBE (Magenta) ratio at
that point is approximately 4:49:47, giving [17.1, 10.6, 8.9] (the HSP of the mix is calculated for you in the
simulation) a respectable distance of 2 from the polymer. The ratio of GBL to DBE gradually shifted because of
the factor of 3 in their relative evaporation rates. Clearly the starting ratio should be changed to reflect this if
the resulting level doesn’t quite give the required compatibility with polymer.
The high level of the low boilers at the end of the curing reaction means that a final bake will be required, but
this would be required anyway as removal of the final % solvent becomes constrained by diffusion rather than
evaporation. The bake will also complete the cross-linking reaction.
Note that in this run we’ve chosen to include the effect of Activity Coefficients – when the presence of one
solvent increases the evaporation rate relative to another. To do this the Activity Coefficients are calculated
from the HSP distance between one solvent and another (or one solvent and the HSP of the rest of the blend).
The calculation is based on the “enthalpic deviation from ideality” term involving the χ parameter. For solvent-
solvent interactions one often-used formula is:
Equ. 14-1 lnγ = ln (φ1/x1) + HSPDistance2 φ22*0.6MVol1/RT
For solvent-polymer interactions (an option in the calculations) there is an additional term. To avoid over-
complicating the calculations, this second formula is used for all the interactions, on the grounds that the errors
in the assumptions will tend to balance out. Responsibility for the use of this method is Abbott’s.
Equ. 14-2 lnγ = ln (φ1/x1) + 1- φ1/x1 + HSPDistance2 φ22*0.6MVol1/RT
where φ is the volume fraction, x is the mole fraction and γ is the activity coefficient. With solvents that are
very close, γ approaches 1, for solvents that are far apart γ might rise to 2 or 3 and, at very low concentrations
might rise as high as 10. These formulae and a warning about the full complexity of multi-solvent/polymer
calculations are found in J. Holten-Andersen, C.M. Hansen, Solvent and Water Evaporation from Coatings,
Progress in Organic Coatings, 11, 219-240, 1983.
You can choose to include the Target polymer in the activity coefficient calculations.

83
You can even do the calculations at an elevated temperature. The relative evaporation rates are calculated by an
approximate fit (from Abbott: RER=0.046*MVol*VP) to the calculated vapour pressures using Antoine
Coefficients which are included for as many of the Optimizer Solvents as possible. This seems to give a better
fit over the range of data than a published method (Stratta, J. John, Dillon, Paul W., and Semp, Robert H.,
Evaporation of Organic Cosolvents from Water-borne Formulations. Theory, Experimental Study, Computer
Simulation, Journal of Coating Technology, 50, 1978, 39-47) which has RER=0.822*MWt ½*VP.
Equ. 14-3 RER = 0.046 MVol. VapourPressure (mm Hg)
At the same time, the temperature correction of HSP is also performed.
Here are the same 3 solvents at 50ºC instead of the nominal 25ºC of the previous plot:

Figure 14-2 The same system at 50ºC


Note that no meaningful Antoine Coefficients (AA, AB, AC) are available for DBE as it’s a mixture so the
values were taken from DPG which has a similar RER at room temperature.
It’s likely that you will disagree with us on the precise choice of solvent blend. And you’ll note that this
simulation says nothing about absolute time. It can’t – all it knows about is relative evaporation rates. The
modeller automatically adjusts itself so that the fall-off of solvent at the start isn’t too rapid or too slow and the
rest of the curves then fit around that. Only you know what air flows you are applying and whether the
temperature of the drying system is lowered by evaporative cooling. Both factors affect the absolute drying
time. (For the 3rd Edition we’ve let you do an absolute time calculation by specifying how long it would take
nBuAc to evaporate under your conditions.) But the point is that we reached our decision via a rational process,
with a “better than nothing” model and when our coating isn’t quite as glossy as we would like, or our solvent
blend is a bit too expensive, or the level of residual solvent is a bit too high, the formulation can be tweaked
rationally.
What would have been the alternative? We know the answer from real-world experience. The alternative is a lot
of trial and error, a lot of confusion, and the expenditure of a great deal of time and energy.
We have to make a really important point about the temperature corrections used for the optimization. You
notice that the HSP of all the solvents are changed by the temperature. This is mostly a thermal expansion effect
– the MVol increases with increasing temperature. Why did we not change the HSP of the target? If, for
example, the solvent blend is a good match at room temperature, it’s possible that it will be a bad match at a
higher temperature. Exactly so! In a series of important papers in the 1960s, Patterson repeatedly pointed out
the existence of the LCST “Lower Critical Solution Temperature”. This, confusingly (don’t blame us for this
mad nomenclature), is at a higher temperature than the UCST “Upper Critical Solution Temperature”. The
84
UCST comes from the 1/RT effect in the Chi parameter and means that polymers are less soluble at lower
temperatures. No surprises there. The LCST is less familiar – the fact that polymers (especially high MWt ones)
start to become less soluble at higher temperatures and eventually come out of solution. The explanation for this
non-intuitive effect lies exactly in the phenomenon we’re discussing. The HSP of the solvents generally
decrease with temperature because of thermal expansion. But the polymer expansion coefficient is much
smaller than that of the solvent so its HSP stays approximately constant. So a perfect HSP match at room
temperature can become a large mismatch at high temperatures.
Don’t blush
This is a good point at which to introduce the new Blushing Estimator in the 3rd Edition. A solvent mix which
gives an excellent coating on a dry day might give a horrible coating on a day with high relative humidity (RH).
Or it might be OK if you dry slowly but horrible if you increase the air flow to dry a bit faster. The reason is
that the evaporation of the solvent removes heat from the solution, causing it to cool. If the surface temperature
falls below the dew point of the air then water can condense out, causing the coating defect often called
“blushing”. The dew point increases on days with high RH, and a faster air flow increases the rate of
evaporation and therefore lowers the surface temperature. Either of these effects might tip the balance between
a good and a blushed coating.
If you increase the airflow to the coating you increase the rate of evaporation, which cools the coating further,
which then decreases the rate of evaporation. Clearly there is a point at which the two effects balance out. This
is the so-called Wet Bulb temperature of the solution. If there’s sufficient heat coming in (e.g. via the substrate)
or if the air flow is not high enough, then the solution may be above the wet bulb temperature. But it will never
be below it. The web bulb temperature is the worse-case scenario.
HSPiP has all the necessary data for the calculation. From the HSP we know (by definition!) the enthalpy of
vapourisation. We also provide the Antoine Coefficients for a large range of solvents – and Y-MB provides
estimates for other solvents. The Antoine Coefficients give the vapour pressures of the various components.
Finally the MVol is needed for knowledge of the actual (as opposed to molar) amount of vapour in the air.
The wet bulb temperature is therefore automatically calculated for any mix where all the data are known. By
entering the RH of the air we can also calculate the dew point of that air. If the dew point is higher than the web
bulb then the wet bulb value is surrounded by red to give you a visual warning of the possibility of blushing.
Only you can tell whether the air flow is low enough or the supply of heat from other sources is high enough to
overcome the blushing. But at least you have a quick way to check whether you are at risk from blushing.

85
Figure 14-3 This 67:33 blend of s-butyl acetate and benzyl alcohol has a wet-bulb temperature below the current dew point of
18.0 degrees – so there is a danger of blushing. Swapping the ratio of the solvents or reducing the RH to 60% (not shown)
removes the possibility of blushing.

86
Chapter 15 That’s swell (HSP and Swelling)
If you look in the polymer data table you can find the same polymer giving different values. This can be deeply
troubling to a first-time user and seems to undermine the whole premise of HSP.
Further thought reveals that some polymers must show different HSP under different conditions. This is such an
important principle that we give it a chapter to itself.
For first-time users we provide an “Instant Guide” set of values for common polymers. Until you’ve built up
your own experience, use these values for the polymers – but remember that they are only there as an instant
guide and that other values might apply for your specific problem, as this chapter emphasises.
Let’s take for example PCTFE, PolyChloroTriFluoroEthylene.
If you calculate a sphere using data from solvents that swell it by >2% you get (though with so few good
solvents, the fit is somewhat arbitrary) [17.8, 1.9, 2.7], typical of a C-Cl polymer:

Figure 15-1 A correlation with PCTFE swelling at 2% absorption


But if you do a plot with those solvents that swell by >5% you also get a very different good fit, typical of C-F
polymers [16, 5.5, 7.0].

87
Figure 15-2 Same polymer, different correlation at 5% absorption
What’s happening is that at low levels of solvent absorption, the solvents associate themselves with the –Cl rich
areas of the polymer. As you go to greater swelling, the solvents have to associate with the predominant C-F
regions.
This must be a general principle. If you test a polymer which contains a small portion of –OH functionality then
at low levels of swelling, alcohols will be very happy to be associated with these regions, so the solvent sphere
is biased towards the alcohol region. But when you start swelling/dissolving the whole of the polymer, the
alcohols are very poor solvents, so the sphere shifts towards a lower δH and δP region.
Similarly, if a polymer contains crystalline and non-crystalline regions, then swelling data at low levels of
solvent will reflect the non-crystalline region and therefore a bias towards whatever functionalities
preferentially reside in that region.
So we can now flip the problem of having different solvent spheres into a distinct advantage. If you find
conflicts in the data, these may well be providing you with fundamental insights into the internal structure of
the polymer. It’s not obvious that PCTFE should have chlorine-rich and fluorine-rich regions, but the HSP data
seem to suggest that that is the case.
The same principles can be applied to the latest nano-scale issues. It is becoming common practice to e-beam
write nanostructures for integrated circuits, photonic crystals and nanobiology. When “negative” resists are
used (i.e. those that become less soluble on exposure) there is a problem of development. You want a solvent
that quickly whisks away the un-crosslinked resin. But such a solvent can readily enter the cross-linked
polymer and cause it to swell. If you write 10nm features, then it only needs swelling of 5nm across both sides
of the feature and the swollen polymers touch across the divide and degrade the quality of the image. One
proposal to fix this is to use solvents just at the edge of the HSP sphere – they will still dissolve the un-
crosslinked resin, but will be unlikely to enter the crosslinked system. We are grateful to Dr Deirdre Olynick
and her team at Lawrence Berkeley National Laboratory for allowing us to reproduce data from their paper that
explores these issues in a profound way: Deirdre L. Olynick, Paul D. Ashby, Mark D. Lewis, Timothy Jen,
Haoren Lu, J. Alexander Liddle, Weilun Chao, The Link Between Nanoscale Feature Development in a
Negative Resist and the Hansen Solubility Sphere, Journal of Polymer Science: Part B: Polymer Physics, Vol.
47, 2091–2105 (2009).
The team first established the HSP sphere for the calixarene resist of interest.

88
Figure 15-3 Sphere for Calixarene e-beam resist
It is interesting to note that they used a sophisticated Sphere algorithm (fully described in their paper) which
included some heuristics that could eliminate false fits. Happily, the values of our straightforward algorithm
match theirs. They were then able to show that solvents closer to the centre of the sphere were better at creating
high contrast images, whilst those near the edge were better at avoiding the problems caused by swelling. A
rational compromise can then be reached on this basis. Importantly, other solvents and/or solvent blends can
then easily be devised on rational principles to improve the process even further. The paper contains much more
of interest and readers are recommended to explore their paper in detail.
Of course kinetics must be part of the optimisation process and it is likely that issues discussed in the Diffusion
chapter will also play a part in understanding. But by establishing the basic thermodynamics of the system,
further optimization can be a more rational process.

89
Chapter 16 Paint failure – the science of blistering
Water blistering in polymeric coatings generally requires the presence of water locally within the film in an
amount close to its saturation solubility. The films swell because of the absorbed water and there is an increase
in the compressive stresses as emphasized by Brunt from the TNO in the Netherlands in the 1960’s.
Hydrophilic components can collect water and initiate blisters. If the adhesion is poor, the blisters that form can
remove the coating from the substrate, either isolated as blisters, or by total delamination of the coating.
Isolated water filled blisters are usually formed in softer coatings, remembering that water has a significant
plasticizing effect, whereas more rigid ones tend to delaminate, not being able to yield enough to accommodate
local blisters. Both of these types of failure are called blisters here. Once formed, blisters can grow during
continued water exposure, for example, by osmotic effects. There are more subtle mechanisms of blister
formation, however, and hydrophilic components are not necessary. The solubility relations of water in the
polymer in question are important, not just at room temperature but also as they are affected by changes in
temperature. The δH parameter changes more rapidly than the other HSP, and the HSP for water approach those
of the polymer more closely as temperature increases so the water solubility increases. As discussed in the
following, this can lead to water blisters if the temperature falls rapidly. In order to minimize or prevent
blistering at substrates the adhesion must be such that water at the interface cannot cause local loss of adhesion.
Anchor-type adhesion or covalent bonding to a substrate is recommended if possible. Physical bonds across an
interface are not nearly so resistant (see the chapter on adhesion). Under special conditions of rapid temperature
changes it is possible for blisters to occur in the middle of polymer films, or even near the air surface, as
described in detail below. This mechanism is responsible for a potential problem of excess water within a
polymer be they elastomers or rigid plastics as described in the following. Water blistering can potentially
occur in films applied to any substrate, but the majority of the practical problems are found at interfaces for
coated metals and wood. The following general cases will be discussed:
• Cause 1: Presence of hydrophilic components
• Cause 2: Substrate below the local dew point
• Cause 3: Rapid temperature changes leading to “fog”
• Cause 4: Inverted primer (normally higher equilibrium water uptake) and topcoat (normally lower water
uptake)
The problem of the whitening of restored paintings is discussed in the context of blister formation in coatings,
even though the “blisters” remain very small, more like a fog. A final section discusses methods to alleviate the
blistering problem.
Cause 1: Presence of hydrophilic components
It has generally been recognized in the coatings literature in the 1960’s and later that the presence of
hydrophilic components could lead to blisters. Water-soluble components of pigments and fillers, hydrophilic
pigments, and salts have been cited. A worst case scenario is the presence of salts in the film or at a substrate.
Sodium chloride, for example, collects water at a relative humidity of 75% or above. Water molecules diffuse
into all coatings (and plastics) at some relatively rapid rate compared to larger molecules. The rate is faster for
films at temperatures above the glass transition temperature, and the amount of water at saturation also
generally increases with increasing temperature as mentioned above. It is only a matter of time before a given
film becomes saturated when it is in contact with liquid water or saturated steam. It should also be recognized
that even on exposure to normal air, there will be a significant amount of water in the film at equilibrium with
the water at some relative humidity in the air. Since this mechanism is fairly obvious, it will not be discussed in
further detail, other than to point out that the hydrophilic sites may include substrate factors such as weathered
wood, rust, or other effect.
Cause 2: Substrate below the local dew point
The satisfactory coating of cold water pipes has always been a problem of some significance, particularly in
warm and humid climates. The condensation of water at or near the pipe can only be delayed by most coatings,
and the anchor adhesion mechanism or covalent bonding, if possible, are suggested for best results. The
blistering of cars can be a result of the same mechanism. On those days where water drips from under carports
and in sheds, the cold metal under the coating on a car can also cause condensation of water from the
increasingly warmer and more humid air present as the day grows older. There is a balance between how
90
quickly the metal can rise to a temperature above the dew point, and how quickly the water can absorb into the
film and diffuse in sufficient amount to the substrate. There are clearly times when the water gets there first to
form blisters, even though temperature change generally occurs more rapidly than water transport.
Cause 3: Rapid temperature changes leading to “fog”
When a film that is saturated, or nearly saturated, at a higher temperature is cooled rapidly, water can remain in
the film in excess of that soluble at the lower temperature. This water precipitates, much like fog. If the film is
sufficiently resistant it may recover after this water finally escapes to the air. If there are hydrophilic
components, the water will preferentially collect at such sites, and blisters are nucleated. The blisters can then
grow on subsequent temperature cycling or because of osmotic effects. The testing of coatings in so-called
blister boxes involves a combination of the causes 2 and 3. The sample is placed at an angle to the horizontal so
that the condensed water can run off. The substrate will be colder than the interior of the cabinet invoking cause
2. At the same time the water periodically running off will induce local temperature changes that presumably
enhance the severity of the test method. An exceptionally severe test of this kind involves putting the films on
top of a container in which there is boiling water.
As shown in Charles M. Hansen, New developments in corrosion and blister formation in coatings, Progress in
Organic Coatings, Vol 26, 113-120, 1995, blisters were formed near the air surface of epoxy-coated thick steel
panels during an attempt to measure the diffusion coefficients for methanol in the coating at 50°C. The panels
were removed from the methanol bath and weighed at room temperature. It only took a few cycles before
sizeable methanol blisters near the air surface were formed. The methanol absorbed near the surface would be
near the saturation value at 50°C, whereas there may still not be methanol at the metal substrate. This
concentration of methanol exceeded what the surface region of the film could truly dissolve upon its removal
from the methanol bath. Blisters formed and grew on subsequent cycling.
Two other situations exemplifying this cause are cited in the Handbook on pages 238-240. These are excess
water in free films of EPDM rubber and in poly(phenylene sulfide) (PPS). The temperature cycling with water
exposure for the EPDM was from 120°C to 15°C simulating a problem of a failure in a gasket subjected to
decontaminating steam with subsequent hosing in a dairy. The PPS study was to demonstrate that even such
rigid polymers could be made to fail by this mechanism. Here the temperature cycling with water exposure was
between 90°C and 23°C using 2 mm films. Normal absorption curves were found initially in both cases, but late
in the approach to equilibrium a more rapid water uptake was suddenly encountered as the polymers started to
have the excess water problem. Control experiments at the higher temperature did not show uptake of excess
water when the samples were left in the test for very long time periods. It took 5 days to rupture the EPDM
gasket in the middle. The excess water started appearing in the PPS film after about 40 days.
Cause 4: Inverted primer and topcoat systems
Blisters are often encountered after repairing older paint, even shortly after the repair work has been completed.
The work in Klaus Lampe and Charles M. Hansen, Blæredannelse i malingfilm (Blistering in Paint Films),
Rapport T 16-83 M, Nordisk Forskningsinstut for Maling og Trykfarver (Scandinavian Paint and Printing Ink
Research Institute), Hørsholm, 1983, 58 pages, helps explain how this can happen. Even though these studies
were on metal substrates, the results are still applicable to coatings on wood.
Coatings were cycled between 40°C and 12°C. The cycled coatings blistered in half the time required for the
non-cycled systems that were held at a constant 40°C. The blistered coatings had absorbed about 5%w water at
the time of blistering. The non-cycled systems blistered when the primer became saturated with water, which
required about 8%w (40°C). A topcoat with low water uptake and low permeability prolongs the time for the
blisters to occur. Longer periods of room temperature drying after oven cure improved blister resistance, since
the films were cured more thoroughly. Clear films had an initial milky appearance with blisters appearing later.
This is a manifestation of the “fog” discussed above. One coating had the same water uptake at three different
(40°C, 23°C, 12°C) temperatures. This coating could not be made to blister. The individual layers in these
systems were between 25 and 50 microns as recommended by the suppliers.
A usual topcoat was applied as a primer and a usual primer was applied as a topcoat over this to see the effect
of what might happen in a faulty repair situation. Blisters appeared rapidly at the substrate in such a “repair”
coating with subsequent rusting. This occurred as a rule when the equilibrium water uptake in the topcoat was
larger than that in the primer. The water in the lower layer could not escape rapidly enough in unfavorable

91
situations such as rapid cooling, and blisters at the substrate were common. The only safe practice when in
doubt is to remove the old paint.
To sum up this section, it can be concluded that blister formation is favored when more water can be taken up at
equilibrium in a topcoat than in a primer. Such conditions can easily be found in repair coatings that are not
tuned to the given paint to be repaired.
Whitening of restored paintings
Older repaired paintings occasionally develop whiteness at the places where they have been repaired. The
reasons for this are thought to be based on the same mechanisms as those described above. There are examples
of closed storerooms where the climate is not controlled and there are examples of paintings on cold walls. In
every case the cause of the whiteness is condensed, fog-like water droplets. The water droplets have forced the
repair paint apart and upon ultimate evaporation there are small holes where the water once resided. These
holes have light scattering properties with whiteness being the visual result. This phenomenon would not even
be recognized in a traditional white coating on steel or wood unless the water contact lasted long enough to
produce true, water-filled blisters or delamination. For colored paints it might lead to pastel color version of the
original color.
The things to think about are maintaining a more stable climate and making sure that the repair paint has lower
water solubility at equilibrium than the original paint. This implies repair coatings that would be characterized
as hydrophobic in nature, and that do not contain hydrophilic entities.
Discussion
The cases and mechanisms described above are helpful in understanding some undocumented observations
made by those in the coatings industry:
Why is pure water more severe in attacking coatings than salt water?
The answer is that the salt reduces the water activity, and the coating absorbs water at equilibrium with this
activity. The water content in the coating only approaches the total water saturation possible when in contact
with pure water, and there is excess capacity to truly dissolve the water freed during temperature cycling with
water exposure.
Why do panels perform better when left alone for the full test period in water exposure tests?
The impatient formulator (or boss) who repeatedly removes panels from such testing causes a temperature
change every time the panel is removed. This is particularly important for higher water temperature tests. This
phenomenon should also be remembered in any cyclic testing procedure with changes between exposure to
water, “sunshine”, and dry periods, with or without temperature change.
Why are there more problems with blistering with darker colors?
Darker colors have higher temperatures than lighter ones on sunny days. The larger temperature change on
rapid cooling (clouds or night) after a moist and sunny period creates a larger amount of water that is in excess
of that soluble at the lower temperatures. If this water cannot diffuse rapidly from a film, blisters will form. The
higher temperature also leads to softer films that are not as resistant to mechanical effects (stresses) that can
lead to loss of adhesion.
The temperature changes required for the formation of blisters in a water saturated film are not large. Common
margins of control over the temperature changes in a hot water bath (perhaps +/- 1°C) are large enough to
induce the effect.
Why is a hydrophilic topcoat the worst possible case? The usual topcoat/primer systems have the topcoat with
less water solubility. Brunt emphasized that there has to be swelling stresses to cause the blisters (and then poor
adhesion). In the usual topcoat/primer systems the swelling in the primer produces tensile stresses in the topcoat
that is being pulled apart by the swelling beneath it. The tensile stresses produce a resistance against blistering.
A 2% linear swelling by water is said by Brunt to be a minimum condition for blistering (though, of course, the
degree of adhesion will affect the resistance to blistering). This would have to be a differential swelling
between the topcoat and primer. The Hansen studies could not blister a coating system with equal water uptake
in topcoat and primer. One can blister primers in the conventional systems, but here the temperature cycling
comes in, and presumably not the swelling differential between the coatings. The excess phase separated water
92
in the primer, probably collected at hydrophilic sites, produces swelling stresses that can lift the whole system
from the substrate, again this being initiated at points of weak adhesion to the substrate.
How to minimize blistering
There are a number of factors that can help to minimize blistering. The mechanisms above explain how these
function.
• Do not apply a coating with high equilibrium water solubility onto one with low equilibrium water
solubility.
• Use anchor adhesion (pretreatments such as zinc phosphate) or covalent bonding to the substrate if
possible. Epoxy coatings may simply delaminate without blisters if the adhesion is not suitable under
conditions that otherwise would form local, water-filled blisters..
• Avoid hydrophilic components within the coating or contamination at a substrate.
• If a coating has the same equilibrium water uptake at the different temperatures of its use, it will
presumably not blister. Just how to create such a coating is not known to the authors, but the ceramics
industry was able to create products that did not change dimension with temperature, so why should the
coatings industry not be able to something similar?
The use of thicker coatings will delay the onset of blistering, all else being equal. This is not cost effective, and
the problem is not solved, although the external conditions may change for the better if the delay is long
enough. It is also conceivable that a particularly thick, water-saturated coating will not be able to lose water fast
enough on a rapid cooling cycle, and blisters would then form near the substrate, since this is where the water
content remains high for the longest time. Delaying blistering is also possible by increasing the diffusion
resistance in the coating. A topcoat with low permeability over a primer with high water solubility will extend
the “safe” period. The film will be able to restore a normal condition when the unfavorable water exposure is no
longer present. A creative suggestion is to include suitable holes in the coating by a controlled mechanism.
Excess water will be able to be accommodated in such cases, at least up to a given amount.
Conclusion
Blisters have been all too common in many coatings. Among the major causes are cold substrates below the
dew point of moist and warmer air. This condition is common after cold nights in the Spring and in the Fall. A
similar situation exists for cold water pipes in warm and moist climates, and special measures must be taken to
improve adhesion. A rapid decrease in the temperature of essentially water-saturated polymers can also lead to
excess water in the bulk of the polymer since equilibrium water solubility is generally lower at lower
temperatures. Excess water is precipitated like fog in the films. This can collect into blisters in weaker films and
even cause delamination in more rigid coatings. This is particularly problematic for repair coatings when a
repair topcoat with high water equilibrium solubility is used on top of a primer or previous coating with lower
equilibrium water solubility. The water in the primer cannot escape readily in the event of a rapid decrease in
temperature since there is too much water to remove from the layer above it with the time allowed to avoid
blisters.

93
Chapter 17 Skin deep (HSP and Skin Absorption)
It’s really rather important to know if a chemical will or will not penetrate into the skin. It can be a matter of
health & safety or it can be a matter of beauty (even if that is only skin deep).
With strong pressure to reducing the amount of testing on animals, it’s important to find alternative ways to
screen chemicals for their ability to penetrate the skin. Having a high scientific confidence that a chemical can
or cannot thermodynamically penetrate the skin can reduce the need to test. If you are confident that it cannot
penetrate then it is unlikely to be a useful beauty aid, but it is equally unlikely to pose a general safety hazard
by skin penetration. This approach is not enough on its own to guarantee efficacy or safety, but it is a very good
starting point.
One possible approach is to find directly the HSP of skin. Clearly this is impossible as skin is a complex multi-
component system. But a reasonable starting point for such measurements is abundantly available – psoriasis
scales. (see Hansen, C.M., and Andersen, B.H., The Affinities of Organic Solvents in Biological Systems, Amer.
Ind. Hyg. Assoc. J., 49, No. 6, 301-308 (1988)). The usual multi-solvent test, using swelling/non-swelling as a
criterion, gives δP & δH values (9.3, 15.6) which are credible and a δD value (>25) which is not.

Figure 17-1 Using file Psoriasis


The reason for this large D and large radius is unknown, but we present the data for you to reach your own
conclusions.
The next approach is to do permeation tests on real skin samples. (See Ursin, C., Hansen, C.M., Van Dyk, J.W.,
Jensen, P.O., Christensen, I.J., and Ebbehoej, J., Permeability of Commercial Solvents through Living Human
Skin, American Industrial Hygiene Association Journal, Vol. 56, 651-660 (1995)). By choosing a suitable range
of solvents it could be possible to see if there is a sensible correlation with HSP, even though we know that
permeation rates also depend on molar volume and shape. Such a correlation is shown below, with the
definition of “good” solvents being those having rapid permeation rates (they are Dimethyl sulfoxide (DMSO),
Dimethyl formamide (DMF), Dimethyl acetamide (DNAc), and N-methyl-2-pyrrolidinone (NMP)), which, for
humans, is “bad”:

94
Figure 17-2 Using file Skin
The values are [17.6, 13.5, 10.2, 4.3] which are reasonable. The Radius is a rather small 4 which means that
rapid skin penetration is quite restricted – which seems to be another triumph for evolution as a large R would
make us rather too susceptible to harm via our skin.
Suppose you wanted a reasonable polymer model for skin penetration. Clearly you need a polymer with HSP
close to that value. In the software you can find one easily. Enter these parameters as the final row in the
Polymers list. Now double-click on that row. The software finds the optimum match (best mutual overlap of the
two radii) between Skin and the other polymers. High in that match is polyurethane. So if you want to test
chemicals for skin penetration, have a go first with a polyurethane. Not surprisingly, the test industry has found
that polyurethane can be a useful test substitute for skin. One example involved tests of adhesion to skin.
Testers preferred to have the adhesive stick to the polyurethane rather than pull out the hairs from their skin.
Now let’s set ourselves the target of protecting our skin from a solvent or chemical which from HSP criteria
would be likely to permeate and which from chemical intuition might be likely to be harmful. By loading the
full Solvent Sphere Data set and then entering the Skin parameters as an additional row, the double-click trick
gives a list of chemicals that match skin closely. From that list, something like Ethyl Isothiocyanate sounds like
something you wouldn’t want to get into you via your skin. Select that row.
Now go to the Polymer form and click the Polymer button. This automatically finds the best match to the
selected solvent. Polyurethanes are near the top so they should not be used for protection as they afford little
added protection against chemicals, even though comfort may be at a maximum. Now go to the bottom of the
table to find the worst match. Of the practical polymers that can be used for gloves there’s no surprise to find
that simple PE or PP gloves should be more than adequate to protect you from this chemical, though they are
often too stiff for comfort and use. The ubiquitous Nitrile glove is also in this region.
Whilst we’re on the subject of gloves, there’s an interesting data set which seems, at first sight, to undermine
HSP thinking on permeability. Here is the entire data set for 1hr breakthrough times for Neoprene.

95
Figure 17-3 A mindless correlation with Neoprene gives very bad results
The fit is awful and out of 66 solvents, 13 of them are “wrong”. So does this mean that HSP are useless?
Any scientific tool used without thought can lead you astray. Arguably the bad fit above is a mindless use of
HSP. Because this correlation is based on breakthrough times which in turn depend on diffusion coefficients,
it’s obvious that molar volume must play an effect. In this case, Kinetics must be significant, whereas HSP, as
stated at the start, assumes pure Thermodynamics.
So let’s try to exclude gross kinetic effects from the calculation. Let’s exclude all very small molecules, say
below a molar volume of 70. And let’s exclude all very large molecules, say, above a molar volume of 200.
What do we then find?

Figure 17-4 A fit using more intelligence over molar volume effects
Now we get a perfect fit with no “wrong” solvents. It’s probable that this is a pretty good set of HSP for
Neoprene. But how valid is what we’ve just done?
96
Only you can decide according to your own application. Fortunately, HSPiP allows you to play “what ifs”. If
you happen to know that there is a very good reason for excluding methanol, but not the other small molar
volume molecules, you can try another fit. Or, as it turns out, the real problem is the apparent slow diffusion of
the largest molecules within the film. Excluding those, on the reasonable grounds that the lack of permeability
was a size effect, gives a value not too different from the one above.
So using HSP involves some judgement by the user. But that’s no bad thing and at least one’s judgements, as in
the above “good” fit, are made explicit. And there is a really important point that should be emphasised. If a
large solvent has a RED>1 you can be reasonably certain that it will not permeate. If a small solvent has a
RED<1 you can be reasonably certain that it will permeate. A small solvent with a RED>1 is problematical, but
you would be wise to assume that it will get through. A large solvent with a RED<1 is likely to permeate if you
give it a longer exposure time, so if you are serious about glove protection you would be cautious about such
solvents. In other words, even though we have admitted some margin of error, there are still quite a lot of things
you can conclude with some degree of certainty. And as ever, consider the alternative. If you didn’t have the
HSP approach, how would you be able to form any sort of scientific judgement?
The lipid-only bricks and mortar myth
The famous Potts and Guy correlation between skin permeability coefficient and LogKow (the octanol/water
partition coefficient) + Molecular Weight has produced conclusions entirely opposite to the ones above. It
seems self-evident to many in the skin permeation community that chemicals pass through the lipid layer, with
the corneocytes being mere passive bricks.
It is slowly becoming clear that this lipid-only route is a myth. This is not to criticise the Potts and Guy paper.
It’s simply that it’s quite straightforward to produce a similar graph (discussed below) “proving” a strong
correlation with water solubility. If this graph had been published instead of the LogKow graph then a water-
only myth might have sprung up instead.
The truth is much more interesting than either myth.
At the heart of the mythology is a split into two different worlds. The first one, represented by the HSP work
above (itself driven by issues of chemical safety), is concerned with the migration of small molecule liquids
(e.g. solvents, acrylates…) into skin. These are often presented to the skin as pure liquids and the fact that they
often swell the skin is good evidence that the skin really does have high values for δP and δH, totally untypical
of a lipid-only route.
The second world is one where the pharma industry need to get large molecules to go through the skin. The
permeation rates are painfully small. “Permeation enhancers” are often used to speed things up, giving, often, a
factor of 2-5 in Jmax, the maximum flux. The literature on permeation enhancers is highly confusing. Because
the lipid-only route is seen as objectively true, then lipophilic molecules should be excellent enhancers. But by
far the best enhancers are water, ethanol, DMSO, NMP etc. Their enhancement is explained away as being
pathological. Ethanol, for example, is often said to be a “lipid extractor” which does the enhancement by
ripping out the lipids, allowing the permeant to go through a destroyed skin. This happens to be nonsense.
Ethanol is a useless lipid extractor, as you would expect from its large HSP distance from typical lipids. Some
classic enhancers such as Azone have a lipid tail, but their HSP are not particularly low. The terpenes can
sometimes be good enhancers – and are seized on as proof of the lipid route. But sadly the evidence shows that
usually they are pretty useless unless combined with other elements of the formulation such as ethanol.
If you read the skin permeation literature it is quite painful to see authors trying to explain results in crude terms
such as “lipophilic” or “hydrophilic”. The term lipophilic is a cause of much confusion. A good way to think
about it is to ask “What is the opposite to lipophilic?” A typical answer is “Hydrophilic”. But next ask “What is
the opposite to hydrophilic?” The answer is ambiguous. It can be “Hydrophobic” or it can be “Oleophilic”.
What is missing from the discussion is the key fact that octanol is not particularly hydrophobic – water is
soluble to 20% in octanol. The Potts and Guy correlation is not with a Octane/Water partition coefficient which
is a much stronger hydrophobic/hydrophilic measure.
Another key fact often missing from the debate is that the lipid bilayer contains ~25-30% cholesterol. The HSP
of cholesterol [20.4, 2.8, 9.4] mean that it insoluble in both ethanol and hexane, but a 50:50 mix is not a bad
solvent (though the δD match is bad). It is really more comfortable in, say, chloroform or 1,4 dioxane than in
any typical “lipophilic” solvent. So the heart of the lipid bilayer itself is not particularly lipophilic.
97
One clue to resolving this situation is mentioned above. The terpenes often give good permeation enhancement
when used with drugs formulated in ethanol. It’s often difficult to untangle precise quantities being used, but it
turns out that many formulations contain approximately equal quantities of terpene and ethanol. If you do a
quick HSP check, it’s no surprise to see that the HSP of the vehicle is close to that of skin. In particular, the
cyclic structures of the terpenes provide a boost to δD necessary to get up to the skin value.
Once the skin permeation literature is approached as an exercise in solubility a lot of things start to make sense.
The behaviour of many permeation enhancers looks a lot more explicable in this light. But let’s go further.
Let’s think through all those pharma molecules.
The Potts and Guy correlation is with permeation coefficient. But we don’t actually care much about that
coefficient. Instead we are interested in Jmax, the maximum flux.
A paper by Sheree Cross’s group in the University of Queensland provides a large dataset of Jmax values:
Beatrice M. Magnusson, Yuri G. Anissimov, Sheree E. Cross, and Michael S. Roberts, Molecular Size as the
Main Determinant of Solute Maximum Flux Across the Skin, The Journal Of Investigative Dermatology, 122,
2004, 993 –999. We are grateful to Dr Cross for providing us with the full dataset to do our own analysis. An
important sentence near the start of the paper neatly summarises our own view of the misdirection caused by
the emphasis on permeability coefficients: “In practice, it is the maximum flux (Jmax) of a solute that is of most
interest in determining the maximal dermal, toxic, or systemic effect. Almost all studies concerned with
predicting skin permeability have focused on skin permeability coefficients from aqueous solution.”
At first sight their title is discouraging. It doesn’t mention solubility at all. But one of the key findings from the
paper is that LogKow is not a useful predictor. Our own plot of the full dataset shows a slightly negative bias
with an r2 of 0.09

Figure 17-5 The non-correlation between Jmax and LogKow

One plot which shows an interesting correlation (but nowhere near as good as the molecular size which they
rightly concluded was the strongest single correlation) show Jmax versus LogS, the water solubility.

98
Figure 17-6 Jmax correlates much better with water solubility!
The more water soluble the compound, the higher the Jmax! This is the plot we mentioned earlier which could
have launched an alternative (and equally mythological) “hydrophilic only” skin permeability hypothesis.
Dr Cross’s team were able to improve their correlation somewhat by including melting point. Why should this
be so? Because melting point, as discussed in detail in the Predictions chapter, has a major influence on
solubility.
Although fitting large data sets like this one is fraught with dangers, it seems to us that the most sensible way to
look at the data is by saying that Jmax is simply Solubility * Diffusion Coefficient/Skin_Thickness. We can’t do
anything about skin thickness so we should focus on the other two factors. We know that Diffusion Coefficient
depends on molecular shape, for which Molar Volume is the best surrogate approximation. That will explain
the big theme of Cross’ paper. Solubility can be predicted from first principles by the following equation:
Equ. 17-1 ln(Solubility)= -C + E –A - H
where C is a term based on the melting point (and is 0 for a liquid), E is the “combinatorial entropy” term
which we can ignore given the margin of error in all these data, A is an Activity Coefficient term and H is a
Hydrophobic effect term when considering solutions in water and alcohols. This too can be ignored in this
simple analysis.
The C term can be estimated directly from the melting point using the Yalkowsky simplified expression:
Equ. 17-2 C = -0.023*(Tm –T)
(The pre-factor of -0.023 comes from a very recent Yalkowsky paper though it is quoted there as -0.01 as the
paper uses Log10 instead of our Ln)
A can be estimated by the HSP distance between the chemical and skin.
We can therefore simply estimate Solubility in the (outer layer of the) skin and also use a simple power-law
dependence of Diffusion Coefficient based on the molar volume. That’s quite a few fitting parameters, but
here’s the fit:

99
Figure 17-7 A “first principles” fit to Jmax. Much more encouraging. Note that the previous correlation with water solubility
is mostly because water solubility contains a large element of the “crystal” term.
That’s not a bad fit given such crude approximations. Inspection of the individual components of the fit show
the following:
• A power of 0.5 describes the molar volume effect on diffusivity
• A “skin” value of [17, 8, 8] best provides the HSP term
• The Crystal and Diffusion terms fight for dominance in a manner that makes intuitive sense from the
calculation. In other words, for each molecule we can rationally decide which component is most
important in deciding the Jmax.
• The HSP activity coefficient correction is often small because many pharma molecules tend to be in the
region of [17, 8, 8] (not surprisingly because they have to be generally compatible with the biological
system). But the influence can be decisive for the non-pharma chemicals of concern to the general skin
permeation community or those concerned with cosmetics which use large concentrations of relatively
simple/small molecules.
The fact that HSP don’t make dramatic differences across the log-log plot does not mean that HSP can be
ignored. Remember that permeation enhancers tend to deliver factors of 2-5 in enhancement. These are trivial
in a log-log plot but can make all the difference for a real drug. Because the crystalline (melting-point) term can
be so significant, it’s not surprising that HSP cannot make a big difference for big pharma molecules. But by
making sure that the vehicle (e.g. 50:50 ethanol:terpene) matches both the skin and the pharma molecule, the
all-important factors of 2-5 can come in to play.
The HSPiP diffusion modeller (see the next chapter) can readily simulate all these effects. In particular there are
two competing issues. The first is concentration dependent diffusion coefficients in the outer layers of the skin.
The more, say, ethanol/terpene there is, the higher the diffusion coefficient. The second relates to the fact that
the permeation enhancers themselves get depleted across the skin. This reverse gradient can be approximated in
the modeller. Here is a typical example modelling some real skin diffusion data and providing a quite
satisfactory estimate of 21 minutes “lag time” (a key parameter in many skin studies) from the magenta
extrapolation of the red flux curve.

100
Figure 17-8 The diffusion modeller coping both with concentration dependent diffusion and a reverse gradient, whilst
providing a classic “lag time” fit to the integrated flux.

To remove all doubt, we are not saying that the above correlation proves that the solubility/diffusion approach
is the best way forward. But we are saying that the LogKow seems to have produced only a confused literature
that is focussing on the wrong parameter (permeability coefficient) rather than the right one (Jmax). And we are
not claiming that HSP are dominant in skin permeability of pharma compounds. But we are saying that the
HSP provide key insights into the modest effects of skin permeation enhancers and provide a coherent,
numerical language for thinking through the effects of the mixtures of components generally found in skin
formulations such as ethanol/terpene mixes and in liquid-based formulations found in cosmetics, fragrances etc.
Formulating For Efficacy
By one of those lucky breaks that makes science so interesting, Abbott encountered Prof Johann Wiechers, a
well-known expert in the cosmetics industry. Johann had been dissatisfied with the “bricks and mortar” model
and came up with an approach he called Formulating For Efficacy based on solvent “polarity”. Polarity was
somewhat ill-defined but the approach seemed to offer a genuinely productive way to optimise the complex
formulations so often used in cosmetics. At the time Abbott and Wiechers met, Wiechers had stumbled upon
HSPiP and had realised that his notion of “polarity” was much better expressed via HSP. They agreed to
combine their very different backgrounds and the software FFE was created (S Abbott, PhD, and JW Wiechers,
PhD, Formulating for efficacy, software, available at www.jwsolutionssoftware.com), using HSP principles to
guide formulators to effective formulations. FFE used the same basic Diffusion model has HSPiP but with
refinements to handle the multiple components in the formulation.
Tragically, Johann died suddenly and Abbott has continued the FFE journey. As a tribute to Johann, the science
behind FFE is described in S Abbott, An integrated approach to optimizing skin delivery of cosmetic and
pharmaceutical actives, Intl. J. Cos. Sci. 34 217-222 (2012). There is now plenty of evidence that this
“solubility” approach to skin permeation (as opposed to the LogP approach) is gaining ground. Whether HSP
are the best way to deal with this complex problem is still debatable – but until some more powerful algorithm
is provided it seems to be the best tool for the purpose.
101
102
Chapter 18 HSP and Diffusion
Diffusion of chemicals through polymers is a highly complex and highly debated issue. In this chapter it is
shown that HSP offer profound insights into these phenomena. Diffusion in polymers as such is not directly
connected to thermodynamics (HSP) but instead has a lot to do with kinetics. There are (at least) three reasons
for facing up to the controversy with the help of HSP. The first has to do with Health & Safety. As other
chapters show, HSP have a lot to offer on the subject of protective workwear such as gloves. The second reason
is that diffusion lets us tackle head on the thermodynamics v kinetics controversy that often gets in the way of
the use of HSP.
The third reason is that in the years since the first version of this chapter was written, HSPiP users have found
the link between HSP and diffusion to be hugely productive for solving problems in all sorts of areas such as
protective garments, environmental barriers, flavour scalping, flavour barriers, food safety, polymer “in use”
certification and many more.
The good news is that although the calculations of diffusivity are rather hard to implement, the theory of what is
going on is remarkably simple. A diffusion modeller is included within the HSPiP package so you can explore
the theory without having to worry about its detailed implementation. In what follows we are referring only to
solvent diffusion under constant temperature and are not discussing diffusion driven by other gradients such as
temperature.
For those who just want a quick method to estimate diffusion coefficients, the official Nordtest Poly 188 found
at http://www.nordicinnovation.net/nordtestfiler/poly188.pdf gives a well-validated methodology that is a
simplified version of what is found in this chapter. It is no coincidence that Hansen was one of the authors of
the Poly 188 test.
Contrary to popular mythology, there is only one type of diffusion that is important for polymers in normal
practice. If you do practical experiments you might be surprised by this statement because it’s rather obvious
that there are at least two distinctly different situations of interest: absorption (solvent going into a polymer)
and desorption (solvent coming out of a polymer). This is obvious to anyone who has done
absorption/desorption experiments because absorption is generally much faster than desorption. Those of you
who know more about the topic will also be aware that some diffusion rates are simple (“Fickian” diffusion)
and some are complex (e.g. “Type II” diffusion).
The response to this apparent complexity is to repeat that there is only one type of diffusion based on just one
simple equation, the diffusion equation (sometimes called Fick’s second law). The apparent differences
between all the different types reflect the fact that different factors involved in this equation are more or less
important depending on circumstances. The reason to stress this simple unity is that each of the different factors
is rather easy to understand. It is therefore rather easy to work out which factors are required to sort out what is
happening in just about any diffusion behaviour. Armed with HSP of solvent and polymer, with some
knowledge of molar volume (integral to HSP) and molecular shape, and with some understanding of whether
(and when!) a polymer is highly crystalline, semi-crystalline or elastomeric (often reflected in the HSP Radius),
you can readily calculate diffusion behaviour. As discussed below, the diffusion rate increases in the order
highly crystalline < semi-crystalline < elastomeric as governed by Factor 5. These categories are gross
simplifications only to be used as rules of thumb – there are certainly cases of crystalline polymers with higher
diffusion rates than semi-crystalline. But these rules are a good starting point. We provide some more rules of
thumb on diffusion rates in the text and in the HSPiP modeller.
Consider the following factors, one by one.
Factor 1. This is the Mass Transfer Coefficient, h. A large h means that if you have a well-stirred liquid in
contact with the polymer surface then you have no shortage of molecules to be able to diffuse into the polymer.
If, on the other hand the polymer surface is in contact with still air, then a boundary layer builds up and solvent
trying to escape from the polymer will see a high concentration of vapour which reduces the gradient for
diffusion and therefore slows down the rate, so h is small. One reason for there being a difference in absorption
and desorption in this case is the simple and obvious fact that the Mass Transfer Coefficient from 100% liquid
in absorption is much higher than that into a boundary layer of air (saturated with vapour) in desorption. But
there also can be times when mass transfer into the polymer from a liquid is limited (e.g. poor stirring of an

103
inert carrier liquid such as water, with fast absorption of a solute). In all of these cases there is rapid diffusion of
the chemical within the polymer and some problem with mass transport more or less exterior to the polymer.
Another factor which can cause large reductions in the mass transfer coefficient is the formation of a highly-
crystalline skin at the surface. The absorbing molecule simply cannot find any access point into the bulk (less-
crystalline) polymer. The reduction in mass transfer coefficient is larger for larger and/or more complex
molecules. Above a certain size the mass transfer coefficient becomes zero – there is no penetration into the
bulk. The most extreme example of this is when two polymers, even well-matched in terms of HSP, simply
cannot interpenetrate. What is surprising is that even some simple molecules (e.g. a single benzene ring or a
modestly branched structure) can have a mass transfer coefficient close to zero for some polymers such as those
with high crystallinity or other forms of close-packing such as are found in the Topas polymers. Chapter 16 of
the Handbook discusses some of these issues in greater detail.
A good method for checking if diffusion is limited by mass transfer is to carry out the experiments on samples
of different thicknesses – the thicker the sample, the less relevant the mass transfer effect becomes – as you can
readily determine using the diffusion modeller.
Given that most of us don’t know what the Mass Transfer Coefficient is for our systems, the modeller uses the
Hansen B value. This is the ratio of the diffusion resistance to the surface resistance. With D0 as the diffusion
coefficient at the lowest concentration encountered, and for a free film sample of thickness L, the diffusion
resistance is (L/D0). Dividing this by the surface resistance (1/h) gives B. Thus,
B= hL/D0
A high B (>10 for a constant diffusion coefficient) means essentially no significant limitation by mass transfer.
Factor 2. This is the local saturated concentration of the liquid in the polymer right at the surface during
absorption. For RED less than 1 for a correlation based on “good” or “bad” solubility, this concentration will be
very high since the solvent can in principle completely dissolve the polymer. It is very difficult to assign an
initial given surface concentration, but it is probably in the range of 50-70%, because at still higher
concentrations, the issue is not one of diffusion of solvent into the sample but diffusion of the sample into the
solvent. For RED larger than 1, the larger the value, the lower the local saturated concentration and therefore
the slower the absorption rate. This is the reason that HSP is so important for understanding diffusion. In the
modeller a simple algorithm has been used to illustrate this so you can compare overall diffusion rates as the
RED changes. The algorithm is for illustrative purposes only – it’s up to you to specify the surface
concentration in any specific scenario.
Once the molecule is inside the polymer, as long as it is within its solubility limit (we’ll explain this in a
moment), HSP play no further role. The rates of diffusion of a low RED and high RED solvent of similar molar
volume and shape are the same. You might be surprised that in an HSP book it is claimed that HSP are not
important for diffusion inside the polymer, i.e. the diffusion coefficient at a given concentration. The
experimental data have confirmed this fact many times. This also means that HSP play no part in classic
desorption experiments to air. Naturally the desorption from one polymer to another (migration) does depend
on the HSP of the second polymer as a large HSP mismatch would mean, as in absorption, a low surface
concentration in that polymer.
Although Factor 2 is about absorption, it’s a good point to discuss why desorption takes so much longer than
absorption. It has been shown that the diffusion coefficient increases exponentially with the concentration of the
solvent. For rigid polymers this increase is a factor of about 10 for each 3%v increase in solvent concentration.
For flexible polymers the increase is a factor of 10 for about 15%v increase in solvent concentration. Whereas
during the whole time of the absorption process, the solvent is largely diffusing in at concentrations
approaching the maximum (surface equilibrium) concentration, and certainly much higher than the lowest
concentration, in desorption most of the solvent diffuses out at much lower (and falling) concentrations than the
initial one. In desorption the concentration at the surface is low (zero) so the process is largely controlled at or
near the exit surface since the diffusion coefficient here is so low.
Before going to Factor 3, let us clarify this statement that “as long as it is within its solubility limit, HSP play
no further role.” If you dip some polymers (e.g. epoxies or polypropylene) into hot water, the solubility is
increased sufficiently for water to diffuse in (the δH of water falls off rapidly with temperature, boosting its
solubility). If you then cool the sample, the water becomes insoluble in the polymer. The individual water
104
molecules can still diffuse (diffusion coefficient is independent of HSP) but when they meet each other, they
phase separate from the polymer. This is the classic case of water blisters. If you hot-dip/cool a number of times
you get more and more water into the polymer, but each time you cool, the water phase separates out into
bigger and bigger blisters. The blisters are very persistent. That is because each blister is a new diffusion
problem from one phase (the water blister) into the other (the polymer). Given that there is a large HSP
mismatch, the surface concentration at the blister/polymer interface is low so the rate of diffusion is low. Those
blisters can be very persistent. So now you can see why it’s important to qualify the statement that HSP have no
effect on diffusion once the molecule is inside the polymer. See p141 of the Handbook or C.M. Hansen, New
Developments in Corrosion and Blister Formation in Coatings, Progress in Organic Coatings, 26, 113-120,
1995 for further details. Incidentally, a beautiful demonstration of HSP co-solvent effects is provided by the
well-known fact that glycol ethers in coatings can produce blisters under aggressive thermal/water cycling
tests. The ethers remain in the coating and during the hot/wet part of the cycle the combined glycol ether/water
HSP is a sufficient match to the hot polymer to allow the water to enter. On cooling the blisters start to form.
The same coatings without the glycol ethers have no blistering because the HSP distance of the water is too
great, even at the higher temperatures.
Factor 3. This is the molar volume. The larger the molar volume, that is, the size of the molecule, the smaller
the diffusion constant. This is a generalisation that is modified by Factor 4, and therefore it must be used as a
guideline rather than a hard fact. In the modeller the rule of thumb (based on the rather small number of studies
in the literature) is used that the log of the diffusion constant is proportional to the molar volume. The constant
of proportionality changes strongly from rigid to flexible polymers. For rigid polymers there is a very strong
dependence, so a doubling of molar volume can result in a 10 to 100-fold reduction in diffusion constant. For
flexible polymers the dependence is weak – a doubling of molar volume may merely halves the diffusion
constant.
Factor 4. This is the molecular shape. A linear, flexible molecule can easily wiggle through a polymer. A rigid
(aromatic) molecule or a highly branched molecule takes much longer to find a space (or “free volume” – see
Factor 5) in which to wiggle. The rule of thumb section below gives some examples of the effects of Factors 3
and 4. A well-known table of solvent diffusion rates from Shell suggests that the combined effects of molar
volume and molecular shape result in a diffusivity order from faster to slower of: Methanol, Acetone, MEK,
Ethyl Acetate, n-Heptane, n-Butyl Acetate, Benzene, 1,4-Dioxane, Toluene, MIBK, i-Butyl Acetate, 2,4-
Dimethyl Pentane, Cyclohexane, Diacetone Alcohol, Pentoxone, Methyl Cyclohexane, Cyclohexanone, Methyl
Cyclohexanone. This table was derived using the technique originally developed by Hansen when he created
the data of Fig 1 of the chapter on glove data.
Factor 5. Informally we can say that Factor 5 is whether the polymer is rigid (slow diffusion), flexible (faster
diffusion), or a quasi-solution – e.g. an elastomer where the polymer is held together by just a few crosslinks
(still faster diffusion). More technically we can talk about polymer “free volume” where the rigid polymer is
below its glass transition temperature (Tg) and therefore has little main-chain segmental motion and the flexible
polymer is either naturally above its Tg or has large free volume for main-chain segmental motion thanks to all
the solvent. It is Factor 5 which causes most of the confusion about diffusion science. The more free volume
there is, (or, equivalently, the lower the glass transition temperatures), the faster will be the rate of diffusion.
The simple view of all this is that when the solvent enters a polymer it starts to plasticize it and diffusion is
faster. As more solvent diffuses into a polymer the diffusion rate can increase by factors of 100’s, 1000’s and
even millions (the solvent itself is increasing the polymer free volume) so the diffusion looks more complex.
This “concentration dependent” diffusion is not some sort of special case or special phenomenon – it is the
general rule for polymers. Sometimes it looks “special” because the concentration dependence is so large – but
this is only because the polymer happens to be rigid and therefore susceptible to a large increase in diffusivity.
This simple view needs to be treated with caution. Some crystalline regions are so impenetrable to solvents that
they act as permanent blocks so there is little increase in diffusion coefficient. Some “amorphous” polymers are
in fact highly rigid so show a large increase in diffusion rate whilst others are highly flexible and therefore
show a small increase. The following figure is for diffusion of chlorobenzene in polyvinylacetate.

105
Figure 18-1 Diffusion coefficients for chlorobenzene in poly(vinyl acetate) at 23°C measured by absorption and desorption
experiments in a quartz balance apparatus as well as with an isotope technique. v f is the volume fraction. The upper curve in
the figure is for diffusion coefficients based on total film thickness. The lower curve is for diffusion coefficients based on dry
film thickness as used in the modeller. It should be noted that the lower curve varies more and more from the upper one as the
concentration of solvent increases. A self-diffusion coefficient for a liquid (vf = 1.0 in the figure) is a fictitious quantity on the
lower curve, although it is used to define the diffusion coefficient in the solvent rich regime.
The data in the figure are the result of combination of absorption and desorption experiments supplemented by
isotope experiments to give a unified view of concentration dependent diffusion in polymers. In every
measurement the observed diffusion coefficient was initially considered as a constant that must be adjusted to
the change in concentration within the film during the whole process. Solutions to the diffusion equation with
different concentration dependencies were generated and compared with that for a constant diffusion coefficient
to develop these “F” factors. The apparent, constant diffusion coefficients are given by squares in the figure
with the corrected values being given by circles. The adjustments are for absorption, desorption, or surface
effects as indicated by the subscripts a, d, and B. Desorption experiments take place largely at local
concentrations within the film that are much lower than the initial concentration that is ascribed to the
experiment. These adjustments are much larger for desorption than for absorption. The correction for surface
effects in the absorption experiment at vf = 0.5 is a multiplier of 250. Such experiments should not normally be
used to measure diffusion coefficients at these intermediate concentrations. The procedure used for these
adjustments is described in more detail in the Handbook. The upper curve is for diffusion coefficients based on
the wet film thickness, while the lower curve is for dry films. It is clear that there are two different regimes,
rigid at lower concentrations, and elastomeric at higher concentrations, separated by the break at about 0.2
volume fraction of chlorobenzene.
Diffusion coefficients at very high solvent concentrations are usually best described based on total film
thickness rather than dry film thickness, since the value for the latter at zero polymer concentration becomes
meaningless. A value at 100% liquid is required to define the diffusion coefficient curve, however, and this
value will be somewhat lower than that found in the literature for self diffusion in the given liquid. Fortunately,
diffusion at very high solvent concentrations is usually so rapid as to not be a significant effect in the situations
of major interest, so smaller deviations in this region are not important. Whether diffusion is very rapid or
“super-rapid” does not really matter since the process is controlled by what happens at (much) lower
concentrations. Usually the surface concentrations at equilibrium for absorption or permeation and the start
concentrations for desorption are sufficiently low to allow neglect of this effect.
The modeller gives you full control over all these factors. It assumes three regimes that change at two critical
solvent concentrations (which you can choose). Each regime has a diffusion coefficient which depends on a D0
106
value (i.e. the value at the lowest concentration for which this regime is applicable) and an exponential “k” x
concentration term which reflects the increase in diffusion rate. The larger the k, the larger the increase in
diffusivity with concentration:
Drigid = D0r exp(kr x concentration)
Dflexible=D0f exp(kf x concentration)
Dsolution=D0s exp(ks x concentration)
Some useful data and rules of thumb for Do, cm2/s
In polyvinylacetate at room temperature (23°C):
Liquid D0, cm2/s
Water 4x10-8
Methanol 4.5x10-10
Ethylene glycol monomethyl ether 2x10-12
Chlorobenzene 1x10-14
Cyclohexanone 1x10-15

Figure 18-2 A typical dependency of D0 on MVol. In this example of diffusion in PVAc, D0 falls by 2
orders of magnitude for each doubling of MVol

Chlorobenzene concentration %v
0.2 1x10-8 (changeover from rigid to flexible-type behaviour)
0.59 3x10-6 (changeover from flexible-type behaviour to solution-type
behaviour)
0.76 9x10-6

Pure solvents (self diffusion unless indicated otherwise):


Chlorobenzene (25°C) 1.7x10-5
Chlorobenzene (10°C) 1.3x10-5
Chlorobenzene (40°C) 2.0x10-5
Ethanol (25°C) 1.2x10-5
Water (25°C) 2.3x10-5
Glycerol in ethanol (25°C) 0.6x10-5

In polystyrene D0r:
Chloroform 3x10-13
107
In polyisobutylene D0f:
n-Pentane 2.5x10-9
Isopentane 1.2x10-9
Neopentane 0.1x10-9

Diffusion coefficients above about 10-8 cm2/s appear to indicate elastomeric behaviour in otherwise amorphous,
rigid polymers, but this value may be lower for true elastomers.

There is an increase in diffusion coefficient:


For rigid polymers: a factor of about 10 for each additional 3%v
For flexible polymers: a factor of about 10 for each additional 15%v

Suggested Diffusion Coefficients


Earlier versions of HSPiP allowed the user to guesstimate diffusion coefficients. This functionality was never
entirely satisfactory and it cluttered up the interface and confused users. Instead an “EPA approved” method of
calculation, at least for some common polymers, is supplied as a starting point. Measurement of diffusion
coefficients is rather easy. Simply dip thin slices of polymer into the solvent and weigh at regular intervals then
use the diffusion modeller to fit the data!
Special Cases and Combinations
By breaking down diffusion into these five factors it becomes easy to disentangle much of the confusion about
special cases such as “Super Case II”. There is really nothing special about these. Typically what is happening
is that the mass transfer limitation (Factor 1) is interacting with the strong dependency of diffusion on
concentration (Factor 5) in a way that is not intuitively obvious. It’s a useful short-hand to call any mass
transfer effect a “surface resistance” but this term is not very insightful. A “surface resistance” from poor
airflow (desorption) or poor stirring (absorption) is very different from a “surface resistance” due to a highly
crystalline skin on an injection molded part.
Further confusion arises when tests are done on very thin parts (or, even, hyper-thin parts when FT-IR
measurements are made on the first few µm of a sample) because then the mass transfer limitations are
proportionally much more significant than on large parts. A polymer showing an “anomalous” diffusion when
tested on thin samples may well give entirely normal diffusion when tested on a thicker part.
That’s all there is to it. The bad news is that there is no simple way to calculate each of the five factors. If you
are lucky enough to have reference values of your particular polymer then you are off to a good start. But the
good news is that with the modeller that captures the essence of each of these factors you can make rapid
progress in understanding whichever system is of particular interest to you. So let’s see what it can do.
Absorption and breakthrough

108
Figure 18-3 A simple absorption and breakthrough plot
Here we have an elastomer with a medium-sized, linear solvent. At low concentrations the diffusion rate is 1E-
07 and above 0.333 volume fraction the rate becomes constant at 3.8E-07. The solvent has a a surface
concentration of 0.24. After 3.1min it has broken through (at a 0.1% level) to the other side of a 0.2mm sample.
Shortly after that, the concentration gradient stabilizes to its final form with the absorption being balanced by
the desorption. The “Square Root” option has been chosen which creates a straight-line in the increase of %
concentration.

109
Figure 18-4 The same absorption but with a slower-diffusing molecule
A larger or cyclic molecule might a diffusion rate a factor of 10 slower, so breakthrough time is 31.1min.

Figure 18-5 A slower-moving molecule but a lower RED number

110
A solvent closer HSP match but the same cyclic structure and molar volume is estimated to breakthrough in
24.1min simply because the surface concentration is estimated to be higher at 0.55.
Desorption

Figure 18-6 A classic desorption curve


The same solvent is assumed to have saturated the block of polymer and is now allowed to desorb via the left-
hand surface (the right-hand being assumed to be blocked). The coloured curves show the solvent distribution
with time, the red curve being the distribution after 100min.

111
Figure 18-7 Desorption by a smaller, faster molecule
The smaller, linear molecule desorbs considerably faster.

Figure 18-8 Same molecule but desorption from a rigid polymer

112
This behavior resembles the formation and drying of a polymer film from solution. Such behavior has been
studied in detail by Hansen in Hansen, C.M., A Mathematical Description of Film Drying by Solvent
Evaporation, J. Oil Colour Chemists’ Assn., 51, No. 1, 27-43 (1968) and in the Doctoral thesis from 1967 that
is available as a PDF file on www.hansen-solubility.com by clicking on the cover page.
In a crystalline polymer, the shape is highly skewed. Because the diffusion rates are relatively high through the
bulk, the profile is rather flat. At the edge, where concentration is very low, the diffusion rate plummets.
“Surface resistance”

Figure 18-9 “Surface resistance” coming from a Mass Transfer limitation in a permeation study
In this permeation example, the “surface resistance” comes because the “B” value (ratio of diffusivity to surface
resistance) has become significant (<=1) at the entry surface. When there is a significant surface condition at
the entry surface, where there is (usually) contact with a liquid there can also be a significant surface condition
at the exit surface, where the contact is with a gas. This possible effect at the second surface is not considered in
the present discussion. The diffusion and surface effects are finely balanced. Solutions to the diffusion equation
for absorption, as modelled here in Figure 14-8 and in Figure 14.9, give the so-called “s-shaped” or “sigmoidal”
absorption curve. This behaviour is often misinterpreted in the literature as being anomalous diffusion since the
absorption is not a straight line versus the square root of time, which is called “Fickian”. The other forms of
“anomalous” diffusion that are incorrectly interpreted in the literature have the coined terminology Case II and
Super Case II. These are discussed in the following. An article in the European Polymer Journal (Hansen, C.
M., The significance of the surface condition in solutions to the diffusion equation: Explaining “anomalous”
sigmoidal, Case II, and Super Case II absorption behaviour, European Polymer Journal 46 (2010) 651–662)
explains these so-called anomalies with the following abstract:
“Absorption into polymers is frequently described by the terms Fickian, sigmoidal (S-shaped), Case II, or Super
Case II. This terminology is used to describe absorption that is respectively, linear with the square root of time,
has a slight delay or S-shape with the square root of time, is linear with linear time, or increases more rapidly
than with linear time. Solutions to the diffusion equation, Fick’s second law, that include a potentially
signficant surface condition are shown to reproduce all of these. Sigmoidal absorption results when the surface
condition is moderately significant for either a constant diffusion coefficient or exponential diffusion
113
coefficients. Exponential diffusion coefficients and a lower surface mass transfer coefficient result in Case II
type behavior, with Super Case II behavior resulting when the surface condition becomes still more significant.
The results reported here are supported by extensive experimental data with reasonable and verifiable values for
the diffusion coefficients and surface mass transfer coefficients.”
The following figures show these effects for absorption.

Figure 18-10 Simulation of water absorption into untreated PVA films (Sigmoidal Absorption).
The S-curvature in the lower right figure matches the experimental data in [Hasimi, A., Stavropoulou K.G.,
Papadokostaki M, Sanopoulou M. Transport of water in polyvinyl alcohol films: Effect of thermal treatment
and chemical crosslinking. European Polymer Journal Vol. 44, 4098-107 (2008)] very well. The initial
curvature is very dependent on the concentration dependent diffusion coefficients used, so improved diffusion
coefficient data, especially at low concentrations may remove any minor differences. The concentration
dependent diffusion coefficients that were used are given in the upper right corner of the figure. The
concentration profiles as a function of time confirm that there is diffusion resistance of significance only up to
concentrations near 0.1volume fraction or less. At higher concentrations diffusion within the film is much faster
than water can get to and through the exposed surface. The experiments are primarily a measure of a mass
transfer coefficient of unknown origin (test setup, surface effects, etc.) as reported in (Hansen 2010) cited
above.

Case II absorption

114
Figure 18-11 Simulation of Case II type absorption

The straight line absorption curve at the lower right is typical of Case II absorption. The h value, 8(10)-6 cm/s,
is reasonable as is the diffusion coefficient profile. The tail at greater than 90% absorption would be reduced for
higher h values, which could also be reasonable. This kind of tail can be seen in the literature for the
polystyrene/n-hexane system in an often cited reference for Case II behaviour [Jacques, C.H.M., Hopfenberg,
H.B., Stannet, V. in Permeability of Plastic Films and Coatings. Hopfenberg, H.B. Ed. New York:Plenum
Press;1974, p.73.]. An initial curvature upward is also possible with an increased h value.

The traditional example of Case II type diffusion is that reported in [Thomas NL Windle AH. Transport of
methanol in poly(methyl methacrylate. Polymer 1978, 255-265]. The reason for this, in addition to the straight
line absorption curve using linear time, is that step-like concentration gradients were reported for methanol in
the PMMA. Iodine was used as a tracer for the methanol and showed horizontal concentration profiles leading
to the assumption of a step-like advancing front for methanol. It has been shown that iodine is not suitable for
this with much slower diffusion than methanol in the polymer. There is a more detailed analysis of the
absorption of iodine and methanol into methanol-swollen PMMA at different temperatures using data from the
studies of Thomas and Windle available on www.hansen-solubility.com as a free download. It is a matter of
chance that the “advancing fronts” of the iodine met at the center of the free film exposed on both sides at the
same time as methanol fully saturated the film. This was found at 30°C but not at higher temperatures. The
30°C absorption curve for methanol is matched in Figure, but the concentration gradients are clearly not step-
like. We have provided a file for loading into the diffusion modeller so you can repeat the calculations reported
in this figure making it possible to follow the process as it unfolds. If diffusion coefficient data at low
concentrations for the methanol in PMMA system become available, these can be entered into the settings
rather than the estimates used for this figure to confirm the result.

115
Figure 18-12 An example of typical Case II absorption of methanol into PMMA at 30°C using data from Thomas and Windle.
The absorption curve at the lower right matches the experimental data as a straight line on a plot using linear time, but the
concentration gradients are very far from a step-like advancing front as indicated by Thomas and Windle. The iodine tracer
method used in this study is not suitable to follow the diffusing methanol with misleading conclusions as a result .
The same authors also measured absorption of methanol into PMMA at 0°C. The technique of a new sample for
each single data point gave somewhat scattered data, but Figure 18-12a shows a good fit and Super Case II
behavior with an increasing rate of uptake at longer times.

116
Figure 18-12a. Methanol into PMMA at 0°C by the same authors. A “Super Case II” example.
Super Case II absorption
As we’ve stressed, there is no need to invoke any new principles to explain exotic behaviours such as Super
Case II. To make it appear, we simply have (a) a large dependence of diffusion rate on concentration as in
Figure 16-11, which means a very low D0, and (b) a significant surface entry resistance. Here’s a specific
example:

Figure 18-13 Simulation of Super Case II absorption


117
A significant surface condition coupled with the measured concentration dependent diffusion coefficients given
in Figure 16-1 above leads to a marked increase in absorption rate well after the absorption process has started.
There is an exponential approach to the equilibrium value at the very end of the absorption process because the
driving force for further absorption at the surface has become small. The significant surface condition in such
cases can probably be attributed to the hindered entry resistance, perhaps like a skin, described below since the
value of h is only 1(10)-7 cm/s, but some combination with one or more additional sources of significant surface
condition resistance is also possible.
It is useful at this point to contrast the absorption for the same diffusion coefficient data but with no surface
resistance. The diffusion coefficients used over the concentration interval from 0 to 20% by volume differ by a
factor of 106. Similar straight line curves on a square root of time plot are found for lower values of this ratio.
This could lead to some confusion with results for a constant diffusion coefficient, where the plot is also a
straight line to uptakes that are relatively high but not as high as shown in the figure. The diffusion coefficient
data used here are those measured for chlorobenzene in PVAc at 23°C.

Figure 18-13a. Chlorobenzene into PVAc


There are different combinations of the parameters that can lead to Super Case II as seen in the following
example.
The traditional example of Super Case II absorption is that given in [Holley RH Hopfenberg HB Stannet V.
Anomalous transport of hydrocarbons in Polystyrene, Polym Eng Sci 1970, 10, 376-382]. The absorption curve
for n-hexane in poly(styrene) given in the figure models the uptake curve very well. The surface mass transfer
coefficient is dominant in this result as seen by the horizontal concentration gradients almost from the start, but
the concentration dependent diffusion coefficients affect the shape of the curve. These are not the true diffusion
coefficients for n-hexane in poly(styrene) since there was retained solvent and orientation in the samples. The h
used here is very close to that derived from the initial absorption rate reported for this experiment. The details
of this and other aspects of Super Case II absorption are given in a free download on www.hansen-
solubility.com.

118
It has been shown in the above that the diffusion equation including concentration dependent diffusion
coefficients and, if necessary, a significant surface mass transfer coefficient, can satisfactorily model what has
been traditionally called “anomalous” absorption in polymers.

Figure 18-14 The hexane/poly(styrene) Super Case II absorption


Significant surface condition caused by hindered entry
The expected phenomena that hinder mass transfer in absorption will also be capable of hindering mass transfer
in desorption, with the opposite sign in the mathematics, of course. These include diffusion through a stagnant
gas layer, heat transfer to or from the given surface, wind velocity, etc. In addition to these easily recognized
effects a more subtle cause of a significant surface condition has been elucidated recently [Nielsen, T. B. and
Hansen, C. M. Significance of surface resistance in absorption by polymers. Ind Eng Chem Res, Vol.44, No.11,
3959-65 (2005), and also in the Handbook]. The table is abstracted from this study.

119
Solvent Apparent h, cm/s Equilibrium uptake, volume fraction
Tetrahydrofuran 2(10)-4-1(10)-3 0.676
Hexane 7.78(10)-6 0.351
Diethyl ether 1.21(10)-6 0.268
Propylamine 1.49(10)-7 0.181
Ethylene dichloride 1.18(10)-7 0.176
Ethyl acetate 1.46(10)-8 0.076
n-Butyl acetate 8.30(10)-10 0.202

Table 18-1 Apparent surface entry mass transfer coefficients and equilibrium uptake for solvents in a COC polymer Topas
6013® (Ticona).
This is an entry or hindered surface passage resistance dependent on the size and shape of the entering
molecule, and, of course, the polymer surface morphology. For smaller molecules such as tetrahydrofuran, n-
hexane, and 1,3-dioxolane there is no significant surface condition effect of this kind for absorption into the
COC polymer Topas© 6013 from Ticona. With more extensive absorption experiments it was obvious that entry
into the polymer was more difficult as the size of an absorbing molecule increased or its structure became more
complicated, such as with side groups or cyclic entities. So-called S-shaped absorption curves, of the type
shown above, with a pseudo time-lag were observed in a number of cases, including absorption of ethylene
dichloride, diethyl ether and n-propyl amine. Solvents containing benzene rings and more complicated
structures, such as acetophenone, phenyl acetate, 1,4-dioxane, and methyl isobutyl ketone are completely
prevented from absorbing into this same polymer, in spite of HSP similarity to those that do absorb. This
comparison indicates that they should readily absorb into the polymer. Their absorption after prolonged liquid
exposures has been so little that it could not be detected. In such cases there is no significant transport
resistance in the external media, so it has been postulated that there can also be a significant surface condition
caused by an entry or surface passage resistance. This type of resistance deserves much more attention to fully
understand what is happening. One can surmise however, that for most polymers there will be smaller
molecules that enter readily, and very large molecules (such as a polymer molecule of the same kind) that will
not be able to enter at all. Between these extremes there will somewhere be a range of molecular sizes and
shapes where entry is possible but becomes retarded since the molecules cannot rapidly find suitable sites to
absorb even though they may be adsorbed. The orientation of adsorbed molecules at such selected sites where
absorption is possible is thought to be a key element in this type of resistance to transport. Thermal treatments
can be expected to affect surface morphology, such as in known from rapid cooling in injection molding.
Because of this, thermal treatments having an effect on surface morphology may also be a factor in this
behaviour.
Further analysis of the data in the Table confirms that, excluding the data for n-butyl acetate and using the
higher value for THF (with such rapid absorption, the errors are large), the plot of log(h) v C (the line is a guide
for the eye) confirms that the amount of solvent that can be absorbed at equilibrium is very significant for the h
value of the system.

120
Figure 18-15 The relationship between log(h) and the equilibrium uptake fraction C for Topas 6012®. Data from Table 16-1
This equilibrium uptake is clearly a function of the HSP relations. Thus, HSP relations are extremely important
for diffusion in polymers affecting the overall concentration gradient, the maximum diffusion coefficients
within the bulk of films, and in many cases also the entry mass transfer coefficient at the surface. What can be
learned from this in practice is that a mismatch in HSP is the best strategy for a good barrier polymer,
recognizing that the size and shape of absorbing molecules can also restrict absorption or even prevent it
regardless of HSP considerations.
Permeation
The modeller automatically calculates permeation rates in g/cm²/s. This is of necessity an approximation
depending on how long you run your simulation – the value displayed gradually asymptotes to the final value as
the system equilibrates.
For those who prefer a plot of the integrated amount permeated that option is available in the 3rd Edition, along
with an extrapolation down the straight part of the curve to estimate the “lag time”. Such plots are common in,
for example, the skin permeation literature. Numerous examples of permeation in chemical protective gloves
are given in Chapter 17.
Summary
With this powerful modeller you can explore a wide range of systems, including the important “Breakthrough”
type of experiments where both kinetic and thermodynamic (HSP) factors play a role. Examples of Fickian and
what is commonly and erroneously called non-Fickian or anomalous diffusion have been given to help guide
your efforts.
The next two chapters treat diffusion in protective gloves in more detail, showing how improved judgment of
glove safety is possible and how one can actually deduce the concentration dependent diffusion coefficients
from permeation data.

121
Chapter 19 It’s your call (Rational Selection of Chemical Protective
Gloves)
If you had to recommend the best gloves to use in situations that could be literally life or death, what would you
do? To be more specific, what gloves would you recommend to hospital staff so they can safely handle
immensely potent cytotoxic drugs of the type used in chemotherapy?

One approach you could adopt is to insist on glove permeability tests for all cytotoxic drugs on all relevant
gloves. But this just isn’t practical.

An alternative approach is to use chemical intuition: “It seems to me that glove X should have no problem with
cytotoxic chemical Y because they are chemically incompatible.” But would you then have the courage to
handle these chemicals all day with the gloves that you have intuited should be OK? We wouldn’t.

In the absence of specific testing, the approach we favour is a rational, numeric approach. The numbers it
produces cannot be accurate, but represent respectable estimates of what can be expected. But we don’t need
high accuracy itself. What we need is a large margin of safety. If we estimate that a potent drug would have a
breakthrough time of 60 minutes, it doesn’t really matter if the real value is 30 minutes or 120 minutes; in
neither case would we volunteer to wear those gloves for a whole morning. But if we estimate that the
breakthrough time is 360 minutes, we don’t mind if it’s really 300 or 600 minutes, the margin of safety (given
that these gloves are worn for significantly shorter periods whilst handling the drugs) is relatively large enough
for us to wear the gloves with confidence.

Happily, we have an excellent methodology to make this numeric estimate. We know from the Diffusion
chapter that Permeability=Surface Concentration x Diffusion Coefficient. And we already know that a large
HSP mismatch between chemical and polymer reduces the surface concentration and therefore reduces the
permeability. A measure of the HSP mismatch is the RED number, so a large RED means low permeability and
therefore long breakthrough time.

We also know some things about the diffusion coefficient. First, because we are handling very potent
chemicals, they will tend to be in low concentrations, so the relevant diffusion coefficient is likely to be the D0
value, the “intrinsic” low-concentration diffusion coefficient. There will never be enough of the chemical to
give the orders of magnitude increase in diffusion coefficient that could give dramatic decreases in
breakthrough time, nor significant swelling of the gloves. Second, we know that the molecular shape and size
makes a big difference. There are no general rules for the effect of molar volume but for the materials used for
gloves the approximation from the previous chapter that doubling the molar volume halves the diffusion rate
would be a conservative choice. We know, too, that linear molecules generally wiggle their way through with
more ease than branched, cyclic and/or aromatic molecules. Once again, we can make a conservative estimate
of a halving of diffusion rate for a given molar volume if the molecule is not especially linear. For many of the
complex cytotoxic molecules we can be confident that this is much too conservative.

We almost have all the tools we need to estimate breakthrough times for cytotoxic molecules through typical
gloves. But we are missing a key bit of data. Although a “large” RED undoubtedly means a “longer”
breakthrough time, how do we actually estimate the breakthrough time for a given RED number? Fortunately
there are HSP correlations for the most common types of gloves that can be used for this purpose.

Let’s take a specific example. From the Attacking DNA chapter we know that Cyclophosphamide has HSP
[17.5, 11.9, 12.6] and molar volume 279.

Now let’s look at a table of glove properties. Along with the glove type and breakthrough times is given the
upper and lower molar volumes of molecules used in the correlation. This is a useful guide for comparing to the
molar volumes of the chemical under consideration. “All” mean that there were no limitations on the solvents.

122
Glove Type - Breakthrough δD δP δH R
NR 20 MIN (58 to 178) 17.50 7.30 6.50 5.10
NR 1 HR (61 to 178) 16.60 9.10 4.40 10.00
NR 4 HR (58 to 178) 19.00 12.60 3.80 13.30
BR 20 MIN (71 to 110) 16.50 1.00 5.10 5.00
BR 1 HR (71 to 126) 15.80 -2.10 4.00 8.20
BR 4 HR (All from 71) 17.60 2.10 2.10 7.00
NAT 20 MIN (61 to 267) 14.50 7.30 4.50 11.00
NAT 1 HR (56 to 325) 15.60 3.40 9.10 14.00
NAT 4 HR (up to 325) 19.40 13.20 7.70 19.00
PVC 20 MIN (61 to 267) 16.10 7.10 5.90 9.30
PVC 1 HR (61 to 267) 14.90 11.10 3.80 13.20
PVC 4 HR (up to 149) 24.40 4.90 9.90 22.70
PVA 20 MIN (All) 11.20 12.40 13.00 12.10
PVA 1 HR (All) 15.30 13.20 13.50 8.80
PVA 4 HR (All) 17.20 13.60 15.40 10.90
PE 20 MIN (All from 40) 16.90 3.30 4.10 8.10
PE 1 HR (All from 40) 17.10 3.10 5.20 8.20
PE 4 HR (All from 56) 24.10 14.90 0.30 24.30
VIT 20 MIN (All) 10.90 14.50 3.10 14.10
VIT 1 HR (56 to 178) 16.50 8.10 8.30 6.60
VIT 4 HR (All to 178) 13.60 15.40 8.60 14.40
NEO 20 MIN (75 to 178) 17.60 2.50 5.90 6.20
NEO 1 HR (69 to 178) 19.00 8.00 0.00 13.20
NEO 4 HR (61 to 266) 14.60 13.90 2.30 15.90
Table 19-1 Representative glove properties

If the Cyclophosphamide is compared to the 20min Nitrile rubber gloves [17.5, 7.30, 6.50] and radius 5.1, the
RED number is 1.5. If the comparison is for solvent breakthrough times of 1 hour then the glove HSP values
are different [16.60, 9.10, 4.40], radius 10.0 and the RED is much reduced, to 0.88 The radius is larger because
the increased time allows less favourable molecules to diffuse through. By extrapolation/interpolation it’s
possible to say that for a RED of 1 the breakthrough time would be 45 minutes. In other words, we have
defined a radius of Nitrile which places Cyclophosphamide in the danger category of “soluble” and therefore
we can assume, all things being equal, that it will diffuse through in 45min if present in the sorts of
concentrations typical of the permeation cell breakthrough experiment.

But what are those concentrations?

Again, we can make rational estimates, but quickly add that these can be improved upon with experimental
swelling data for a given glove type.

123
Breakthrough time Extension of Estimated uptake, Estimated Dmax/D0
in Permeation cell breakthrough time %
for very low
concentrations
<20 m 4.0 25 50
20-40 m 1.9 15 10
40-60 m 1.5 10 5
60-240 m 1.1 5 2
>240 m 1.0 <3 1
Table 19-2 Estimated values of uptake and Dmax/D0

For very quick permeators, the glove material could easily hold 25% of the permeant. For very slow permeators
the glove will probably hold <3%. We can then use the typical curves of diffusion coefficient v concentration to
estimate the Dmax/D0 at those concentrations and, therefore the extension of breakthrough time for very low
concentrations.

The above argument is sound, but the actual numbers cannot possibly be accurate. But we don’t need extremely
high accuracy. If we estimate that our intrinsic breakthrough time is 45min and if we think (from the above
table) that we therefore have a factor of 1.5 which takes us to 70min this really doesn’t alter our judgement very
much. Whether it’s 45min or 70min we’re not going to use those gloves for a whole morning.

And when the gloves give us >240min we don’t care if the extra safety factor is 1.1 or 1.2, we’re already in
quite a good safety zone.

If we continued with the Cyclophosphamide example we would add our final factor, the “shape/size” factor.

This is clearly not a linear molecule – it’s both branched and cyclic, but then so are some of the molecules used
in the breakthrough experiments. So there’s no room to give a significant change of D0 on those grounds. Its
molar volume is about 50% larger than any used in the breakthrough experiments so maybe we can add a factor
of 1.5 to the breakthrough time. We can finally arrive at an estimate of 45 x 1.5 x 1.5 = 100min.

And because it’s good practice to wear two sets of gloves (in case one gets accidentally holed), and because
breakthrough time goes as thickness squared, two gloves take the breakthrough time to beyond 200 minutes.

For Natural Rubber (latex) gloves the basic estimate is <20min. Even allowing for some adjustments upwards,
it seems clearly inappropriate to think of using latex gloves, even two pairs.

For Butyl gloves, the starting estimate is >240min because Cyclophosphamide’s RED number at 20min is 2.8
and at 4 hours it’s 2.0. There is no extra 1.5 for solubility because a 240min diffuser is already at low
concentrations. So we can add a 1.5 for shape/size giving use >360min or >6 hours. That’s not bad.
124
Polyethylene has such a large HSP distance from these sorts of drugs that such gloves are obviously excellent
barriers from a diffusion point of view. However, they are easily ripped and can’t be recommended for such
critical use.

As this may be a matter of life and death, let’s repeat ourselves. The calculations on the Cyclophosphamide and
Nitrile gloves cannot possibly be highly accurate, but they are good enough to say that Nitrile gloves are not
recommended for long-term use. But for short-term use, with a rule that two pairs of gloves should always be
worn (and extra rules if there is any chemical incident), they look OK. Similarly, the calculations with the Butyl
gloves cannot possibly be accurate, but they are more than good enough to say that there is a considerable
margin for safety for a whole morning or for cleaning up significant spills. And that’s all we need.

Of course it’s not up to the HSP scientist to make the final recommendations. A large number of other
considerations have to be taken into account. For example, whilst Butyl gloves are excellent barriers, they
aren’t good for delicate handling of medicines. Latex gloves are very comfortable, but the barrier properties for
these sorts of chemicals are much too poor. Nitrile gloves are very comfortable and are a better barrier than
latex. So a hospital committee might decide, for example, that Nitrile gloves are a good compromise choice
provided that (a) two pairs are worn, (b) the outer glove is replaced after 30min and (c) if there is any serious
incident (e.g. a spill of the chemicals) the user swaps to Butyl gloves.

If you are disappointed by this approach then think about how expert committees reach their opinions on
exposure limits of chemicals. They almost never have enough good data to reach a definitive and accurate
assessment of the specific risks of a specific chemical. But usually they don’t need that accuracy. Instead they
need a defensible set of numbers to say that the risk level is in this range rather than that range. From those
numerical judgements all sorts of practical consequences can then flow. They have to make such judgements. If
they say that all chemicals are dangerous then we can’t live a practical life. If they say “we don’t have good
enough data to form any judgement” then we live a lottery life.

If experts say “no glove is 100% guaranteed to handle all possible cytotoxic chemicals” then their life-saving
capabilities for cancer patients will never be practically deliverable by medical professionals. If experts say “we
can’t calculate any glove to high accuracies, so just make your own judgement” we are asking the medical
professionals to take unnecessary risks, or to wait an excessively long time for experimental results.

So we do the best we can, with the rational tools at our disposal. If you can think of a better method than the
one described here, we’d be happy to put a note in future editions of the eBook saying “The HSP estimator
method has now been superseded”. Till then, we think it has much to recommend it.

Some like it hot

The above analysis may sound a little academic. Yet whilst we were writing it, Hiroshi hit a painful problem.
Unfortunately he hadn’t carried out a proper HSP risk assessment before undertaking a task, and his hand was
in pain for 3 days afterwards.

You see, Hiroshi loves cooking with chillies. He decided to make a large amount of chilli sauce and ground up
this large supply of chillies.

125
He decided that PE gloves would be a good barrier to the capsaicin in the chillies, but quickly discovered that
they were easily damaged, so he swapped to latex gloves. After some time handling the chillies he found that
his hands were hurting from capsaicin that had got through the gloves.

We then decided to work out what gloves he should have used. Happily, we’d done the work already. When we
loaded the .mol file for capsaicin into the Y-MB estimator (it’s included in the Examples folder if you want to
try it yourself), we found the estimate was similar to Cyclophosphamide and the other cytotoxic chemicals.

Therefore we can recommend that the next time he has to make chilli oil, he should use two pairs of Nitrile
gloves or, if he doesn’t mind the discomfort, PE on the inside and latex or Nitrile on the outside.

Whilst we’re on the subject of chillies, we can address another important question. If you accidentally eat too
much hot chilli, what is the best way to remove the pain?

Because the HSP of water is too far from capsaicin, the old favourite, cold beer, is clearly useless. Ethanol is
not a great match, but is much closer, so a sip of neat vodka will be helpful. You often hear people say that
capsaicin is “soluble in oils”. This is only partly true. Simple oils and fats such as olive oil or lard ~ [16, 1, 5]
are too far away in HSP distance to be very effective. Indeed, Hiroshi experimented with extracting capsaicin
with olive-oil and found it made a very weak solution. However, they are better than water so that’s one
possible reason why milk and yoghurt are so often recommended as a good way to remove the sting of chilli. It
seems likely that the proteins in milk are a reasonable HSP match with capsaicin and maybe that’s the real
reason that milk/yoghurt are recommended. However, if it’s the fats that are important, remember not to use
low-fat milk/yoghurt during your chilli crisis.

126
Chapter 20 Gloves: from minimum data to maximum insight (Exploring
Glove Permeation)
The real world is often full of incomplete data. Like it or not, we have to make decisions with the data we’ve
got. This chapter extends the theme from the previous one and shows how HSPiP can take imperfect data and
with some good theory, a good modeler and some persistence, create some solid understanding of an important
topic – the solvent resistance of gloves.
There are two inter-related criteria for judging the safety of gloves: the breakthrough time and the steady state
permeation rate (SSPR). As we will see, ASTM judge the breakthrough time when the permeation rate reaches
a certain limit. To understand the importance of HSP for the performance of protective gloves, we’ll remind
ourselves of the equation for the SSPR.
SSPR = Dav(C1 – C2)/L [g/cm2/s ]
Dav = D0Dlm [cm2/s]
C1 is the surface concentration (g/cm3 dry polymer), C2 is assumed to be zero, D0 is the diffusion coefficient
(cm2/s) at essentially zero concentration (or where the diffusion coefficient is indeed a constant), and L is the
(dry) film thickness (cm). The log mean average of the diffusion coefficients at C1 and C2, is called Dlm. This
accounts for changing diffusion coefficients within the film from Dmax at the exposed side to D0 at side 2 for a
typical exponential dependence on concentration:
Dlm = (Dmax/D0 – 1)/ln(Dmax/D0) [dimensionless]
HSP immediately affect the SSPR because the surface concentration is higher for a better HSP match. This
surface concentration is almost always assumed to be the equilibrium concentration as measured by swelling
tests (but see below for acetonitrile in Viton for a case where it is not).
Although we know that HSP don’t affect D directly (for similar molecular shapes, a molecule with a bad HSP
match will diffuse at the same speed as one with a good match), D can increase greatly as the concentration
increases, so a good HSP match indirectly leads to a high D through this effect.
When D is concentration dependent, calculations are best done with a full-scale modeler such as the one in
HSPiP. But if you are keen to do things from formulæ then it is possible to calculate the steady state permeation
rate using those given above.
The breakthrough times also have the double dependency on polymer solubility and therefore a double
dependency on HSP. Fortunately the HSPiP diffusion modeler can readily handle this complex situation. The
main problems are connected with getting the correct data to put into the calculations. There is very little
information in the protective clothing literature on diffusion coefficients, much less how they change with the
local concentration.
Estimates of the equilibrium swelling concentrtion and D are required to understand what is going on within the
glove, and when to be concerned with events on the “safe” side of the protective elastomeric film. When such
data are lacking, simple but judicious use of HSP correlations of swelling, breakthrough times, or permeation
rates are fully justified, because of the very close connection between the HSP and these phenomena. That’s the
logic behind the choice of gloves in “It’s your call”.
Help in estimating D0 for a given chemical can be found in Figure 1. This figure ranks (from top left to bottom
right) the amount of solvent retained in vinyl and acrylic films for very long times (months). This is, in effect, a
ranking of their diffusion coefficients at low concentration and essentially the same ranking will be found in
films made from different polymers. It is the activation energy required to cause polymer chain segmental
motion that determines how readily suitable free volume becomes available for movement of the solvent
molecule. These activation energies will be different for different polymers, but the ranking of solvent diffusion
coefficients will not change significantly. The larger, cyclic and branched molecules will require more free
volume in order to move in any polymer so there is a larger activation energy for these molecules than for
smaller and more linear molecules. Figure 1 provides a method to interpolate/extrapolate diffusion coefficients
(D0) on a relative basis in any polymer. The key thing to remember is that although the relative order will
remain constant between polymers, the size of the effect will be highly polymer dependent. In a very rigid
polymer the difference in diffusion rates between large and small molecules will be much larger than for a
127
highly flexible polymer. The ranking by Shell reported in the Factor 4 section of the Diffusion chapter provided
independent confirmation of this fact.

Cl
H3C CH3
O
+
N
O O H3C O CH3
O
Cl O
O
H3C O CH3 CH3
HC H3C
H3C3 CH3
CH3

OH O H3C
HO
O
O CH3
+
N
CH3 O O
H3C
CH3 CH3
O
O CH3 HO CH3
+ O
N
CH3 CH3 O O O

O
OH
CH3 O HO H3C
O CH3
O
H3C CH3

Figure 20-1 Relative solvent retention in vinyl and acrylic films as a function of the size and shape of the solvent molecules.
(Source: Hansen, C.M., Doctoral Dissertaton, Technical University of Denmark, 1967, available as a PDF file on www.hansen-
solubility.com by clicking on the cover page). Cyclohexyl chloride is retained more than cyclohexanone which is retained more
than diacetone alcohol, etc. with methanol being the solvent that gets out fastest.

Analysis of permeation data to estimate the diffusion coefficient concentration profile


Diffusion coefficients must be low if a glove is to offer good protection. This implies that either the HSP match
must be (very) poor to keep the surface concentration (equilibrium swelling amount) down, or else the
challenge chemical has a large molecular size.
Because we don’t have all the required data to hand in order to explain permeation rates and breakthrough times
we have to find a rational process for providing adequate fits to whatever data are available. The trick is to find
(concentration dependent) diffusion coefficients and (hopefully from independent measurements) surface
concentrations that fit both the equilibrium permeation rates and the breakthrough times. Experience shows that
the room for adjusting the parameters to attain good simultaneous fits is surprisingly small, giving confidence
that the resulting values are meaningful.
When inputting the concentration-dependent diffusion coefficients there is no strict requirement to enter values
for concentrations above the surface concentration. However it is good practice to make sure that the curve
covers the whole range smoothly up to 100% solvent which of course is limited to some value below that for
the self-diffusion coefficient since the diffusion coefficients used in the modeler are based on dry film
thickness. As shown in the first figure of the diffusion chapter, the diffusion coefficients based on dry film
thickness will be somewhat lower than those based on total film thickness. At 100% solvent the diffusion
coefficient based on a dry film thickness becomes meaningless. This does not significantly affect the results
within the concentration range of usual interest, but should be recognized when solvent concentrations become
very high.
We have used permeation data reported in the Chemical Protective Clothing Performance Index, Second
Edition, by Krister Forsberg and Lawrence N. Keith (abbreviated F/K). Data from other sources have been used
as a supplement to these. The major problem with the F/K data is the lack of information on the surface
concentration from equilibrium swelling experiments. A second problem is that the criterion for the
breakthrough time is not given. The work reported here started with a breakthrough time defined at having a
128
concentration on side 2 of the glove equal to 0.1% of the equilibrium solubility. This was supplemented by use
of the ASTM F739-99a criterion for breakthrough which is a permeation rate of 0.1μg/cm2/min. This is equal to
1.67E-09 g/cm2/s, the units used by the HSPiP modeler.
In the figures below the red line reports the permeation rate at the given time. These values can be read
accurately using the mouse over the graph, allowing interpolation to find the ASTM breakthrough time. Values
defining the diffusion coefficient profiles were systematically changed until simultaneous agreement was found
with both the reported breakthrough time and reported permeation rate. Excellent agreement of this kind could
be found in all cases reported below for Nitrile, Butyl, Neoprene, Viton, and Natural Rubber gloves. In one case
discussed in detail below (acetonitrile in Viton) it was necessary to include a significant surface condition effect
to greatly improve the match with the literature data. A significant surface condition delays the breakthrough
while not (necessarily) affecting the steady state permeation rate. It was possible to match permeation data in
F/K for natural rubber in some cases, but the data for methanol, as discussed below, are difficult to make
consistent.
Our use of specific glove types in the following examples is merely so that the interested reader can identify the
specific data from F/K. They, in turn, emphasize that the “same” glove (e.g. “Nitrile”) from different
manufacturers can have different properties – though we’ve tried to ensure that our examples are of gloves that
are within the normal range expected of each type.
In all the figures that follow, the blue line gives the percentage of the saturated amount in the glove. The red
line follows the permeation rate to the equilibrium condition.
Toluene permeation in Nitrile glove Ansell Edmont type 37-165
Insight into the mechanism by which the Ansell Edmont Nitrile glove type 37-165 protects against toluene is
provided by Figure 2. The F/K data are: Breakthrough time 61 minutes, permeation rate 4000 mg/m2/min, and
thickness 0.64 mm. The equilibrium uptake of toluene in this glove type is 0.67 volume fraction (private
communication from Dr. Jaime Lara, IRSST, Montreal). Figure 2 was generated using these data by assuming
different diffusion coefficient profiles as a function of concentration. The range of possibilities among the
variables is surprisingly limited. Emphasis was given to the perfect calculated match of the experimental
permeation rate of 6.68E-06 g/cm2/s. Less emphasis was given to a simultaneous match of the breakthrough
time because the experimental determination can hardly be expected to exactly match the 0.1% concentration
assumption for breakthrough which gave a time of 64.3 minutes. The results found when matching the data
using the ASTM breakthrough criterion are discussed below. The agreement is surprisingly good in both cases.
A very interesting aspect is the diffusion coefficient profile. A straight line on the log D versus volume fraction
plot starting at D0 equal 1.15E-08 cm2/s and ending at 1.1E-06 cm2/s at the surface concentration was required
to match the performance data. The whole permeation process occurs in an elastomeric diffusion region. It can
also be seen that further extrapolation of this straight line to the pure liquid gives a diffusion coefficient near
1E-05 cm2/s which is very close to that expected for the pure liquid. This self-consistency is also an important
consideration for a correct estimate.

129
Figure 20-2 HSPiP model for the permeation of toluene in the Ansell Edmont Nitrile glove type 37-165. The literature data
indicate an equilibrium permeation rate of 6.68E-06 g/cm2/s versus 6.57E-06 g/cm2/s in this analysis. The literature indicates a
breakthrough time of 61 minutes by an unspecified experimental technique while this calculation gives 65.1 minutes. The
criterion here is when the concentration at side 2 reaches 0.1% of the surface concentration (saturated swelling amount) on
the exposed side.
Dibutyl phthalate permeation in Nitrile glove Ansell Edmont type 37-165
Figure 3 indicates what might be expected from a larger challenge chemical such as the plasticizer dibutyl
phthalate. Here the D0 is taken as 1E-09 cm2/s extending linearly as the logarithm of D to the pure liquid where
the diffusion coefficient is taken as being similar to that for many viscous liquids at 1E-06 cm2/s. The surface
concentration is based on an assumed degree of swelling of 500% that is typical for “good” solvents for this
elastomer. The breakthrough time is 6.3 hours compared with the F/K data for the system which simply states
>360 minutes. This shows that such low diffusion coefficients are not usually of greatest interest to protective
clothing. To reduce costs there is room for adjusting other parameters such as film thickness, composition,
crosslinking, etc. while still maintaining acceptable safety. What this calculation has shown is that D0 is not
larger than 1E-09 cm2/s, and probably is much lower depending on the actual breakthrough time.

130
Figure 20-3 HSPiP model for the permeation of dibutyl phthalate in the Ansell Edmont Nitrile glove type 37-165. The
literature data indicate an equilibrium permeation rate of <1.5E-05 g/cm/s which is easily satisfied by the 4.58E-06 g/cm/s
found here. The literature indicates a breakthrough time of >6 hours by an unspecified experimental technique while this
calculation gives 6.3 hours. The criterion used here is when the concentration at side 2 reaches 0.1% of the surface
concentration (saturation swelling amount) on the exposed side. The diffusion coefficients could be lower than those indicated
in the figure.
Figure 3 also provides insight into what happens with a short term exposure for such large molecules. There is a
possibility for considerable absorption at shorter times, which also strongly suggests that desorption of
plasticizers into challenge liquids will also occur to a significant extent near the surface, even for shorter
exposure times. The reason for this remark will become clear when we discuss the Butyl glove below. The first
curve in the lower left hand figure for concentration versus distance is for an exposure time of about 12
minutes, where the uptake is 15% of the saturation value with a penetration depth at about 20% of the film
thickness (0.64 mm).
The F/K data collection does list one Nitrile glove type, North LA.142G, where the breakthrough time for
dibutyl phthalate is given as >960 minutes for a thickness of 0.33 mm. An analysis similar to the one above
suggests the D0 in this case is less than 4E-13 cm2/s for the same surface concentration. It is quite possible that
this value is too low. The long breakthrough time with the thinner film could be caused by a lower surface
concentration possibly supplemented by a stiffer glove material. Lowering the surface concentration to 0.6
volume fraction and use of a D0 equal to 1E-10 cm2/s gives a breakthrough time of 18.3 hours, just exceeding
the breakthrough time requirement.
Methanol permeation in Nitrile glove Ansell Edmont type 37-165
The permeation of methanol in the Ansel Edmont Nitrile glove type 37-165 is analyzed in Figure 4. The F/K
data source gives a breakthrough time of 11 minutes with a permeation rate <9000 mg/m2/min, which converts
to <1.5E-05 g/cm2/s. The equilibrium swelling (data, again, from Dr Lara) is 0.32 volume fraction which is
taken as the surface concentration. The self-diffusion coefficient for methanol is close to 1.8E-05 cm2/s. Figure
4 shows a breakthrough time of 11.3 minutes and a permeation rate of 3.09E-06 g/cm/s. This permeation rate is
far below the limit given above, and presumably is a good estimate of the actual rate. This is, again, an
advantage of the mathematical modeling, because some permeation rate values may still be important, but may

131
not be measurable. These can be reasonably estimated with limited knowledge of the system. D0 is found as
2.4E-07 cm2/s, in agreement with expectations related to very small, linear molecules.

Figure 20-4 HSPiP model for the permeation of methanol in the Ansell Edmont Nitrile glove type 37-165. The literature data
indicate an equilibrium permeation rate of <1.5E-05 g/cm2/s which is in agreement with the 3.09E-06 g/cm2/s in this analysis.
The literature indicates a breakthrough time of 11 minutes by an unspecified experimental technique while this
calculation gives 11.3 minutes. The criterion here is when the concentration at side 2 reaches 0.1% of the
surface concentration on the exposed side.

Toluene permeation in Butyl rubber glove Best Company 878


Butyl rubber consists of about 98% isobutene and 2% isoprene with some variation in composition between
different sources. It is widely used for chemical protective gloves because of excellent resistance to commonly
used chemicals. These calculations used the data of Evans et.al.(Katherine M. Evans, Wumin Guo, and James
Hardy, Modeling Solubility Parameters and Permeation Data of Organic Solvents Versus Butyl Gloves from
Four Manufacturers, J. Appl. Poly. Sci., Vol. 109, 3867-3877 (2008)). This is a very complete study and again
there is significant variation among the suppliers. We use their toluene data for the Best company type 878
glove in order to show the differences between Butyl rubber and the Nitrile glove studied above. The targets are
a breakthrough time of 26.8 minutes (ASTM criterion) and a permeation rate of 1.29E-05 g/cm2/s. The
equilibrium solubility is essentially the same as in the nitrile gloves at 0.865 volume fraction. Because of the
almost identical solubility, it can immediately be seen that the diffusion coefficients will be higher for this
Butyl glove than for the Nitrile glove discussed in the above. This is because the breakthrough time is shorter
for a somewhat thicker film (0.75 mm versus 0.64mm).
A factor of some apparent importance for the performance of this glove type is that Evans et.al. found about 19
percent weight loss after careful removal of all solvent absorbed during the permeation testing. The glove
contains about 19% plasticizing material. During a permeation test some of this plasticizing material will be
partly removed from the exposed surface, thus reducing the inherent diffusion coefficient(s) attributable to the
original material in this region.

132
The diffusion coefficients at concentrations just under the surface concentration were lowered in an attempt to
simulate removal of plasticizer from the surface region of the glove. This was done by using a break in the
diffusion coefficient profile at a concentration below the surface concentration. This lower slope at higher
concentrations allows better fits of the experimental data than for the previously used straight line profiles,
indirectly confirming the loss of some plasticizer.

Figure 20-5 Permeation of toluene in the Best company glove type 878. The break in the diffusion coefficient profile gives a
better fit of the data, and could reflect migration of plasticizer from the glove to the toluene. The ASTM criterion give a
breakthrough time of 26.7 minutes versus the target 26.8 minutes. The target permeation rate of 1.29E-05 g/cm2/s is well
matched by the calculated permeation rate of 1.31E-05 g/cm2/s
Figures 6 and 7 repeat the examples reported in Figures 2 and 4, but with the ASTM criterion as a target for the
breakthrough time. For comparison the ASTM breakthrough criterion would give breakthrough times for
toluene in the Nitrile glove in Figure 2 equal to 56.8 minutes and for methanol in the Nitrile glove in Figure 4
equal to 5.7 minutes.

133
Figure 20-6 Repeat of the permeation calculation for toluene in the Ansell Edmont Nitrile glove type 37-165. The ASTM
criterion for breakthrough gives 60.7 minutes by interpolation of the red curve to a permeation rate of 1.67E-09 g/cm2/s. The
target is 61 minutes. The diffusion coefficient profile is essentially unchanged from that used above in Figure 2. The
equilibrium permeation rate is 6.4E-06 g/cm2/s. This is very close to the target value of 6.68E-06 g/cm2/s. Note that the grid
line density is extended to 300.

134
Figure 20-7 Repeat of the permeation calculation for methanol in the Ansell Edmont Nitrile glove type 37-165. The ASTM
criterion for breakthough times gives 11.3 minutes to compare with the 11 minutes given in the Forsberg/Kieth performance
handbook. The permeation rate of 1.67E-06 g/cm2/s is much lower than the <1.5E-05 g/cm2/s required by the data. D0 has been
decreased from 2.4E-07 cm2/s to 1.3E-07 cm2/s to accommodate the change in the breakthrough criterion while maintaining
the same equilibrium permeation rate.

Acetone permeation in Natural Rubber gloves


There are data in the F/K Chemical Protective Clothing Performance Index for the permeation of many
different chemicals through natural rubber gloves, but many chemicals degrade the natural rubber. It should
perhaps be noted again that there is wide variation in performance among the gloves. Solubility data provided
by Dr. Lara have been used to try to select a system to study where the solubility is not too high. The glove type
for which solubility data is available is Ansell 356, but this glove type is not included in the F/K source. This
glove has a solubility of acetone equal to 0.162 volume fraction.
Figure 8 below matches the target breakthrough time (5 minutes) very well at 5.3 minutes with a permeation
rate of 1.71E-06 g/cm2/s where the target is 1.84E-06 g/cm2/s. This is good agreement, and shows that
reasonable estimates can be used to develop the diffusion coefficient profiles in natural rubber.

135
Figure 20-8 Permeation of Acetone in natural rubber. F/K data used are for Ansell Edmont 46-320. The breakthrough time
for this glove type in the F/K data is 5 minutes for a permeation rate of 1.837E-06 g/cm2/s.
Methanol permeation in Natural Rubber gloves
The F/K data source shows a wide variety of data for the permeation of methanol through different glove
qualitites. Ansell Edmont glove 46-320 was selected for this study. The reported breakthrough time is 20
minutes and the permeation rate is 1.34E-07 g/cm2/s for a thickness of 0.31 mm. The results in Figure 9 for a
surface concentration given by Dr Lara as 0.086 volume fraction are in excellent agreement with the literature
data. The diffusion coefficient profile yielded an ASTM breakthrough time of 20.0 minutes. The calculated
permeation rate is 1.32E-07 g/cm2/s.

136
Figure 20-9 Permeation of methanol in Natural Rubber. The given diffusion coefficient profile yielded an ASTM
breakthrough time of 26 minutes versus the reported value of 20.0 minutes. The calculated permeation rate is 1.33-07 g/cm2/s
versus the reported value 1.34E-07 g/cm2/s.
It should be noted here that the methanol permeation data for some glove types could not be matched. An
example of such a situation is Ansell Orange 208. The reported breakthrough time is 17 minutes, not far from
that in the previous example, but the permeation rate is reported as 1.5E-09 g/cm2/s. This permeation rate is
lower than the ASTM breakthrough criterion of 1.67E-09, and could simply not be matched by a reasonable
diffusion coefficient profile. This reported permeation rate is 89 times lower than that reported for methanol in
the natural rubber used in the previous example, and this for comparable breakthrough times. No explanation of
this is obvious.
It is concluded, however, that permeation through natural rubber gloves can be modeled by the HSPiP software,
with any exceptions to reasonable results requiring an explanation.
Methylene chloride permeation in Viton gloves
The data source for the following calculations for methylene chloride (Figure 10), chloroform (Figures 11a and
11b), and acetonitrile (Figure 12) in Viton is:
Katherine M. Evans and James K. Hardy, Predicting Solubility and Permeation Properties of Organic Solvents
in Viton Glove Material Using Hansen’s Solubility Parameters, Journal of Applied Polymer Science, Vol. 93,
2688-2698 (2004). This is a very careful and thorough study providing the data needed.
The breakthrough time for methylene chloride in the given Viton glove material was 38.4 min with a steady
state permeation rate of 1.15E-06 g/cm2/s. Figure 10 shows that these values can be duplicated with the
diffusion coefficient profile given in the figure. The ASTM breakthrough time is calculated as 38.3 minutes
with a permeation rate of 1.13E-06 g/cm2/s.

137
Figure 20-10 Permeation of methylene chloride in Viton. The agreement is excellent with the reported breakthrough time of
38.4 minutes being matched with 42.1 minutes, and the reported steady state permeation rate of 1.15E-06 g/cm2/s was matched
with a calculated 1.13E-06 g/cm2/s.

Chloroform permeation in Viton gloves


Figure 11 gives a diffusion coefficient profile for chloroform permeation in Viton. There is a calculated ASTM
breakthrough time of 6.87 hours versus an experimental value of 6.90 hours and a steady state permeation rate
of 6.18 g/cm2/s compared with the value from Evans and Hardy of 4.15 g/cm2/s. It has not been possible to find
a set of parameters to better match both of these simultaneously. It is possible that a treatment similar to the one
presented in the following for the permeation of acetonitrile could improve the agreement, but at the same time
there is better agreement with another source of permeation data in this system also reported by Evans and
Hardy as shown in Figure 12. In this second case the calculated value of the ASTM breakthrough time is 9.3
hours versus an experimental 9.5 hours. The calculated steady state permeation rate is 4.53E-08 g/cm2/s to be
compared with an experimental value of 4.67E-08 g/cm2/s.

138
Figure 20-11 Permeation data reported by Evans and Hardy for a previous study for chloroform in Viton. In this case the
calculated value of the ASTM breakthrough time is 9.1 hours versus an experimental 9.5 hours. The calculated steady state
permeation rate is 4.58E-08 g/cm2/s to be compared with an experimental value of 4.67E-08 g/cm2/s.

139
Figure 20-12 There is a calculated ASTM breakthrough time of 7.1 hours versus an experimental value of 6.90 hours and a
steady state permeation rate of 6.17 g/cm2/s compared with the value from Evans and Hardy of 4.15 g/cm 2/s.
Acetonitrile permeation in Viton gloves
Evans and Hardy report an ASTM breakthrough time of 13.8 minutes and a steady state permeation rate of
1.95E-05 g/cm2/s for acetonitrile in Viton. Figure 13 gives the closest fit found for these values without trying
to explore what a significant surface resistance might do. The data in Figure 13 reproduce the breakthrough
time precisely at 13.8 minutes, but the calculated steady state permeation rate is 1.53E-05 g/cm2/s. While this
might be satisfactory and probably within experimental error, an attempt was still made to explore the effect of
a significant surface resistance.

140
Figure 20-13 Permeation of acetonitrile in Viton. Data of Evans and Hardy are reproduced almostexactly for the ASTM
breakthrough time of 13.8 minutes but the steady state permeation rate is low at 1.58E-05 g/cm2/s compared with the
experimental 1.97E-05 g/cm2/s.
Figure 14 shows the results of a significant surface effect. The change in the input values is a big reduction in
the B value which in turn gives a limiting mass transfer coefficient, h of 1.27E-04. It is relevant to compare the
estimated diffusion coefficients for acetonitrile (molar volume 52.6) with those of methylene chloride (molar
volume 63.9). One would expect similar diffusion coefficients based on similarity of size (and shape) when
comparing these two solvents. D0 for methylene chloride is estimated in Figure 10 as 1.1E-08 cm2/s. D0 for
acetonitrile estimated with the inclusion of a significant surface effect is 9.5E-09 cm2/s , which is close to this,
while without the significant surface effect, the D0 is 2.2E-09 cm2/s . This expected similarity also speaks for
the presence of a significant surface effect. The much larger equilibrium absorption of the acetonitrile is the
main difference between these two solvents in the present context. This allows for very rapid diffusion within
the Viton, and the supply through the surface cannot keep up with the removal from the surface.

141
Figure 20-14 Permeation of acetonitrile in Viton modeled with a significant surface resistance generated with a surface mass
transfer coefficient, h, of 1.27E-04 cm/s. The ASTM breakthrough time here is 14.1 minutes compared with the experimental
13.8 minutes and the steady state permeation rate is 1.88E-05 g/cm2/s compared with 1.97E-05 g/cm2/s. The surface
concentration at the steady state permeation condition is 0.832 times the solubility (0.528 volume fraction).
If one accepts this result, the effects of film thickness can be explored by maintaining the same h, and changing
the thickness. The effect of putting on two gloves, for example can be estimated by entering a thickness of 0.82
mm and, neglecting any resistance between the gloves. Here one finds an ASTM breakthrough time of 39.1
minutes and a steady state permeation rate of 1.32E-05 g/cm2/s. The surface concentration is 0.882 times the
solubility. The rule of thumb that the breakthrough time will be four times as large for a doubling of the film
thickness would have indicated a breakthrough time of 4 times 13.5 which gives 54 minutes or 13.8 times 4 as
55.2 minutes. The reason for this difference is the surface condition has less effect on thicker films on a relative
basis, and the higher concentrations in the film lead to higher diffusion coefficients. The gloves are not quite as
safe as expected based on the rule of thumb.
A similar analysis of the steady state permeation rate would suggest that doubling the film thickness would cut
the permeation rate in half. This is 0.98E-05 g/cm2/s compared with the 1.32E-05 g/cm2/s estimated by this
calculation. Again the glove(s) are not as safe as predicted by the rule of thumb because there is more solvent in
the glove, and the diffusion coefficients are higher than in the reference measurement.
These calculations should clearly be tested by experiment.

142
Figure 20-15 Permeation of acetonitrile in Viton modeling the effect of wearing two gloves.
The ASTM breakthrough time is 39.1 minutes and the steady state permeation rate is 1.32E-05 g/cm2/s for a
surface concentration that is 0.882 times the maximum amount that the Viton film can dissolve.
The time-lag
The time-lag calculation, now available in the Third Edition, was not used in any of the previous figures in the
chapter. The values that would have been found are reported in the following table for permeation in Viton
gloves only. The study by Evans and Hardy is the only source of the experimental information used for this
purpose, and this is for Viton gloves only. The time-lag is the time value found by extrapolating the
accumulated permeation rate back to zero amount permeated as seen in Figure 17-16.
Figure Solvent Experimental Time-lag Calculated Time-Lag
18-10 Methylene chloride 55.7 min. 72.6 min.
18-11 Chloroform 677 min. 1194 min.
18-12 Chloroform 661.5 min. 984 min.
18-13 Acetonitrile 25.6 min. 15.1 min.
18-14 Acetonitrile 25.6 min. 16.1 min.
18-15 Acetonitrile - -
18-16 Acetonitrile 25.6 min. 25.2 min.
Table 20-1 Time-lags, experimental and calculated
From the above table only marginal change can be seen comparing the time-lag of 15.1 minutes from Figure
18-13, where there is not significant entry surface resistance in the calculation, to the 16.1 minutes found in

143
Figure 18-14, where there is. Both are still well below the experimental 25.6 minutes. The calculated and
experimental steady state permeation rates and breakthrough times match well for the calculations in the latter
figure, however. The time-lag data for the other figures is calculated as being much longer than measured. The
reason for this is not known, but there is a strong suspicion that this is caused by surface resistance, perhaps
mainly on the exit side of the samples in permeation cells as discussed in the following.

Figure 20-16 Test of the lag-time calculation using the system acetonitrile in Viton as in the previous figures. The
experimental time-lag is 25 minutes in adequate agreement with the calculated 20 minutes, and calculated breakthrough time
of 13.1 minutes matches the experimental value of 13.5 minutes. However, the steady state permeation rates are not well
matched with 1.26E-05 g/cm2/s calculated versus 1.95 g/cm2/s found experimentally.
Figure 17-16 gives data that match the lag-time and breakthrough time very well, including a significant surface
resistance on the entry side, but the steady state permeation rate is much too slow. Matching the experimental
and calculated permeation rates in Figure 17-16 would require a significantly higher surface concentration
while keeping the same diffusion coefficient profile that gave the good matches in the time-lag and
breakthrough time. This is not possible. Attempts to reconcile this in various ways did not lead to an improved
result. It should be noted, however, that the there are only two relevant entries in the F/K database for
acetonitrile in Viton (both for North F-091). The one lists degradation and the other a more rapid breakthrough
(6 minutes) than in the example above, although the film thickness is smaller (0.26 mm). This leads one to
suspect that the breakthrough and time lag data reported in Figure 17-16 are more reliable than are the steady
state permeation data, and that some degradation in the Viton may be responsible for the higher experimental
steady state permeation rate than is predicted by these calculations.
At this point in time there is too much variation in literature data and too many variables to properly assign
accurate values to all of the relevant parameters, but each of the parameters discussed has, or easily can have
significance, particularly for more rapid permeation in thinner samples.
Breakthrough times for low concentration exposures
It is possible to predict effects that are very difficult to measure when reliable diffusion coefficient profiles have
been established. One of these is how quickly breakthrough occurs for a very limited exposure to a challenge
chemical. This is not easily estimated from general permeation data which are typically measured at high
concentrations with pure liquid exposure in permeation cells. The following tables report breakthrough times
144
for low (constant) surface concentrations for the Nitrile glove type studied in the above using the diffusion
coefficient profiles for toluene and methanol that were established in Figures 2 and 4. Like all numerical
modelers there is some trade-off of accuracy (grid size) and speed so the values reported below are somewhat
grid dependent. However it should also be noted that differences in breakthrough times for toluene with
different types of Nitrile gloves far exceed any numerical uncertainties. For toluene the variation is from a low
of about 4 minutes up to a high of 61 minutes. For methanol the variation is from a low of 7 minutes to a high
of >480 minutes. One must be very careful with generalizations. The following is an indication of what is
generally expected, but cannot be considered a complete analysis.
C1 Breakthrough time Dmax/D0
0.00 - 1.00
0.01 496 1.04
0.02 486 1.13
0.05 452 1.39
0.10 402 2.0
0.67 64.3 (61) 86.9
Table 20-2 Breakthrough times in minutes for toluene in the Nitrile glove Ansel Edmont type 37-165 for
given values of the surface concentration, C1, in volume fraction.

C1 Breakthrough time Dmax/D0


0.00 - 1.00
0.01 21.3 1.04
0.02 21.3 1.08
0.05 20.3 1.25
0.10 19.3 1.54
0.32 11.3(11) 4.08
Table 20-3 Breakthrough times in minutes for methanol in the Nitrile glove Ansel Edmont type 37-165
for given values of the surface concentration, C1, in volume fraction.
These data emphasize once more the importance of keeping diffusion coefficients low, one way or another, to
keep breakthrough times long.
Practical implications
It’s a good idea to remind ourselves why we’ve gone to such lengths to analyze all that data. The reason is that
it matters! The safety of real people with real chemicals depends on these analyses.
The problem is that there are many more chemicals than there are people to do the careful measurements of
breakthrough times and permeation rates. We therefore need rational ways to arrive at the best choice of glove.
We’re already off to a good start. From the above analysis we can say that the majority of the D0 values of
interest to glove safety will lie between the limits of about 1E-09 cm2/s for larger molecules and about 1E-07
cm2/s for smaller molecules with an average near 1E-08 cm2/s. Figure 1 gives you an idea of how to estimate
relative D0 values if (as is usually the case) you don’t have the resources to measure them yourselves.
It is also clear that a greater HSP distance is required for safety when the molecules are smaller or more linear.
Solvents with smaller molecules often appear as outliers in simple HSP correlations, getting through the glove
more quickly than expected. Conversely experience has also shown that the phthalate plasticizers, for example,
do not get through given glove types quickly, in spite of good matches in HSP with the glove materials. The
reason for this is clearly portrayed above as a very low D0.

145
For molecules beyond simple solvents and plasticizers, a method to improve the simplest HSP analyses in such
situations has been given in Chapter 12. This uses the example of Neoprene gloves with a 1 hour breakthrough
time. HSP can lead to reasonable predictions of breakthrough times for complex chemicals. If you have the
luxury of more data then you can use the techniques described in this chapter to provide even better
information.
Wearing two pairs of gloves, or a gloves of twice the thickness, gives four times the protection against diffusion
and gives extra protection against accidental tears or punctures.
There will be times when simple estimates can lead you astray. As the acetonitrile/Viton example shows,
without knowing that there is a strong surface resistance, the “twice the thickness gives four times the
protection” guide leads to a false conclusion. However, most of us would not want to use gloves that so
obviously loved to absorb so much solvent. Once again, a large HSP mismatch is a much better idea.
Because D0 is usually lower in more rigid polymers than in more flexible polymers, the cautious user might like
to go for a glove made from, say, polyethylene. But of course its very rigidity (lack of comfort) and the ease
with which it is torn are often good reasons for not using it.
In summary, as in the previous chapter, the intelligent use of whatever data is to hand can lead to reasonable
and defensible conclusions. In the absence of any other data then the best rule of safety is to maximize the HSP
distance between the glove material and the challenge chemical with a glove material that is as flexible and
comfortable as possible.

146
Chapter 21 Saving the planet (Finding Improved Environmental
Solutions)
We all try to do our bit to save the planet. Although a few grand gestures can make a difference, mostly we
have to do it one step at a time.
HSP are proving versatile in their ability to contribute to lots of small steps. So here we have a selection of
examples.
HAP substitutions
The USA lists chemicals considered to be Hazardous Air Pollutants. If you happen to be using one of these as a
solvent then there is considerable pressure for you to substitute it with a non-HAP variant. How do you do this
sensibly?
The key is functionality. You need to have the same solvating power. It’s unlikely that there is a single solvent
that can be used to replace your current solvent, so you need to find a mix of solvents that will do the same job.
Without HSP your task is very tricky.
Suppose you are using diethanolamine, with a low (i.e. serious) HAP reporting level. What mixture of solvents
will match its particular HSP? The Solvent Optimizer quickly gives you an idea.

Figure 21-1 Approximating Diethanolamine


The target values for diethanolamine are [17.2, 10.8, 21.2] and by clicking the “2” button you instantly find that
the 2-best solvents for giving a match are a 77/23 mix of ethylene glycol and cyclohexanone. Unfortunately,
ethylene glycol is also on the HAP list so you want to exclude that from the list. Ctrl-clicking on ethylene
glycol grays it out and re-clicking “2” gives you a methanol/benzyl alcohol mix, but methanol is also on the
HAP list, so you Ctrl-click that and try again to find that a Dipropylene Glycol (DPG)/ethanol mix is a good
match. Now it’s unlikely that you would have thought of such a mix, and you may not even be familiar with
DPG. But who knows, this mix might be just what you’re looking for.
The point of this example is that with HSP (and some convenient software tool) it becomes routine to think
through these tricky substitution issues in a methodical and environmentally-friendly manner.
VOCs (Volatile Organic Compounds)
The rational approach to reducing the harmful effects of VOCs is to find chemicals of the correct chemical
compatibility and with both a low vapour pressure and a low reactivity with ozone (as measured by the Carter
MIR) and with OH radicals (Log[OH]).
HSPiP can give you all of this in one module. The Y-MB calculator not only estimates the HSP of a chemical
of interest but uses its neural networks to estimate vapour pressure (at 25º) as a convenient judge of how low
the vapour pressure might be, plus it calculates the Carter MIR and Log[OH]. As an additional guide, from the
vapour pressure an estimate of the RER (Relative Evaporation Rate – nBuAc=100) is made via the empirical
formula RER=0.046*MVol*VapourPressure. We’ve been asked why MVol is necessary for calculating the
RER. The Vapour Pressure is, essentially, in moles. If you have two chemicals with the same vapour pressure,
147
but one has twice the molar volume, then there will be twice as much of this second chemical in the air. It’s as
simple as that! Another question is how the RER based on nBuAc=100 is related to that based on ether. The
answer is that the ether-based system is the time needed to evaporate rather than the rate. nBuAc takes 12 times
as long to evaporate as ether so
RER-based-on-nBuAc=12/Ether-based-system (or 1200 if ether is 1 in its system).
Bio-solvents
What better way to replace some bad solvent than by some bio-solvent? Well, there are arguments for and
against bio-solvents, but at least we can agree on an objective value – their HSP. If the HSP is wrong then
however bio-friendly they are, they simply won’t be able to do the job.
A typical example is the proposed use of FAME solvents – Fatty Acid Methyl Esters. We use the data from a
typical paper on the subject where FAME solvents were tested for applications of epoxy resins, Y Medina
Gonzalez et al, Fatty Acid Methyl Esters as Biosolvents of Epoxy Resins:A Physicochemical Study, J Solution
Chemistry, 36, 437-446, 2007.
One of the target resins was an epoxy resin called Bisphenol A diglycidyl ether. The authors used a classic 40-
solvent HSP Sphere process to determine values shown in the table. The table includes our S-P estimates of the
HSP for two of the FAME materials.
Chemical δD δP δH R (or Distance)
Epoxy 17.2 6.9 8.2 24
Methyl caprylate (C8) 15.5 2.8 5.6 6
Methyl myristate (C14) 15.4 2.2 4 7.2

Table 21-1 FAME solvents compared to an Epoxy


This would seem to suggest that FAME would be good solvents for the epoxy and that there wouldn’t be a huge
difference in solubility between the C8 and C14 versions. But the solubilities turn out to be disappointingly low
(though they at least diminish as chain length gets longer, as would be expected). Is this a failure for HSP? The
large R for the Epoxy looks a little odd. With a more typical value the distances would be closer to the border
of the sphere and the relative insolubility would be more explicable. And, of course, the rather large molar
volumes of the FAME solvents mean that kinetically they will be rather poor – one of the key problems that
beset the change to “greener” solvents.
Rational extraction
Work carried out in Professor Staci Simonich’s team in Oregon State University T. Primbs, S. Genualdi, and
S.M. Simonich, Solvent Selection for Pressurized Liquid Extraction of Polymeric Sorbents Used in Air
Sampling, Environmental Toxicology and Chemistry, 27, 1267–1272, 2008 is a good example of how a
complex environmental sampling procedure can be rationalised via HSP. Anthropogenic SOCs “Semivolatile
Organic Compounds” travel around the world from their sources in one region to provide pollution in another
(often pristine) region. It is a substantial challenge to collect them then to extract them quantitatively. She uses
pressurise fluid extraction, but her problem is to ensure extraction of the SOCs without extracting polyurethane
(PU) which happen to be one of the key components of the system for trapping particulate and gaseous
materials from the air stream.
By plotting a range of typical SOCs, along with typical solvents used in pressurised fluid extraction, choosing
PU in the Polymer form and clicking the Solvent Match button, the 3D plot makes it clear that dichloromethane
is too close to the PU to be of use in extraction. The 50/50 hexane/acetone looks, from the H-P plot to be a
potential problem with the PU but the RED number shows that it too is safely outside the R of this PU and so
should be useable. The favoured solvents, hexane, acetone and 75/25 hexane/acetone seem, from a visual
inspection of the 3D plot to be adequate for extracting the SOCs themselves.

148
Figure 21-2 Using file OSU Environmental
Predicting problems
The problem of contaminants reaching the Arctic via routes such as the one above is a major one. It has been
known for some time that pentachlorophenol [21.5, 6.9, 12.8] not only gets to the Arctic but also accumulates
to high levels in polar bears. One of us, many years ago, spotted that the much-used fire-retardant tetrabromo-
bisphenol A [20,2, 9.1, 13.8] is very close (Distance=3.5) to pentachlorophenol and would therefore be highly
likely to start accumulating in polar bears. Unfortunately for the bears, and those who might eat the bears or the
seals with the same problem, this prediction has been vindicated. Of course the bromo flame retardants are now
being phased out, but the point of this small section is that HSP predictions can be used with some confidence
on issues that take a long time to resolve experimentally. Clearly one can’t make long-term environmental
decisions based on HSP alone, but when difficult decisions have to be made, the insights from HSP can be
crucial.
Plastic containers
Although the environmental benefits of plastic containers are contentious, there are plenty of times when they
are environmentally preferred to glass or tins. If you want to sell something for the first time in plastic, you
need to be confident that the contents will not damage the plastic by dissolving it or causing environmental
stress cracking. Because it’s relatively easy to test for such issues, there aren’t too many mistakes made. What
is not so obvious is when you have aqueous solutions of potent chemicals such as weedkillers. There’s no
chance of the solution dissolving the container or producing stress cracking. The problem is that any diffusion
of the weedkiller through the walls means potential skin contact for the user or leakage into the environment. A
casual examination of a typical tin of weedkiller tells you that one common ingredient is paraquat. From the
standard table we find its values are [19.5, 8.8, 5.9] Now let’s try to put this in a PET bottle. There are a number
of different PET values, but the R-PET data in the Polymers table is [19.1, 6.3, 9.1, 4.9] and the distance
between the paraquat and the PET is 4.1, inside the radius and therefore a warning that PET would not be a
good idea. To check out the polymers in general, select paraquat in the solvents table then click Polymers
within the Polymers form.

149
Figure 21-3 Selecting Paraquat from the main table then clicking Polymers
You will find that the other PET values (such as Mylar PET) also have a Score (in this case, the RED number
of the paraquat) less than 1. Nearer the top are polymers such as PMMA and PVC ( both REDs~0.35). Near the
bottom of the table are the high RED polymers such as Polypropylene and Polyethylene, so you could be
confident of no diffusion of the weedkiller through these bottles. If you think that the worry about dilute
chemicals in PET bottles is excessive, just think what happens if the lid is left off so that the water evaporates
and concentrates the chemical. What remains could be potentially very active in getting through the PET.
Bio-polymers
We know that common polymers have land-fill degradation times of 100’s of years. Bio-polymers can have
degradation times of months to a few years under the right conditions. But industries aren’t going to shift easily
to these new polymers. They have to know much more about them. One key thing that’s needed is an idea of
their compatibilities with solvents, plasticizers and other polymers. And the shortest route to identifying these is
via HSP. We use, for example, data from the paper A. Agrawal et al. Constrained nonlinear optimization for
solubility parameters of poly(lacticacid) and poly(glycolic acid)—validation and comparison, Polymer, 45
(2004) 8603–8612.

150
Figure 21-4 Using file PolyLacticAcid
Armed with this correlation, [18.7, 7.7, 7.0, 9.3] and with the solvents sorted by their RED number, it’s easy to
determine that a high-boiler such as NMP or a low boiler such as 1,3 Dioxolane would be a good choice of
solvent depending on the type of application. If solvent data hadn’t been available, using the polymer SMILES
for PLA, XC(C)C(=O)OX, gives the value [17.7, 8.7, 10.2] which would have been given a reasonable choice
of practical solvents. One of us (Abbott) has written a review (Poly(lactic acid): Synthesis, Structure,
Properties, Processing and Applications, Eds. R. Auras; L-T, Lim; S.E.M. Selke, H. Tsuji, Wiley, 2010) on
how HSP provide a coherent framework for explaining many real-world issues confronting this relatively novel
polymer – such as choice of plasticizer, filler dispersion/compatibility and odour/fragrance permeability.
Pervaporation
One way to remove contaminants from water is to apply high pressure to the contaminated liquid as it passes
through a tube surrounded by a polymer membrane through which the contaminant can diffuse but the water
cannot. After the permeation through the membrane the solvent then evaporates, hence the term pervaporation.
A specific example comes from Professor Yoram Cohen’s group at UCLA. They wished to remove
tetrachloroethylene and chloroform contaminants from water. The technological innovations in producing a
robust membrane supported on a ceramic tube need not concern us here. The key was the choice of the polymer
for the membrane.

151
Figure 21-5 Pervaporation through PVA
The polymer used was PVA (polyvinylacetate). As the data show, both chloroform and TCE have a RED <1
and are therefore likely to be able to pass through the membrane. Indeed, their pervaporation experiments
showed good pervaporation removal of the two contaminants in their system. Note the final solvent in the list. It
is not included in the plot but its RED number shows that it too should be removable by pervaporation. Indeed,
in a later paper, Cohen’s group showed good speed and specificity for removal of methyl-t-butyl ether which is
a modern-day contaminant in groundwater because of its use as an octane enhancer in petrol/gasoline as a
“safe” replacement for lead tetraethyl.
Soil chemicals
It’s often important to know whether a chemical will tend to stay attached to soil particles or prefer to partition
into the water trickling through the soil. The soil-water partition coefficient (relative to organic carbon), Koc,
has been measured for many compounds and a number of techniques for estimating Koc have been suggested. A
paper by Prof Schüürmann’s group at Leipzig has provided an overview of the various datasets available and
compared their own methods to those in the literature: G Schüürmann, R-U Ebert, R Kühne, Prediction of the
Sorption of Organic Compounds into Soil Organic Matter from Molecular Structure Environ. Sci. Technol.
2006, 40, 7005-7011.
Using the CAS numbers from their data table it was straightforward to identify 80 chemicals within our own
dataset for which there was Koc data. By using the simple formula (another variation on the χ parameter form)
Equ. 21-1 log Koc =-0.327 + 0.0008 * MWt * (4 * (δDc-δDs)2 + (δPc-δPs)2 + (δHc-δHs)2) ½
a respectable R² of 0.835 was obtained. In this equation, “c” is the chemical and “s” is the soil. δDs, δPs and δHs
were obtained by fitting as [13.8, 20.8, 9.2].

152
Calculated and Experimental Koc y = 0.8345x - 2E-05
2
R = 0.8348

5
Calculated Koc

0
0 1 2 3 4 5 6
Experimental Koc

Figure 21-6 Calculation of soil sorption


This quick excursion into soil science is not intended to prove that HSP are good predictors of partitioning into
soil, nor is any technical explanation being offered - it would need a much more extensive analysis. But it is
interesting that once again the same basic HSP equation seems to work adequately when applied to such a
complex problem. Note that the fit uses Molecular Weight and not Molar Volume. When the latter is used, the
fit is less good.
Radon
Homeowners in some parts of the world are unlucky. Their houses are built on granitic rocks that contain a lot
of radon from natural radioactive decay. It is therefore desirable to put a barrier film underneath a house to
inhibit the ingress of radon. Can HSP help with the problem? Surprisingly the answer is “yes”. The starting
point for thinking through the problem is the well-known fact that EVOH is an excellent oxygen barrier. Why is
this very modest polymer such a good oxygen barrier? Nothing about its polymer structure makes it an obvious
block to oxygen. And why is something like polyethylene such a poor oxygen barrier? As the Diffusion chapter
shows, permeation depends on diffusivity (which depends on polymer structure) and on solubility. And this is
why EVOH is so good. If you load Gases.hsd you find the following table:

153
Figure 21-7 Gas HSP from Gases.hsd
The HSP of oxygen are [14.7, 0, 0]. The HSP distance from polyethylene (PE) is a modest 5, inside the sphere.
The distance from EVOH is a massive 20. No wonder that EVOH is such a good oxygen permeation barrier.
The table also includes Radon at [17.2, 0, 0]. The distance from EVOH is 17.5 and from PE is 3. There is a
literal take-home message from this: don’t protect your house from Radon with PE film. However, EVOH is a
useless barrier for water as it readily swells. So you can’t protect your home with a pure EVOH barrier – you
need a water barrier to protect the EVOH. Try a PE-EVOH-PE laminate as the HSP calculations are totally
opposite – the distance of PE from water is >20 so the water solubility, and hence permeability is very low.
HSP have one more thing to contribute. That laminate is unlikely to exist because the distance between EVOH
and PE is so large. Therefore you’ll need a tie layer of intermediate HSP.
It’s also a good idea to load Gases.sof into the Solvent Optimizer, enter the HSP of your polymer as the Target
then click on the Sol. Graph button to see the relative solubilities of the different gases in your polymer. It gives
you a good overview of the extent to which different gases are likely to want to permeate through your
polymer.
Geomembranes
If some land has been contaminated with chemicals and new buildings (e.g. houses) need to be built on the land
then it is important to be sure that the chemicals will not pose a long-term hazard. One way to do this is to
provide a geomembrane which reduces the permeation rate to a minimum for all the relevant chemicals.
Classically this has been done with HDPE, but HSP users will realise that this cannot possibly be an adequate
barrier for many contaminents such as BTEX (Benzene, Toluene, Ethyl benzene, Xylenes). But then no single
barrier polymer will ever be good enough as a good barrier for BTEX will tend to be a poor barrier for
contaminants with high δP and δH values. The answer, as with radon, is a barrier containing multiple polymers.
A specific use of HSP in designing and validating such a geomembrane is provided by Steven Abbott, Charles
Hansen, Richard Menage, Steve Wilson, A robust methodology for reducing organic chemical vapour
contamination risks in buildings, Geosynthetics International, in press 2013 and a software tool based on HSP
and a multilayer diffusion model, Puraflex (2011) Puraflex Permeation Modeller Software, www.puraflex.com.
Rational substitution and Read Across
It is very difficult to substitute one molecule with another or do a full comparison of two molecules in order to
“Read Across” the health and safety values from one known compound to an unknown one. In HSPiP we bring
together many of the strands of this and other chapters to make such comparisons powerful and easy. By
entering the SMILES values of two chemicals, the Y-MB method provides a whole range of predicted values.
154
Of course comparison of HSP is vital and the HSP Distance is automatically provided. Important values such as
vapour pressure and RER (Relative Evaporation Rate) are key for those interested in VOCs, along with the
MIR and OH values for VOC reactivities. Various partition coefficients are estimated: log Kow (octanol/water),
log Koc (soil absorbtion) and BCF (Bio Concentration Factor). Remember that the RER calculated is based on
nBuAc=100. If you need the ether-based system then the formula is
RER-based-on-nBuAc=12/Ether-based-system (or 1200 if ether is 1 in its system).
Armed with all this information, HSPiP users are in much better shape to make informed choices affecting
health, safety and the environment. See the Predictions chapter for more information on each of the predicted
parameters.
QSARs
If you want to know how safe or unsafe a chemical is in terms of toxicity to aquatic life, rats or humans, and if
you have a large dataset of tox data, it is possible to make some hypotheses about which factors might be
important in controlling toxicity. HSPiP’s QSAR tool, discussed in a later chapter, is a powerful way to look for
correlations. One trick in the tool is simply to plot the data against a single parameter – this is easily done by
clicking on the column header. Because people often believe that LogKow is important you can instantly check
for a correlation by clicking on the LogKow header. In one of the datasets provided (rat LD50 for alcohols and
phenols) the graph shows a V-shape, with medium values being more toxic than higher or lower. But if you
click on the MWt column you see exactly the same shape. LogKow is strongly size-related. Not surprisingly,
using HSP or HSP Distance one can find some good correlations.
Incidentally, both the EPA and REACH refer to Hiroshi’s Pirika site as a source of LogKow estimated values.
You will be pleased to know that the same prediction algorithm is used within HSPiP!

155
Chapter 22 HSP for ionic liquids (How to Assign HSP to New Materials)
There is a popular notion that one contribution to saving the planet is to use ionic liquids as “green solvents”.
This chapter describes a work in progress and, because of this, is in a different format. We would love to
provide the HSP community with a definitive list of ionic liquid HSP, but the data are simply not there to
produce such a list. We hope that this chapter will stimulate debate within the ionic liquid community and,
especially, create more datasets from which HSP can be derived.
This chapter is written in a form that is intended to serve as a guide for determining the HSP of given chemicals
using ionic liquids as examples. The reason for this choice is that ionic liquids combine special properties in
one molecule and there is a need to discuss methods to measure or estimate their HSP. Relevant aspects of the
HSP methodology are presented and then applied to data from the literature to estimate the HSP for several
ionic liquids. A well-chosen ionic liquid can be a reaction medium for polymerization or dissolve a polymer,
and later be recycled. There is a lot of interest in the 1,3-dialkylimidazolium-based ionic liquids as solvents in
the polymer industry. In addition to pure scientific interest, there are also uses or potential uses for ionic liquids
as plasticizers, ionic polymer liquids, and in separation processes. Ionic liquids, usually amine salts with
organic acid moieties, have been used in the coatings industry for many years. It is thought instructive to see
what these uses have been and what they may tell about the current usage of these special liquids in polymers.
A major application in coatings is to have “ionic” moieties as local regions covalently bonded on polymers,
such as water-reducible acrylics, that are not in themselves soluble in water. The acrylic acid additions are
generally neutralized with amines. This improves the stability of the polymer micelles in water since the
organic salts as such are not soluble in the polymer micelle, and necessarily reside at the water/polymer
interface in a stabilizing function, enhanced by the electrostatic repulsion in the water with its high dielectric
constant. In contrast to this, great care is taken to remove all water in studies of the solvent properties of ionic
liquids for polymers.
In other cases amine salts provide good pigment dispersion stability since the energy relations of the local
“ionic” regions are similar to those of especially inorganic pigment surfaces. Lecithin, for example, is a
traditional dispersing agent of this kind that is used in solvent based systems with the “ionic” portion being
attached to a pigment surface with nothing in the system to being able to dissolve this essentially permanent
anchor away from its desired location. Anionic (or cationic) surfactants are also used for dispersing pigments,
often in connection with traditional nonionic surfactants. As is well known, these have hydrocarbon tails and
water soluble heads in the same molecule. Sodium salts of organic acids are frequently used in such
applications. One can therefore expect relatively high HSP for the heads of such salts, matching to some extent
the HSP of water or pigment surfaces. This has been confirmed earlier in terms of what are now called HSP (1).
Another study of some interest involved a base phthalocyanine pigment that was surface treated in different
ways to give quality differences labeled as +++, ++ and down to ---. These characterizations were found from
measurement of electrophoretic mobility in a given solvent. Those pigments with a strong “+” characterization
suspended primarily in solvents with high δH while those with strong “-“ character suspended primarily in
solvents with high δP. These pigment suspension results in a set of standard solvents provided data for HSP
assignment of these same surface treatments. Those pigments with little or no character of the “+” or “-“ scale
did not readily suspend in any solvents, making a characterization with HSP difficult.
Consideration has been given in the above to the “ionic” character of local regions of larger molecules or at
surfaces. The focus of current interest is solubility where it is the HSP of the whole “ionic” molecule that has
interest. The amine salts discussed above have never been thought of or used as solvents, although they may
have acted in a swelling capacity at times without this being recognized. Table 1 shows experimental results
from page 337 of the second edition of the Hansen handbook for reference. A suspicion developed over the
years but never proven is that larger metal ions increase the δD parameter more than smaller ones, without any
significant effect on the δP and δH. Among the organic liquids the larger moieties, such as Cl, Br, and aromatic
rings, for example, do give higher δD.

Salt D P H R
Mg(NO3)2.6H2O 19.5 22.1 21.9 13.2

156
Formic 17.2 21.5 22.5 -
acid/DMEOH
Acetic 16.8 19.8 19.8 -
acid/DMEOH

Table 22-1 HSP for given salts. All units are MPa½. DMEOH is dimethyl ethanolamine.

How can HSP be assigned to ionic liquids?


Solubility relations – what do they dissolve or do not dissolve?
A list of what a solvent dissolves or does not dissolve is important to assign its HSP. The HSP of the solvent
should be such that they are located within the HSP sphere for solubility of those solutes they dissolve, and
outside of such HSP spheres for those solutes that they do not dissolve. Interpretation of data reporting what
they dissolve is (usually but not always) straightforward, whereas the meaning of data for what they do not
dissolve is often complicated by the effect of their molar volumes. This non-solubility data is also significant
since this helps to eliminate large regions of HSP space from consideration in assigning HSP when using this
methodology. When it is reported that an ionic liquid dissolves 3% cellulose of a given degree of
polymerization at 70°C, insight is given into its HSP, but not necessarily to precise values. All three of the HSP
must be very high compared with other liquids. The HSP of cellulose are not determined with accuracy due to
the lack of data for its solubility in organic solvents. In general ionic liquid solubility data does not include
results for polymers for which the HSP polymer sphere characterization is available. This prevents the
satisfactory use of many of the available solubility data. There is also considerable variation in the HSP data for
acrylic polymers, and solubility also depends to a certain extent on the molecular weight of the polymer, with
more solvents for low molecular weights. Larger V may prevent solution that would otherwise be inferred by
the HSP alone, so care is also required here. Lower molecular volume ionic liquids are generally preferred in
practice, all else (HSP included) being equal. The HSP of given ionic liquids can either be too high or too low
for a given use, but the lower molecular volume species can be expected to produce lower viscosity room
temperature liquids, which seems to be the goal for many purposes.

Comparison with solubility/miscibility profiles of reference liquids


The approach used by Hansen to assign HSP to amine salts of organic acids in (1) was to determine their
miscibility is a series of well chosen solvents, just like in a traditional determination of the HSP of a solute. The
miscibility profiles of the acids and bases, as well as given solvents with much higher HSP were also
determined. When the miscibility profile of a given salt was found to be the same as that of a given reference
solvent, the salt was assigned the HSP of that solvent. This procedure has the potential for some modest error
due to possible errors in the HSP of the reference solvent as well as the fact that other HSP values, close to
those of the reference solvent, could also give the same miscibility profile. The boundary solvents found most
useful to show differences in behavior, and therefore to allow such characterizations, were 2-ethyl butanol, n-
amyl alcohol, 2-nitropropane, nitroethane, aromatic solvents, aliphatic solvents, n-butyl acetate, and ethyl
acetate. Other solvents were also tested but were not found useful for this purpose, since it is differences in
behavior that are interesting in these procedures. The amine/organic acid salts with acetic acid, for example,
tend to have HSP in the range of [17.1, 22.5, 23.3] MPa½ giving a δt equal to 36.6 MPa½. These HSP are used in
the following to derive group contributions for amine/organic acid salts in connection with a molar volume, V,
for the n-butyl amine/acetic acid salt approximately equal to 139 cc/mole. The salt of dimethyl ethanol amine
with p-toluene sulfonic acid was assigned HSP equal to [17.2, 21.5, 22.5] MPa½ giving a δt equal to 35.6 MPa½
by the same method. V here is estimated as 216 cc/mole. These estimates of molar volume include measured
shrinkage of about 10% compared with the sum of the molar volumes of the reactants. These group
contributions are useful for estimates of the HSP of related salts. In the case of the salt of methanesulfonic
acid/dimethyl ethanolamine a solid was formed that was not soluble in any of the above test solvents indicating
still higher values, although the molar volume of this salt is also lower. This would lead to higher HSP for the
same cohesive energy of the salt moiety.

157
HSP relative to the solid/liquid boundary at room temperature
Another comparison can possibly be made by recognizing that there is a boundary between solids and liquids in
HSP space. This boundary is not sharp, but can provide information about the approximate values expected for
a solid with a melting point at room temperature, for example. Examples of interest for comparison here are
given in the following Table 2.

Solvent MPt °C D P H MVol


Dimethyl 18.5 18.4 16.4 10.2 71.3
Sulfoxide
Sulfolane 27.6 17.8 17.4 8.7 95.5
Dimethyl 109 19.0 19.4 12.3 75.0
sulfone

Table 22-2 Comparison of melting point and HSP for reference solvents.

HSP for amine/COOH salts


The HSP given above for these salts on average can be used to derive group contributions for use with other
salts. These can be found from the equation where the units of the solubility parameters are MPa½ and the V are
in cc/mole, as the cohesion energy attributable to the given group:
GCi = (δi)2V
For an average of organic amine/acid salts (using n-butyl amine/acetic acid as a model):
GCD = (17.1)2(139) = 40,645 joules
GCP = (22.5)2(139) = 70,368 joules
GCH = (23.3)2(139) = 75,462 joules
The sum of these contributions is 186,475.
These GC can then be used to estimate HSP for amine salts of given V, or perhaps for segments of
molecules where molecular breaking gives a V that can be used to find the local HSP, since this may
also be of interest. These GC are the cohesive energy due to the respective types of interaction,
dispersion (D), dipolar (P) and hydrogen bonding (H). The HSP methodology is quantitative to the
extent data permit assigning accurate values.

HSP for dimethyl ethanolamine/p-toluenesulfonic acid salt


The same general equation given above can also be used for this salt. Corrections for the alcohol group are
included as 2900 J for the polar contribution and 19,400 J for the hydrogen bonding contribution.
GCD = (17.2)2(216) = 63,901 J
GCP = (21.5)2(216) - 2,900 = 96,946 J
GCH = (22.5)2(216) - 19,400 = 89,950 J
Even though the HSP are similar to those above, these group contribution values are higher because the V of
the salt is larger. The total cohesive energy found by this method is 250,797 joules.

HSP for other ionic liquids


Methods of characterization

158
Assigning reasonably accurate HSP to other ionic liquids is difficult since no complete systematic studies of
their behavior have been performed that would have allowed this. Information in references (2-13) has been
used to try to estimate HSP for representative ionic liquids for which the very limited data allow this. As stated
above, these estimates are probably more indicative than precise. Strictly theoretical studies can always be
questioned, and the experimental studies are not sufficiently complete. A problem with experimental
determination of the HSP of liquids in general is that they are usually miscible in most of the test solvents, and
this does not provide sufficient differentiation for an accurate assignment, if any assignment at all. Calculations
based on group contributions or other methods of estimation available with the HSPiP software/eBook are
likewise a possible approach, but nothing really replaces experiments, if they are possible-

Inverse Gas Chromatography (IGC)


A most promising methodology to avoid the traditional problems of assigning HSP to liquids is inverse gas
chromatography (IGC), assuming a sufficiently broad selection of test solvents is used to generate the data. The
liquids are borne on substrates as thin films. Most IGC studies do not include liquids with high δP, in the faulty
belief that alcohols are sufficient for this purpose.

The intrinsic viscosity method


The intrinsic viscosity method used in (6) appears to be a simple and reliable way to determine the HSP of
liquids that are not readily characterized by traditional methods. Higher intrinsic viscosity is associated with a
closer HSP match, and viscous test liquids can also be included. The method involves extrapolating the intrinsic
viscosity (viscosity of solution/viscosity of solvent minus 1) of very dilute solutions to zero concentration.
Capillary tube viscosimeters are usually used. A suitable range of solvents is also required for this method.
Suggestions for added test solvents of this kind are easily found with the HSPiP software using the HSP
estimates for some ionic liquids that follow

Reaction rate method


A new method to assign HSP is to study the reaction rate in a large number of liquids for the [4+2]-
cycloaddition of singlet oxygen with 1,4-dimethylnaphthalene and derivatives (10,14). This method is very
promising, giving results that appear to be correct because of agreement with other methods of characterization.
The reaction rates reported in (14) could be used with the HSPiP software to find given HSP characterizing
what is presumed to relate to the solubility of an intermediate in the reaction as discussed below in detail. These
HSP are also in agreement with the HSP of commonly used reaction solvents, being only very slightly higher.
The solvents used in (14) appear to be adequate but it might be wise to supplement them with others in the
region of interest.

Nomenclature
There is a special nomenclature that is used to describe the ionic liquids. This is briefly explained for those
ionic liquids discussed in this chapter in the following. The cation is usually put in brackets with lower case
letters. The anion is not always put into brackets, and generally follows usage any chemist will recognize.

im 1,3-dialkylimidazolium
m methyl
e ethyl
b butyl
h hexyl
o, oc octyl
dd dodecyl
[bmim] stands for 1,3-butyl methyl imidazolium, for example.

159
The ionic liquids for which HSP are specifically estimated are:
[bmim]Cl, [bmim]PF6, [omim]PF6, and [bmim]BF4.

Review of the content of the references


Winterton (2) provides an extensive review of ionic liquids with data of the kind that are useful to assign HSP
to given ionic liquids if there is enough diversity in the results. There is also a brief discussion of solubility
parameters and an indication that fresh insights might be found by application of solubility parameter methods.
This report is an attempt to do just that. There is also mention of work that estimates the Hildebrand parameter,
δt, from intrinsic viscosity measurements in different liquids and reaction rate constants. When analyzing data
for test liquids one should recognize that assigning HSP outside the range of an average of these is not possible
unless some method is used to extrapolate into this region, such as is done with the HSPiP software.
Kubisa (3) mentions a long list of solutes where ionic liquids have been shown to be good solvents including
inorganic and metalorganic compounds, enzymes, and polymers such as methacrylates (in particular). The main
kinds of ionic liquids discussed are [bmim]PF6, probably because of its low melting (glass transition) point of -
61°C and [bmim]BF4. There is little help in assigning HSP to these, although the qualitative solubility of a
series of methacrylate polymers in [bmim]PF6 is given. This allows a potential check on HSP estimates for this
ionic liquid.
Kim et.al. (4) describe studies of the solubility and reactions of polyepichlorohydrin in particular, but provides
solubility data for seven polymers in three ionic liquids [bmim]Cl, [bmim]BF4, and [bmim]PF6. This is the kind
of data that is desired, but the present data are of limited value since they are taken at 70°C, and there is only
one polymer for which HSP have been determined and one (cellulose) for which estimates of the HSP can be
made with considerable uncertainty. The number of useful systems that show differences in behavior, and their
HSP relations with these, are the important data that determine the accuracy of HSP assignments. These
solubility data must also be considered with caution since they are based on the stability of a solution in the
ionic liquids. This solution was made by including a good, volatile solvent that was subsequently evaporated to
form the solutions that were then observed for a long time (weeks).
Vygodskii et.al. (5) report numerous polyamide and polyimide polymerization reactions with different 1,3-
dialkylimidazolium-based ionic liquids. These can provide different molecular weight polymers, with those
with highest molecular weight being those where the dialkyl groups are C4, C5, and C6 aliphatic chains with Br
as counterion. The solvents that can otherwise be used in the reactions with greater or lesser success are N,N-
dimethylformamide (DMF), N,N-dimethylacetamide (DMAC), N-methylpyrrolidone (NMP), nitrobenzene
(NB), and m-Cresol (MC). One anticipates from these data that the HSP of these ionic liquids will be within the
HSP region defined by these liquids, or perhaps have slightly higher HSP approaching the solid region, and this
indeed turns out to be the case as seen below. This HSP region is often referred to as the reaction solvent
region, in agreement with some of the uses of the ionic solvents. The molecular solvents may require more
extreme conditions than for the ionic liquids, but they are more or less successful as solvents in the cases
reported. The solvent used for the intrinsic viscosity measurements on thoroughly dried polymers to find an
indication of their molecular weights was concentrated sulphuric acid in all cases but one (where NMP was
used).
Lee and Lee (6) measured intrinsic viscosities for a number of ionic liquids in up to 15 different solvents, but
only reported the intrinsic viscosity data for two liquids and up to 9 solvents in a figure that is not easy to
interpret. A table with the full data would have been extremely helpful to evaluate the quality of the estimate,
and probably even to estimate the HSP using the HSPiP software. The Hildebrand parameters discussed in the
above for organic amine/acid salts are 36.6 MPa½ for n-butyl amine/acetic acid and 35.6 MPa½ for dimethyl
ethanol amine/p-toluene sulfonic acid. The test liquids that best matched the miscibility profile of the n-
butylamine/acetic acid salt were ethanolamine (δt=31.7) and triethanolamine (δt=36.6), with the latter being
accepted as correct. These values are somewhat higher than those reported by Lee and Lee for 8 of the more
frequently mentioned ionic liquids having cations of the 1-alkyl-3-methylimidazolium type. These are in the
range of 25.4 to 29.8 MPa½. It is stated that the intrinsic viscosity method is reliable since, in control
experiments with a set of well-selected solvents, it accurately found δt for dipropylene glycol (δt=26.4) and 1,3-
butanediol (δt=29.0). These have viscosities similar to some ionic liquids, and δt very close to those of the ionic

160
liquids. This speaks well for the intrinsic viscosity method in itself, but care must be taken since solutes may
have δt and HSP that are out of the range attainable by the test solvents. This article also gives the molar
volumes and densities for the ionic liquids studied. This is a very useful help for those who may continue this
study. The δt reported in this article for [bmim]PF6 (29.8 MPa½) and [omim]PF6 (27.8 MPa½) have been used as
a reference values in the following.
Weingärtner (7) does not report any data useful for present purposes, although the review covers other aspects
of ionic liquids very well.
Derecskei and Derecskei-Kovacs (8) give molecular modeling calculations of the densities and cohesive energy
density for [emim][bmim][C6mim] all with anions [CF3C02], [(CN)2N] and [N(SO2CF3)2] abbreviated [Tf2N].
These agree reasonably well with “experimental” values. This data is worthy of comparison with other
approaches for finding the cohesive energy density and total solubility parameter. There is also a split of the
total solubility parameter into “Coulombic” and “H-bonding” contributions that is difficult to use in the present
context.
Xu, et.al. (9) provide glass transition temperatures for a number of ionic liquids in a plot showing a minimum at
about 250 cc/mole. These temperatures are from about 0°C down to about -100°, making direct relations to
room temperature behavior rather remote.
Swiderski et.al. (10) determined cohesive energies and Hildebrand solubility parameters for a number of
common [bmim] ionic liquids that appear to agree well with other estimates in the literature. The anions were
[BF4], [SbF6], [PF6], [(CF3SO2)2N], [CF3SO3]. [bm2im][(CF3SO2)2N] was also studied. The methodology was a
previously determined procedure and equation developed using 28 solvents that related molecular oxygen
reaction rates with 1,4-dimethylnaphthalene to the Hildebrand parameters (14). The δt found for [bmim][BF4]
and [bmim][PF6] were 31.6 MPa½ and 30.2 MPa½ in good agreement with the values found by Lee and Lee (6),
and have been used as reference values below. This method appears to be quite reliable. Correlation with HSP
rather than the lesser meaningful δt would have been preferred, at least in the present context. 24 of the solvents
used in (14) were correlated with HSP using the HSPiP software with the following results:

Rate x10-4 M-1s-1 δD δP δH R0 Fit G/T


k>86 21.47 15.16 17.44 13.9 0.964 4/24
k>65 19.06 16.46 15.07 11.2 0.936 5/24
k>40 20.41 15.99 14.63 12.7 0.976 6/24
k>26 20.73 16.17 14.02 13.4 0.999 8/24
k>7 21.02 15.85 15.70 17.5 1.000 18/24
Average 20.5 15.9 15.4 - - -

Table 22-3 HSP correlations of the rate constant data given in (14). G is the number of “good” solvents out of a total number,
T.
With this positive result, it would seem logical to extend this method with solvents having higher HSP, perhaps
at the expense of some having lower HSP to lessen the work. The results found by this method have been
incorporated into the estimates for the HSP of the ionic liquids analyzed below.
The average of the HSP correlations given in Table 3 is in the higher HSP region of the HSP sphere defined by
typical reaction solvents such as DMSO, DMF, DMAC, and NMP. This can be seen from the figure on page
278 in the handbook. This is a little closer to the solid region than any of these molecular liquids. This clearly
points to a generality involving the enhanced solubility of an intermediate species. The HSP for the ionic
liquids discussed in detail below somewhat higher than this likewise just above the HSP of the typical reaction
solvents. This gives added support to this conclusion.
Mutelet et.al. (11) used inverse gas chromatography (IGC) in an attempt to evaluate both the Hildebrand and
Hansen solubility parameters. Data and results are reported for [mocim][Cl], [emim][(CF3SO2)2], and
[bmim][PF6]. The authors report differences in values of the Hildebrand parameters found in the literature for
[bmim][PF6] depending on method of determination that vary from 22.02 calculated in this work, a Monte
161
Carlo simulation giving 26.70, and other experimental data giving 18.60, all in J½cm-3/2. This study is not useful
to assign HSP since the solvents used did not sufficiently cover the required HSP space (the same old story).
There is a marked lack of solvents with high δP, since the authors erroneously assumed alcohols covered the
highly “polar” liquids.
What can be gained in this context from the 1969 article by Hansen (12) is that HSP could be assigned to
proteins (zein from corn) and inorganic salts, and that mixtures of inorganic salts in given liquids could dissolve
certain polymers such as Nylon 66 (15% calcium chloride in methanol) and zein (10% calcium nitrate in
gamma butyrolactone, 20% magnesium nitrate in 1-butanol). The “ionic” moiety need not be chemically
attached to the organic solvents. The nitrates and chlorides are those types most readily soluble in organic
solvents. The nitrate and chloride entities have considerable HSP leverage in “salts” to reduce their HSP. This
statement is based on the fact that the HSP of the nitro compounds and chlorinated solvents are well removed
from regions where all three HSP are high. These anions reduce HSP in an inorganic salt to a greater extent
than other common anions thus allowing organic solvents to dissolve them.
Polymers used in this study to help assign HSP to ionic liquids
Sheets were made to note potentially relevant information for each ionic liquid for which there was hope for a
HSP characterization. It became evident that such would be impossible for most of the ionic liquids at this time
(September 2009), even those discussed in many sources. This is primarily because the solvents used to
generate the data or methodology used did not sufficiently cover the required HSP space. The data related to
polymerization are also of such a nature that it can be difficult to assess just how to use it. Miscibility relations
are reported for some monomers and many polymers. Most of the polymers discussed in this literature have not
been assigned HSP. The miscibility of the resulting polymers in these polymerization studies depends on a
number of factors such as the temperature prevailing at the time of judgment.
Polymer δD δP δH R0
PMMA #66 18.64 10.52 7.51 8.59
PAN #183 21.7 14.1 9.1 10.9
PS (p.102-G, 21.3 5.8 4.3 12.7
Handbook)
Dextran C 24.3 19.9 22.5 17.4
(Cellulosea)
PEG-3350 17.3 9.5 3.3 10.0
PEO #209 21.5 10.9 13.1 15.9
b
(heated solns.)

Table 22-4 HSP relations for polymers of significance in assigning HSP to the ionic liquids. Units for δ are MPa½.
a
The HSP for soluble cellulose are thought to be (much) higher in all three parameters, with a correspondingly
larger R0 that would not encompass what dissolves Dextran C. Dextran is an amorphous polymer resembling
amorphous cellulose in structure. This estimate is used for guidance only with something reported as a solvent
for cellulose expected to have a relatively low RED, and non-solvents for cellulose may still have RED
somewhat lower than 1.0 based on this correlation.
b
This correlation with heated solutions is based on very old so-called Solvent Range data and is highly
questionable. It is included here for the sake of completeness, and because some data used for comparison also
involve heated solutions.
Among the numerous HSP correlations for PMMA its solubility relations are thought best represented by
correlation #66 in the Hansen handbook. The PAN correlation is thought to be quite good. There are numerous
correlations for PS in the handbook, with this one being based on the authors own experience with a very high
molecular weight type suitable for injection molding. Solvents that only swell this PS may well dissolve lower
molecular weight polystyrenes.
The solubility of cellulose is so important commercially as well as to assigning HSP that an extended
discussion is given in the following. Hansen has made estimates of the HSP for cellulose using noncrystalline
162
Dextran C as a model for soluble cellulose (13). Dextran A with a molecular weight of 200,000 to 275,000
dissolved in the same solvents as Dextran C with a molecular weight of 60,000 to 90,000, but with obvious
differences in viscosity (12). The Dextran C model is also thought useful in the present context. It is based on
solubility in ethylene glycol, glycerol, formamide, ethanolamine, and dimethyl sulfoxide. Later the HSP
assigned to lignin monomers placed them on the boundary of this correlation (13). N-methyl morpholine-N-
oxide [19.0, 16.1, 10.2, V:97.6] also conformed to these HSP as a good solvent (RED:16.7/17.4) for soluble
(amorphous) cellulose. The HSP of cellulose will be higher than the values for Dextran C, but presumably have
the same relative values and with a radius of interaction that is larger than that for Dextran C but not
encompassing the liquids mentioned above.
There are two HSP sets for PEG that have not been published until this time. These are for PEG-3350 solubility
and PEG-200,000 for swelling and solution. The HSP for the higher molecular weight species were not used
here but are given for reference as [17.0, 11.0, 5.2, R:8.2]. To avoid future misunderstandings it should be noted
that chloroform and trichloroethylene were the only solvents dissolving the higher molecular weight PEO, and
that the lower molecular weight PEO also dissolved in three solvents (water, formamide, and methanol, the
only alcohol to dissolve) with RED greater than 1.0 and molar volumes less than 41 cm3/mole. The PEO data
must be considered with care since this polymer forms special structures that may significantly affect solubility
results.
HSP assignments to ionic liquids
It has only been possible to assign HSP to the four ionic liquids listed in Table 5 in spite of all the data in the
literature. The methods discussed above will hopefully be used in the future to improve this situation. The data
used for assigning the respective HSP are discussed in the following. Fortunately these four ionic liquids are
representative and their HSP can be used with care together with other data, especially the Hildebrand
parameters given in (6,10), to estimate HSP for others.
Ionic liquid δD δP δH δt V, cc/mole
[bmim]Cl 19.1 20.7 20.7 35.0 175.0
[bmim]PF6 21.0 17.2 10.9 29.3 207.6
[omim]PF6 20.0 16.5 10.0 27.8 276.0
[bmim]BF4 23.0 19.0 10.0 31.5 201.4

Table 22-5 Estimated HSP for given ionic liquids, MPa½.


[bmim]Cl
[bmim]Cl is interesting because at 70°C it dissolves 3% cellulose having a degree of polymerization of 223
(molecular weight about 36,000) as described above (4), although the solutions were observed for weeks after
initially being helped with a volatile solvent. Other related ionic liquids produce swollen solutions with this
cellulose. Other polymers [bmim]Cl dissolve include poly(vinyl phenol), poly(vinyl alcohol),
poly(epichlorohydrin), and smaller amounts of poly(chloromethyl styrene). It does not dissolve poly(ethylene
glycol) (PEG) or poly(methyl hydrosilane). HSP are only available as estimates for one of these, PEG.
Using the interactions with cellulose modeled as described above and PEO gives what might be considered as a
minimum HSP estimate for [bmim]Cl as [19.1,13,13, δt:26.5]. The RED to Dextran C is 15.7/17.4 and to PEO
is 10.9/10.0. If one uses 75,000 J as a group contribution, derived from the δP and δH from the n-butyl
amine/acetic acid salt data, the estimate is [19.1, 20.7, 20.7, δt:35.0]. The HSP could easily be higher than these
the currently accepted values. There are no data we are aware of at the present time to eliminate this
uncertainty. Based on the estimated HSP for the following ionic liquids, there is no apparent need to require
that δP be equal to δH. RED values in parenthesis for the polymers of significance here are Dextran C (0.61) and
PEO #209 (0.82) and PEG-3350 (2.10). PMMS (1.94), PS (1.78), and PAN (1.32) are predicted as being not
soluble. The HSP assignment fits the data available, but other HSP can also do this, so adjustments may be
required in the future.
[bmim]PF6

163
[bmim]PF6 has been assigned a Hildebrand parameter equal to 30.2 MPa½ by the oxygen reaction rate method
(10) and 29.8 MPa½ by the intrinsic viscosity method (6). The δt estimated here equal to 29.3 MPa½ is
considered close enough to these. The HSP estimates are [21.0, 17.2, 10.9] MPa½. [bmim]PF6 is plasticizer for
PMMA (2). PAN is soluble in the presence of unknown amounts of sulfolane, which also is a solvent for PAN,
reducing the value of this finding in the present context. Cellulose is stated as being insoluble while PEG
maintained solubility at 33% for a prolonged time after an initial solution was made in a volatile solvent (4).
Low molecular weight PS is also soluble in this ionic liquid (2). There is considerably more information on
polymer solubility in the literature, but the solubility relations of the polymers involved have not been
characterized by suitable HSP studies. The solubility relations (RED) for the polymers listed in Table 4 are
PMMA (1.03), PAN (0.35), PS (1.04), Dextran C (Cellulose) (0.78), PEG-3350 (1.31), and PEO #209 (0.42).
These data support [bmim]PF6 as a potential plasticizer for PMMA, as being soluble in PAN/sulfolane, and
being able to support low molecular weight PS. The RED for the Dextran C correlation is higher than that for
[bmim]Cl but still indicates a potential interaction is not far off, and the PEO/PEG data support any conclusion
desired. The data all appear to be in reasonable agreement with the assigned HSP.
[omim]PF6
The only help in determining the HSP for [omim]PF6 is that low molecular weight PS is soluble and there is a
higher solubility of PMMA than in [bmim]PF6. This is reasonable enough for the increase in length of a side
chain from butyl to octyl. The estimated δt is 27.8 MPa½ in complete agreement with the identical value found
from intrinsic viscosity measurements (6). The RED numbers for PMMA (0.82) and PS (0.98) are in agreement
with all of these data as well.
[bmim]BF4
[bmim]BF4 has been assigned a δt of 31.6 MPa½ based on the oxygen reaction rate method. This is matched by
the δt of 31.5 found from the HSP [23.0, 19.0, 10.0] that also are supported by what few data are available.
PMMA is not soluble (RED=1.47), polyimides and polyamides are compatible in agreement with chemical
resistance correlations in the handbook, and PEG is miscible in the heated sample experiments reported in (4).
The RED numbers here are 0.58 for the correlation with heated solutions of PEG and 1.63 for the PEG-3350
correlation. These HSP could be confirmed by further testing, such as RED for PAN being 0.515, or perhaps
using other polymers with known HSP solubility correlations.
Discussion of the HSP assignments
The HSP assigned to the four ionic liquids in Table 5 appear to be reasonable. They are presumably not precise
but there is general agreement among the available independent methods of assigning Hildebrand parameters
with these HSP. It is also significant that these HSP are slightly higher in all three parameters compared with
the HSP region defined by the usual reaction solvents, DMSO, DMF, DMAC, and NMP. This is clearly seen in
the figure on page 278 of the handbook. One of the major uses of ionic liquids is as a reaction solvent. The HSP
correlation found for reaction rate data in (14) are also close to the HSP of the reaction solvents and ionic
liquids, giving a convergence of HSP for all of these.
These HSP are somewhat dependent on how one evaluates the data, but values near [20.4, 16.0,14.6] are a
reasonable average. There is obviously a region of HSP at high values of δP and δH not too different from the
HSP of the common reaction solvents that is characteristic of the positive solubility effects in this reaction as
well.
The reaction rate constants are for the [4+2]-cycloaddition of singlet oxygen with 1,4-dimethylnaphthalene and
derivatives in different solvents. This finding is considered significant in itself, in addition to the help this
technique can give for future HSP assignments of ionic and other liquids of interest.
There are now three methods that can be used to assign HSP to ionic liquids in addition to the traditional ones
based on what is miscible with what. This offers great promise for the future, but in each case careful selection
of test solvents is mandatory for an accurate HSP assignment. Where HSP spheres are calculated, it is boundary
solvents that have greatest interest, while in the more direct methods, it is suggested that solvents with HSP
closer to those of the sample be used.
Conclusion

164
HSP have tentatively been assigned to [bmim]Cl, [bmim]PF6, [omim]PF6, and [bmim]BF4 based on solubility
relations and proposed Hildebrand parameters that were found in the literature. The solubility relations with
given polymers aid in arriving at the tentative assignments, but useful data are surprisingly few. Three other
methods are suggested as offering great promise for determining the HSP of such solutes. These are IGC,
intrinsic viscosity (6), and oxygen reaction rate measurement (10,14). In each of these a series of test solvents is
used to determine interactions with the liquid sample. Liquids are difficult to characterize with HSP by more
traditional methods. The intrinsic viscosity is highest in the best solvent(s), as is the reaction rate constant in the
oxygen reaction rate method. The HSP of a significant number of test solvents should be close to the
(estimated) HSP of the liquid being studied in all of these methods, whereas in a traditional solubility parameter
study based on solubility, it is the boundary solvents for the HSP sphere that are more important. Boundary
solvents can also become important for these newer methods, however, if the data are entered directly into the
HSPiP sphere optimizing program as done in the above. This more direct analysis is particularly promising,
also for finding additional test solvents that help define the situation more accurately.
It may be that one or more of these methods will also allow greater insight into the region where materials are
solids and HSP is very much higher than for test liquids. HSP characterizations of materials in this region are
very uncertain being based on extrapolations using techniques and equations that have been found very useful
in the liquid region, but lack studies providing convincing proof for their use in the solid region.
NADES
A recent development that produces very interesting, safe, low-cost solvents are the NADES: Natural Deep
Eutectic Solvents, as developed by Young Hae Choi’s team at U. Leiden15. By mixing unlikely materials such
as choline chloride and lactic acid an interesting eutectic is formed. Unfortunately, many such eutectics are
solids at room temperature and useless as solvents. A surprising number are liquids at room temperature but of
such high viscosity that they are of limited practical use. Adding a lot of water creates, not surprisingly, merely
a dilute aqueous solution of no great interest. But it turns out that 1:1:1 (or similar) ratios of components to
water creates very interesting solvents capable of dissolving interesting materials such as: Rutin, Starch
(PropyleneGlycol:CholineChloride:Water, 1:1:1); Taxol, DNA (LacticAcid:Glucose:Water 5:1:3); Quercitin,
Ginkolide B (Xylitol:CholineChloride:Water , 1:2:3). The viscosities of these mixtures are typically 30-100x
larger than water.
Clearly the NADES approach opens up many interesting avenues for solubility work. But there are two
problems. The first is to find eutectic/water mixtures that are sufficiently low viscosity, without destroying their
interesting solubility characteristics. The second is to predict the solubility properties of those mixes. We can’t
help with the first, but by adding the NADES.sof with our estimates of the HSP of many of the relevant
NADES components it’s possible for users to create their own mixes and estimate their HSP. How good these
estimates will prove is too early to tell. Within the NADES file water is included not with the classic δH=42
value but the 1% value which seems to be useful so often in these sorts of applications.
Obviously these are early days for NADES work and HSP. As with everything HSP, we welcome user
feedback on how useful, or otherwise, this approach turns out to be. If/when we get measured values of some of
the key sugars and/or the choline or betaine chlorides we will update the table.
References:
1. Hansen, C.M., Einige Aspekte der Säure/Base-Wechselwirkung (Some Aspects of Acid/Base
Interactions), Farbe und Lack, Vol. 83, No. 7, 595-598 (1977).
2. Niel Winterton, Solubilization of polymers by ionic liquids, J. Mater. Chem., 16, 4281-4293 (2006).
3. Przemyslaw Kubisa, Application of ionic liquids as solvents for polymerization processes, Prog. Polym.
Sci. 29, 3-12 (2004).
4. Byoung Gak Kim, Eun-Ho Sung, Jae-Seung Chung, Seung-Yeop Kwak, and Jong-Chan Lee,
Solubilization and Polymer Analogous Reactions of Polyepichlorohydrin in Ionic Liquids, Journal of
Applied Polymer Science, 114, 132-138 (2009).
5. Yakov S. Vygodskii, Elena I. Lozinskaya, and Alexandre S. Shaplov, Macromol. Rapid Commun. 23,
No. 12, 676-680 (2002).
6. Sang Hyun Lee and Sun Bok Lee, The Hildebrand solubility parameters, cohesive energy densities and
internal energies of 1-alkyl-3-methylimidazolium-based room temperature ionic liquids, Chem.
Commun., 3469-3471 (2005).
165
7. Hermann Weingärtner, Understanding Ionic Liquids at the Molecular Level: Facts, Problems, and
Controversies, Angew. Chem. Int. Ed. 47, 654-670 (2008)
8. Bela Derecskei and Agnes Derecskei-Kovacs, Molecular modeling simulations to predict density and
solubility parameters of ionic liquids, Molecular Simulations, 34, Nos. 10-15, September-December,
1167-1175 (2008).
9. Wu Xu, Emanuel I. Cooper, and C. Austen Angell, Ionic Liquids: Ion Mobilities, Glass Temperatures,
and Fragilities, J. Phys. Chem. B, 107, 6170-6178 (2003).
10. Konrad Swiderski, Andrew McLean, Charles M. Gordon and D. Huw Vaughan, Estimates of internal
energies of vaporization of some room temperature ionic liquids, Chem. Commun., 2178-2179 (2004).
11. Fabrice Mutelet, Vincent Butet and Jean-Noël Jaubert, Application of Inverse Gas Chromatography and
Regular Solution Theory for Characterization of Ionic Liquids, Ind. Eng. Chem. Res. 44, 4120-4127
(2005).
12. Hansen, C.M. The Universality of the Solubility Parameter, Industrial and Engineering Chemistry
Product Research and Development, 8, No. 1, March, 2-11 (1969).
13. Hansen, C.M. and A. Björkman, The Ultrastructure of Wood from a Solubility Parameter Point of View,
Holzforschung, 52, 335-344 (1998).
14. Jean-Marie Aubrey, Bernadette Mandard-Cazin, Michael Rougee, and René V. Benasson, Kinetic
Studies of Singlet Oxygen [4 + 2]-Cycloadditions with Cyclic 1,3-Dienes in 28 Solvents, J. Am. Chem.
Soc. 117, 9159-9164 (1995).
15. Yuntao Dai, Jaap van Spronsen, Geert-Jan Witkamp, Robert Verpoorte, Young Hae Choi, Natural deep
eutectic solvents as new potential media for green technology, Analytica Chimica Acta 766 (2013) 61–
68

166
Chapter 23 Cleaning by numbers (HSP for Surfactants)
This chapter is co-authored by Dr Richard Valpey III of SC Johnson. We are grateful to Richard for his expert
technical input and to his company, SC Johnson, for giving us permission to use their surfactant HSP data.
Responsibility for errors in this chapter remains with Abbott as per his guarantee.
The good thing about surfactants is that there are so many to choose from. The bad thing about them is the same
– that there are so many to choose from. Many would-be users of surfactants despair at having to sort through
so many different surfactants in search for the perfect one.
The search is helped somewhat by well-known numbers attached to surfactants such as HLB (Hydrophilic-
Lipophilic Balance), Aniline point, KB (Kauri Butanol) value. But these provide surprisingly little scientific
insight into a specific surfactant.
HLB was originally determined by a time consuming determination of emulsion stability. Griffin measured the
stability of two types of emulsions (oil-in-water and oil-out) formed by a series of oils in the presence of
surfactants. He then fit the results to a systematic ranking and called it the hydrophile lipophyle balance (HLB).
It is time consuming because approximately 75 emulsions were made for each HLB determination.
Becher suggested that HLB relates to free energy according to the following equation
C 2 Gm , l C 2 Gm , h
HLB = C 1 + +
RT RT
Where: Gm, l and Gm, h are the free energy of micellisation associated with the lipophilic and hydrophilic
moieties. C1 and C2 are scaling factors.
In its original form, HLB was a relative effectivity index, ranging from 0 to 40. Griffin acknowledged its
limitation to nonionic surfactants.
Davies proposed eliminating this limitation by computing HLB based on the structure of the surfactant by
assigning group numbers (GN) to various moieties according to the following equation:
HLB = 7 +  GN
The Davies method, which finds use in emulsion technology, produces negative HLB numbers, particularly
when the lipophilic contribution is sufficiently large.
Despite difficulties in handling negative numbers and poor correlation to ionic surfactants, HLB is the most
widely used tool for selecting surfactants.
In 1978, Little suggested a tool that overcomes these two difficulties. He proposed the following relationship
between the Hildebrand Solubility Parameter δ and HLB. This method which was originally tested with
nonionic and anionic surfactants, correlates poorly with cationic surfactants.
54( − 8.2)
HLB =
( − 6.0)
Given that HLB themselves frequently offer little insight to specific problems, and given that we know the
limitations of the Hildebrand parameter, this correlation is not of much help.
Surely it makes sense to provide users with chemical insights into the functionality of the surfactants via HSP.
There has been remarkably little work on this approach but by combining the earlier work of Beerbower with
the recent work of Valpey we can make some progress.
The key fact is that we can think of surfactants as having 3 sets of HSP. The first is the hydrophobic portion.
The second is the hydrophilic portion. And the third is the (weighted) average of the two just as with any
mixture of solvents. The last is particularly important even if you don’t use it directly. Because it is a weighted
average, it provides some of the insights from an HLB. So important is this weighting that we’ve added it to the
software so it’s easy to do.
Here is a list of surfactant HSP partitioned in the above manner:

167
Surfactant δD δP δH
SLES hydrophobe 16.0 0 0
SLES hydrophile 20.0 20.0 20.0
SLES Average 16.7 8.1 8.1
APG hydrophobe 15.5 0 0
APG hydropile 23.4 18.4 20.8
APG average 18.8 7.6 8.6
Span 80 hydrophobe 16.1 3.8 3.7
Span 80 hydrophile 18.1 12.0 34.0
Span 80 average 16.1 6.1 13.2
Alkyl sulfosuccinate hydrophobe 16.0 0 0
Alkyl sulfosuccinate hydrophile 20.0 17.0 9.7
Alkyl sulfosuccinate average 19.2 3.4 1.9
ST-15 hydrophobe 16.4 0.0 0.0
ST-15 hydrophile 16.2 7.8 10.4
ST-15 average 16.3 3.6 4.8

Table 23-1 Estimated values for the three characteristic sets of HSP for typical surfactants. There is not yet experimental data
to verify these estimates.
Let’s look more closely at Span 80 – Sorbitan oleate.
The “oleate” part can be imagined as methyl oleate with HSP of [16.2, 4.9, 0.5], or simply as [16, 0, 0]
representing the pure hydrocarbon chain. We’ve chosen a group contribution method that gives the values
shown below. The sorbitan can be calculated as [17.5, 10.3, 20.8] with a molar volume ~ 150.

Figure 23-1
The weighted average (calculated by summing the individual energies then dividing by the combined molar
volume) is therefore biased towards the hydrophobic end – giving [16.9, 7.1, 11.9]. If Span 20 were considered,
sorbitan monolaurate, then the individual HSP don’t change much, but the reduced molar volume of the laurate
moiety (~215) shifts the average to [17.1, 8.4, 13.6].

Figure 23-2
168
No doubt you’re starting to see the problems with this approach. There are a few assumptions that have to be
made. Where do you draw the line between hydrophobe and hydrophile? How do you estimate the HSP and
molar volumes for the chunks into which you’ve divided the molecule?
We’ve attempted to answer some of those questions for you by providing our best estimates of many of the
common groups used in surfactants. By selecting one of the hydrophobes and one of the hydrophiles (it’s up to
you whether that combination can actually exist) you at least have a reasonable starting point for your own
explorations. But our values are only for guidance, you should use your own judgement for your particular
surfactants.
The Y-MB calculation of the surfactant HSP is also shown in the examples above. Wherever there are
meaningful SMILES values for the head and tail and also meaningful Y-MB fragments available, the head and
tails SMILES are stuck together into a single SMILES and sent to Y-MB.
At this stage in surfactant research we have no good data to give you. Instead we’ll go out on a limb and make
some predictions. Let’s take 5 standard “soils” (see the Handbook for an explanation of these 5. The HSP
numbers in the Handbook differ from those shown here)
No. Soil δD δP δH
1 ASTM Fuel “A” 14.3 0 0
2 Butyl Stearate 14.5 3.7 3.5
3 Castor Oil 15 6 8
4 Ethyl cinnamate 18.4 8.2 4.1
5 Tricresyl phosphate 19 12.3 4.5
Table 23-2 HSP of typical soils
Now let’s calculate the distance between each of these soils and 5 surfactants
Surfactant 1 2 3 4 5
SLES 14.5 10.6 8.4 2.8 2.6
APG 14.6 10.7 7.8 4.6 6.2
Span 80 15.0 10.5 5.6 10.4 12.2
Alkyl 10.5 9.5 10.7 5.5 9.3
sulfosuccinate
ST-15 7.2 3.8 4.8 6.3 10.2
Table 23-3 Calculated distances of some surfactants to representative oils
If you believe this approach to surfactants, then from the table you can instantly work out that each stain has an
optimal surfactant. Stain 1 would best be removed by ST-15, though the distance is so large that it might not
work at all. ST-15 will be also be best for stains 2 and 3, with more success, and SLES would be best for stains
4 and 5.
The “if” at the start of the previous paragraph is rather important. Classic thinking about surfactants tends to
assume that the hydrophobic tail does the interaction with the soil and the hydrophilic head does the interaction
with the water so that the tail + soil get swept away. There is an obvious problem with this classic thinking. The
tails of most surfactants are remarkably similar and therefore the cleansing power should be fairly similar as
long as the head is swept away in the water. The very large differences in cleaning power of different
surfactants are therefore not naturally explicable using such simple ideas. The HSP model suggests an
alternative approach to rational removal of soils.
Of course, the classic model includes the formation of micelles as the actual cleaning agents, and the different
chains give different critical micelle concentrations and, therefore, different behaviour in the cleaning
environment. Notions such as Critical Packing Parameter depend strongly on the relative size of head and tail.
The simplistic HSP approach says nothing about this important element of surfactant behaviour. But of course
169
the different chain lengths will also have different HSP and molar volumes, which, in turn, determine their
solubility in water and, even more importantly, their relative solubility in the two phases and thus the delicate
balance which appears as the PIT – Phase Inversion Temperature. Perhaps the most interesting aspects of HSP
and surfactants will be their use in non-aqueous dispersions, where matching HSP of the surfactant ends to the
HSP of the respective phases would seem to be a helpful exercise.
Nevertheless, we’re happy to predict that an intelligent use of HSP will prove highly insightful for many
cleaning applications. We have some evidence from our own commercial activities that these predictions are
indeed helpful. But despite the fact that this approach was first suggested by Beerbower many years ago, only
recently has it been looked at with fresh eyes and more powerful ways of predicting HSP values. The Inverse
Gas Chromatography (IGC) technique discussed in the Chromatography chapter gives hope that the HSP of
numerous surfactants will be measured experimentally, which will be an important addition to our knowledge
base. We are confident that we will be hearing more about coming clean with HSP.
Update for the 4th Edition
The big advance in surfactant theory has come from the HLD theory of Salager as extended by Acosta to form
HLD-NAC. This simple, numerical approach offers great power and puts to shame naïve ideas of HLB and
invalidates many of the formulation ideas behind CPP (Critical Packing Parameters). This eBook is not the
place to discuss HLD-NAC. The free software and apps provided at
www.stevenabbott.co.uk/PracticalSurfactants gives formulators a quick way to learn how to apply the theory.
But HLD-NAC requires knowledge of the “oil” with which one is formulating. In particular, it requires the
Equivalent Alkane Carbon Number (EACN) for the oil. HSPiP can now estimate the EACN from a SMILES
input. The estimation is only as good as the dataset used to model it. There is an unfortunate lack of reliable
EACN values across a wide range of molecules. As HLD-NAC becomes more used and EACN values for more
oils are measured the estimation scheme will become steadily more reliable.
In the examples above the EACN for Ethyl Laurate is estimated as 5.7.

Update for the 5th Edition


There is a fundamental flaw in the estimation of HSP of surfactants as surfactants. Because a surfactant acts at a
surface, with one half in one medium and the other half in the other medium, the “mean field” theory behind
HSP (and most solubility ides) no longer applies. Abbott has worked with experts in other solubility
methodologies and so far there is no hint of a robust theory for understanding/modelling surfactants, especially
when intermediate molecules, so-called polar oils, become involved. For example, is octanol a simple alcohol,
an oil or a (poor) surfactant. The answer is all three depending on the context, and when octanol is added to a
water/oil/surfactant system it is not clear what it is doing.
There are further complications when concentrated surfactants start to form mesophases – again these are
effects beyond current solubility theories.
So we currently advise HSPiP users to treat any HSP calculations of surfactants with great caution!

170
Chapter 24 Chromatography – HSP creator and user (Retention Times
and HSP)
Chromatography is about controlled interactions between chemicals in one phase and those in another. It is
therefore not surprising that HSP can play an important role in understanding the outcomes of a
chromatographic process.
One important aspect of chromatography is as a creator of reliable HSP. Given how hard it can be to get HSP
for chemicals and polymers, it’s good to know that those with the right chromatographic kit can determine HSP
relatively easily.
HSP can also be used to explain and, more importantly, predict retention times in some chromatographic
processes. So chromatography can be both a creator and user of HSP.
Creating HSP with chromatography
The archetypal method is Inverse-phase Gas Chromatography (IGC) where the solid phase is made from the
polymer to be investigated and the retention times are measured for a series of solvents with known HSP. The
closer in HSP the solvent is to the solid phase, the more the solvent will tend to linger within the solid phase, so
the retention time will be higher. However, as we will see, IGC also depends on other factors such as vapour
pressure.
The key issue is how to extract HSP from the retention data. The first step is to convert the retention times into
specific retention volumes Vg. These can then be converted into Chi parameters. Finally, the HSP can be found
by doing a linear regression fit to the formula relating Chi to HSP. The papers of Professor Jan-Chan Huang’s
group at U. Massachusetts-Lowell are good examples of this, J-C Huang, K-T Lin, R. D. Deanin, Three-
Dimensional Solubility Parameters of Poly(ε-Caprolactone), Journal of Applied Polymer Science, 100, 2002–
2009, 2006. Their data can be recast into a format that uses the standard HSP distance (in this case, distance-
squared). The famous factor of 4 is automatically included. Careful analysis by Professor Huang’s group found
that by having no factor of 4 the errors were rather large and they included a factor as a fitting variable. Its
optimum values varied from 2.6 to 4. The group were very much aware of the theoretical value of 4 but
postulate that at the higher temperatures of these experiments the lower values may be justified.
Here is the fit for their data on polycaprolactone, using the revised version of their approach
Equ. 24-1 Chi = C1 + MVol *C2 * (4 * (δDp-δDs)2 + (δPp-δPs)2 + (δHp-δHs)2)/(4RT)
Where constants C1, C2 and the Polymer HSP (δDp, δPp, δHp) were the fitting parameters.

171
Figure 24-1 IGC fit of Polycaprolactone Huang Probes data
From this fit, the values [19.2, 4.4, 5.3] are close to Huang’s optimized fit (using 3.2 instead of the factor of 4)
of [18.5, 4.2, 5.6]. The fitting program contains extra information based on Huang’s analysis. An Entropy value
captures the size of the RTη term (using an approximation of 100 for the Molar Volume of the stationary
phase). And a 90% Confidence Interval is estimated for δD, δP and δH using an approximation to the method
described in Huang’s paper.
The program offers a choice of fits to give the user some idea of the robustness of the results. The “C2*RT ~1”
option forces C2 to be close to its theoretical value of 1/RT. Alternatively you can optimize to the minimum
Standard Deviation. You can also explore the robustness of the fit manually be entering your own values into
the δD, δP and δH fields.
Note that the Temperature (°C) option is used so that solvent HSP are calculated at the 70°C of the dataset as
discussed below. When translated back to 25ºC the values become [20, 4.5, 5.7], though because of the low
thermal expansion coefficient of polymers it can be argued that the HSP do not change much with temperature.
Similarly the 70ºC PDMS data set (PDMS.hsd) gives fit around [15.6, 0, 0]

Figure 24-2 IGC fit of PDMS data

Huang’s analysis is of data from M. Tian and P. Munk, Characterization of Polymer-Solvent Interactions and
Their Temperature Dependence Using Inverse Gas Chromatography, J. Chem. Eng. Data 39, 742-755, 1994.
HSD files for all the polymers in this paper are included with the HSPiP as a contribution to the use of IGC for
HSP purposes. Note that the Chi parameters are calculated in the Tian and Munk paper. See below for doing
your own Chi calculations.
If that were all there is to it, then everyone would be using IGC to measure HSP of their polymers. One issue is
that IGC has to take place at a temperature where the polymer is suitably “liquid”. In the example above, the
HSP for the polymer are those at 70ºC. The fitting procedure compares the polymer HSP with the solvent HSP.
But of course, the validity of the calculation depends on having good values for the HSP of the solvents at these
elevated temperatures. The standard approximation for calculating HSP at different temperatures comes from
Hansen and Beerbower:
Equ. 24-2 d/dT δD = -1.25 δD
Equ. 24-3 d/dT δP = -  δP/2
Equ. 24-4 d/dT δH = -(1.22 x 10-3 + /2)δH
172
 is the thermal expansion coefficient. For the IGC technique to become more popular we need ready access to
reliable calculations of HSP at different temperatures for the common solvents so that data from different IGC
users will be comparable. HSPiP includes those calculations based, wherever possible, on known temperature
dependent s. The full Sphere Solvent Data set contains the appropriate coefficients for calculating  at a given
temperature from the constants a, m and Tc (the critical temperature):
Equ. 24-5  = a*(1-T/Tc)m
Note that this equation is only valid over a defined temperature range. For solvents with no constants or outside
their valid range, HSPiP uses a user defined default value for .
However, many molecules which are adequately “liquid” at room temperature can be used as the “solid” phase
for IGC. The paper by K. Adamskaa,, R. Bellinghausen, A. Voelkel, New procedure for the determination of
Hansen solubility parameters by means of inverse gas chromatography, J. Chromatography A, 1195, 2008,
146–149 measures the HSP of a group of excipients: Cetiol B, Lauroglycol FFC, Labrasol, Miglycol and
Tween 80. They used a similar fitting procedure for deriving the HSP. Professor Voelkel kindly sent the Chi
data for Tween 80 and here is an Excel fit of their data:

Figure 24-3 IGC fit of Tween 80 data


The fitted HSP for Tween 80 [19.9, 5, 6.9] gives a surprisingly high δD value – for reasons that are unclear at
this stage of the science of interpreting these data.
The other issue is that the calculation of Chi from Vg requires knowledge of the partial pressure of each solvent
at that temperature plus the second virial coefficient. This is an important point. If we measured the retention
time of a new solvent on a range of IGC columns, we could do exactly the same calculation to determine the
HSP of that solvent. But how many of us know the partial pressure and second virial coefficients of our new
solvents?
As an aid to the IGC community, and as an attempt to provide a common standard for these tricky calculations,
a calculator for converting specific retention volumes, Vg, to Chi parameters is provided. A representative
range of solvents is provided, spanning a good range of HSP space. For each solvent Tc, Pc and Vc (the three
Critical Parameters) and Antoine A, B, C parameters are provided. These in turn are used to calculate P0 (the
saturated vapour pressure) of the probe solvent and B11, the Second Virial Coefficient. It turns out that there are
a number of alternative methods for estimating B11 from the Critical Parameters, and they give different results.
Happily, the exact value of B11 is not hugely important so we don’t have to worry too much. To help users, two
representative methods are used. The first has been popularised by Voelkel and uses Tc and Vc. The second is
coded BNL after Blu, Nesterov and Lipatov, which uses Tc and Pc. In general the BNL method gives larger
173
(more negative) B11 values than the Voelkel method. The program shows you the B11 values (and P0) so you
can see for yourself. You can find out all about these methods (and others) from A. Voelkel,, J. Fall, Influence
of prediction method of the second virial coefficient on inverse gas chromatographic parameters, Journal of
Chromatography A, 721, 139-145, 1995. The only other inputs are the density of your test material and its
“molar volume”. Both values are controversial. The density, of course, changes with temperature. But as the
molar volume of the solvents also change with temperature (and are not explicitly calculated) the assumption is
made that the relatively small density effects cancel out. The “molar volume” is often an estimate – what, for
example, is the molar volume of a polymer? The recommended guess is to put in the molecular weight of a
typical monomer unit. Errors in this parameter mostly result in an offset of the calculated Chi values rather than
a significant change in their relative values. The calculation of HSP is relatively insensitive to this offset.
Because you can instantly see the effects of varying the density and the “molar volume” you can make your
own judgement of how valid the calculated HSP really are. It’s important to note that these sorts of judgements
about B11, density and molar volumes are generally not explored in IGC papers. By making them explicit, and
by encouraging the use of the same basic dataset, we hope that the IGC community will start to standardise on
one approach so that calculated HSP are relatively more meaningful.
So although IGC has considerable promise for determining HSP of polymers and looks excellent for
intermediate molecules such as excipients, oligomers and surfactants at room temperature, its use for taking a
set of known stationary phases and deducing HSP of solvents looks like something that is not for the faint-
hearted.
There’s one key point about IGC. It’s relatively easy to get good fits from just about any data set. But the value
of that fit is questionable unless the test solvents cover a wide region of HSP space. This is the same issue as
other HSP fitting. Having lots of data points in one small region is far less valuable than having the same
number of data points spread out over the full HSP space. One IGC paper used 5 similar alkanes, 2 similar
aromatics, 3 chlorinated hydrocarbons and one simple ester. This is a very small portion of HSP space! So if
you want really good IGC data make sure you challenge the material with solvents from the full range of
alkanes, aromatics, ketones, alcohols, esters, amines etc. etc. The wider the range, the more likely it will be that
your calculated HSP are meaningful. That’s why in the Vg to Chi calculator we’ve included not only solvents
typically found in IGC papers but a few more that will help cover a broader HSP range. By providing a
consistent set of Antoine and Critical Parameter values we will have saved IGC users a lot of work!
Update for the 5th Edition
Extensive work with Dr Eric Brendlé of Adscientis in France has shown that IGC is a very powerful tool for
measuring the HSP of the sorts of excipients commonly used in cosmetics and pharmaceuticals. By examining
a very extensive dataset of 30+ molecules we were able to identify some procedural issues with the IGC
technique (to state the obvious, the wrong support material can cause major artefacts) and with some subtle
details of the IGC optimisation software. The 5th edition IGC tool now allows you to load Vg data providing the
solvents are named exactly in the manner of the Vg conversion table. It then becomes very easy to go from Vg
to HSP. It has taken a long time, but the promise of IGC as a general-purpose tool for measuring the HSP of
liquids is finally taking shape – as long as the liquid doesn’t volatilise on the column! So it can only be used for
low boiling solvents, oligomers etc. and not for simple solvents.
GC prediction via GCRI – Gas Chromatography Retention Index
It would be very useful if we could predict in advance the exact retention time of a chemical on a GC column.
But of course this is not directly possible as retention time depends on numerous specific variables in the GC
analysis: column type, column dimensions, temperatures, flow rates etc.
A useful substitute is the GCRI. We know that the retention time of straight chain alkanes form an orderly
progression from CH4 to C2H6 to C3H8 … And if we give each of these alkanes a retention index = 100 x
number-of-carbon-atoms we can say, for example, that if a chemical elutes at precisely the same time as
heptane then its GCRI is 700.
If the chemical elutes somewhere between hexane and heptane then, by definition, the GCRI is somewhere
between 600 and 700. Kovats proposed a formula for calculating the GCRI. Suppose the lower n-alkane elutes
at time A, the higher (n+1)-alkane elutes at time B and the chemical elutes at a time X between A and B then:
Equ. 24-6 GCRI = 100 * ( log(X)-log(A))/(log(B)-log(A)) + n * 100

174
If, in the example above, hexane eluted at 60 seconds, heptane at 80 seconds and the chemical at 70 seconds
then GCRI = 100 * (1.845-1.778)/( 1.903 - 1.778) + 600 = 654
If we can predict GCRI from the chemical structure then we can provide an accurate estimate of the retention
time in a specific column provided we know the retention times of the alkane series.
The simplest way to estimate GCRI is to say it depends only on the boiling point of the chemical. This turns out
to be an impressively good predictor. The reason for this is simple – at the very low concentrations of the
chemicals in the GC the behaviour of the gases is close to “ideal” so molecules hop on and off the support more
or less according to their volatility.

But “impressively good” isn’t “good enough”. In the figure for GCRI values on a DB1 (Polysiloxane) support
the general trend with boiling point is clear, but when you look more closely at the graph you see that for a
given boiling point the GCRI can vary by 200-300 – much too large an error to be a useful predictive tool.

Clearly the assumption that the chemical is behaving in an “ideal” manner is not working. And of course we
know that different chemicals will interact more or less strongly with the support phase. And a good measure of
that interaction is the HSP distance.
It then seemed obvious to add a correction term that depended on the classic HSP distance calculation involving
δD, δP and δH and the values for the chemical, c, and support s. When this was tried, the results were worse
than using just the boiling point! What had gone wrong? One hypothesis is that the δD value is already
“included” in the boiling point (for example, Hildebrand and Scott have shown that for simple alkanes, δD
²=RT+17.0*Tb + 0.009*Tb2). If the δD term is removed from the distance calculation so that:
Equ. 24-7 Distance² = (δPc-δPs) ² + (δHc-δHs) ²

175
then the fit becomes excellent:

The correlation formula used (via a Quantitative Structure Property Relationship model) is:
Equ. 24-8 GCRI = -220.27+209.51*(BPt*0.000519+1)^8.52* (DistPH*-0.0372+1)^0.156
Well, it’s almost excellent. The main exceptions are the carboxylic acids, shown in blue. Perhaps this is because
the acids tend to form dimers:

We’d like to claim that we have a universal formula for GCRI. However, the only other extensive dataset
available to us, using the DX1 column, has a slightly different fitting formula which also requires the MVol.
We would need more datasets before being able to resolve this dilemma. So the GCRI modeller lets you choose
between these two (quite similar) columns. If you try other values for the HSP of the columns we can’t say
what your predicted results will be like. If you have access to other GCRI datasets for other columns, we’d be
delighted to fit them and try to get a more generalised rule.
In the HSPiP software we implement the correlation via the Y-MB functional group method using neural
network fits for the boiling point and the HSP, though manual entry of the values is also possible if they are
known independently.

176
In this example the GCRI for a branched aldehyde/alcohol is calculated as 951. If you wanted to test the purity
of this compound in a GC column you would simply set up conditions that gave you reasonable elution times
for nonane and decane and you would find your compound nicely resolved. If you thought that your sample
would have impurities with one less and one more carbon in the main chain, then the GCRI values for these two
molecules are quickly calculated as 881 and 1029 respectively so you would set up your GC column for
conditions where octane to decane (or undecane) eluted in reasonable times. If you found a peak at around 1054
some simple experimentation with likely contaminants show that it is likely to be the dialcohol:

If you wanted to explore the effect of changing to a support with different HSP values then it is simple to find
that, for example, that the original molecule will move from 951 to 1020 for a slightly more polar column.
Changes in the δD value make, of course, no difference to the calculated GCRI.

177
HSPiP, therefore, provides you with a very powerful predictive tool for your GC work.

HPLC
There are good theoretical reasons for believing that High Performance Liquid Chromatography (HPLC) should
conform exactly to HSP for systems other than those that depend on ionized (acid/base) or size-related
separation. The principle is simple: the analyte has to “choose” between residing in the stationary phase of the
column and the mobile phase of the eluent. And the “choice” is a simple HSP calculation. The parameter of
interest is k’, or “capacity factor” (which is calculated from the retention time and the column volume, so a
large k’ represents a large relative retention time), and the key formula is:
Equ. 24-9 ln k’ = C1 + C2 * MVol * (Distancema – Distancesa)
where C1 and C2 are constants, m=mobile phase, s=stationary phase and a=analyte, MVol is the molar volume
of the analyte and the distances are the classic square distances:
Equ. 24-10 Distance2ma = 4*(δDm-δDa)2 + (δPm-δPa)2 + (δHm-δHa)2
Equ. 24-11 Distance2sa = 4*(δDs-δDa)2 + (δPs-δPa)2 + (δHs-δHa)2
What is surprising is that these simple formulae have been used very little. Why is this? Sadly, the pioneering
papers were written on the basis of rudimentary Hildebrand solubility parameters. As we now know, without
the partitioning into D, P and H, the resulting correlations simply aren’t reliable enough and the technique of
correlating with solubility parameter quickly fell into disrepute. Later attempts failed because reliable values of
HSP were hard to find and relatively small errors in HSP lead to large errors in the predicted k’. For example,
Ethyl benzene and Methyl benzene (Toluene) are very close in HSP space and if we were interested in them as
solvents we would not be too bothered if a parameter were out by 1 unit. But for HPLC, Ethyl benzene and
Toluene are classic test analytes which separate readily and although the MVol surely has the most significant
effect (123/107), a small error in the HSP can rapidly be amplified by the squared effect in the two distance
calculations.
An excellent data set covering a good range of analytes, columns and solvent mixes comes from Professor
Szepesy’s group at the Technical University of Budapest. The group have explicitly rejected solubility
parameters because they were not of sufficient predictive value.
It was a good test to use the most recent HSP data set to see if it could do a good job in fitting the data. This is a
tough challenge as there is a lot of high-quality retention data to be fitted using pure thermodynamics with no
adjustable parameters. At the start of the fitting process the key unknowns are the HSP of the column materials
and the HSP of the eluent. The prime data set uses 30/70 acetonitrile/water and it was straightforward to
confirm that by using the 30/70 mixture of the standard acetonitrile and water values there was an adequate fit.
The data could then be processed to find that [15, 0, 1] provided an adequate fit for the first stationary phase
tried – reasonable for a classic C18 column. Other columns in the data set surprisingly gave [15, 0, 1] as about
the best fit, including one –CN column that is supposed to be more polar. The differences in k’ values between
178
the columns depended strongly on the Slope factor (C2). The variation from column to column made no
obvious sense to us till the correlation was found with the %C values (% Carbon) provided for each column in
the data set. Large %C’s gave large slopes (and therefore large separations).
The challenge now got tougher given that the fits had to work not only for one solvent blend but for a range.
This was initially unsuccessful. The changes in k’ values were larger than the change in the HSP of the solvent
blend would naturally produce. The literature provided various suggested fixes for this much-observed
phenomenon but nothing simple and elegant suggested itself. The other problem was that the basic theory said
that the Slope value should be a constant, equal to 1/RT whereas it depended on the column and was a factor
~25 smaller than 1/RT.
In the end a simple solution emerged that gave adequate data fits across 60 data sets: 5 columns, two different
solvents (acetonitrile, methanol) and 6 solvent ranges (20, 30, 40, 50, 60, 70% solvent). The Slope term was
modified to C2 * (1-0.013 * % Solvent). So far there is no theoretical explanation for this fudge factor nor for
the fact that C2 is out by a factor of ~25, but the fit to a large number of totally independent data sets seems to
be compelling evidence that the approach has many merits.
Having now fitted many 10’s of data sets with a considerable degree of success, we need to point out that
HPLC systems do not operate in the sort of HSP zone to which we are all accustomed. The reason for this is
explained in a groundbreaking paper on using solubility parameters (alas, ahead of its time) by Schoenmakers’
team in Chromatographia 15, 205-214, 1982. They show that the stationary and mobile phases are very far
apart in HSP space. This amplifies the separation and (therefore) selectivity. What follows from this is that the
HSP of the stationary phase have a surprisingly small effect on the separation. Changes in the mobile phase
have a significant effect, but even this is modest when you consider the vast change in HSP between, say, 100%
acetonitrile and 100% water. The clever thing about HPLC is that the only thing which makes a big difference
to retention time is the molar volume and HSP of the analyte. It’s worth repeating that the modest effect of
mobile and stationary phase and large effect of HSP falls very naturally from Schoenmakers’ analysis.
What’s not so obvious is the very significant effect from molar volume. The large separation between toluene
and ethyl benzene or between the various parabens also included in the Szepesy data set comes almost entirely
from the molar volume effect. One obvious explanation for this is some sort of diffusion limited or size-
exclusion effect. But that’s not what the basic theory says. The molar volume term comes in simply because the
thermodynamics say that interaction energies are HSP/Molar_Volume. This large effect and its scientific
explanation are a particularly satisfying vindication for the HSP approach.
Here are two plots that show the reasonable degree of success of the technique. The data are for columns at
either end of the spectrum. The first is a classic hydrophobic C18 column and the second is a classic polar CN
column. The results are based on an optimum HSP of [15, 0, 1 ] value for both stationary phases as discussed
above. For the CN column the retention times are rather small, particularly for the left-hand side of the graph.
In the original paper, the authors discarded the data points for Caffeine and Hydroquinone because their
retention times were too small. In each case the plots are of actual v predicted ln k’. Below the plot of the fit is
an indication of what the idealised HPLC trace would look like with the fitted values. The value of this will
emerge in a moment:

179
Figure 24-4 A reasonable correlation for the C18 column:[15, 0, 1]

180
Figure 24-5 The correlation for the CN column: [15, 0, 1] It’s noisier partly because it spans only a small range of column
volumes. The data set is slightly different from the one above, reflecting slightly different experimental protocols in the original
papers
The software allows you to interact freely with the data. You can create mobile phase parameters directly by
choosing the solvent (Acetonitrile, Methanol or THF) and the relative percent with water – or you can directly
enter your own values. The Calculator button automatically finds a slope/offset combination that give you a
slope of 1 for the best fit line. You can alter these values manually if you wish.
The output includes the plot itself (moving the mouse above a data point reveals the solvent, actual and
calculated data) then the intercept, slope, R² (goodness of fit) Standard Deviation and then the range over which
the separation works in terms of k’ and in terms of column volumes. You can see that the –CN column fits all
the chemicals in a range of 1-2.018 column volumes whereas the C18 covers 1-7.076 column volumes – giving
you a choice of speed versus resolution.

181
The same C18 column run with 60% acetonitrile gives a relatively good fit. No fitting parameters were changed
compared to the 30% values:

Figure 24-6 The correlation for the C18 column at 60/40 acetonitrile/water
The fact that the same fitting parameters work for a large shift in the solvent blend indicates the power of the
technique.
You can play live with these data yourself (plus the other HPLC data we have provided) and reach your own
conclusion. One thing that playing “live” means is that you can alter the individual HSP from the standard
HSPiP main form and see the effect on the HSP fit. We did this ourselves. The first was to fit p-Nitrophenol. It
was obvious on the plot that the fit was very poor and we played with the values till the fit seemed reasonable.
Only then did we realise that the published value for p-Nitrophenol was, regrettably wrong; the δP of 20.9
simply makes no sense. The second was to play with Caffeine - the point nearest the 1 column volume elution
point and therefore the least accurate. The official value is an estimate. Perhaps the HPLC data are telling us to
182
revise that estimate. The final adjustment was to Pyridine. The published value was re-checked and makes good
sense, but was way off in the plots. When we looked at other examples of Pyridine in HPLC we noted that it
was normally eluted as Pyridinium as the mobile phase is typically buffered to pH 3.0. Although the Szepesy
data set was explicitly run unbuffered, we suspect that the value we have entered is for Pyridinium rather than
Pyridine.
At this point the reader can go in two directions. The first is to say “They’ve fiddled the data to get the answer
they wanted”. The response to that is that most “fiddling” with the data, although it could improve any single
plot, generally made the plots over the full Szepesy data set positively worse. The second direction is to say
“This looks like a powerful technique for understanding and predicting HPLC behaviour”. That’s why we’ve
added the k’ range output (so users can think which column and/or solvent blend to use) and also the predictor.
You can see in the screen shot how an analyte of [19.5, 13, 15] with a molar volume of 230 would behave in
each of these columns.
The second direction makes sense. HSP and HPLC are a powerful combination – for predicting/understanding
HPLC and for generating HSP values for use in other applications.
That’s also why the simulated elution plot has been included. When you change the % solvent (or the solvent
itself) you can see how the eluted peaks are predicted to change. You can quickly see how two peaks close
together can be separated or vice versa as you change the elution conditions. Similarly, if you click the Show
Solvent Ranges you get a graphical idea of how the k’ of the individual analytes vary over the full 0-100%
solvent range.
Expert HPLC users will object that the predictions of k’ are unlikely to be precise enough – they are, after all,
predictions of ln k’, which can easily compress errors. This may well be true. But look at what the HSP
approach offers. With one rather simple formula, with very few fitting parameters (and maybe after more
investigation, no fitting parameters at all), a user or manufacturer can understand numerous fundamental
aspects of what is going on in a separation. It seems, at the very least, that the HSP-HPLC approach should be
investigated more deeply with alternative data sets.
Finally, here’s a data set prepared independently for us by Andrew Torrens-Burton, an expert Analytical
Chemist at MacDermid Autotype. He gathered data for three different types of analyte, deliberately avoiding
the aromatic test analytes used in the Szepesy data. The standard values for the C18 column calculated from the
Szepesy data set were used for this analysis except that the slope is bigger – because this column happened to
be twice the length of that used by Szepesy:

183
Figure 24-7 The correlation for the set by Andrew Torrens-Burton

NMR Solvents
This is a very short section. We have not, so far, attempted any correlations between NMR and HSP, though it
is clear that, for example, polymer-polymer interactions in solution will depend strongly on polymer-solvent
interactions which can, in turn, be predicted via HSP.
But we see a very powerful use of simple HSP ideas to help the practical NMR spectroscopist.
One typical challenge is to find a solvent for carrying out NMR experiments for a given chemical. There are
relatively few H-free solvents (i.e. deuterated chloroform, benzene etc. plus carbon tetrachloride) and if your
chemical is not very soluble in any of them then you have problems.

184
But we know from HSP theory that a chemical can be insoluble in each of two solvents yet be soluble in a
mixture, if that mixture is a good match for the HSP of the chemical.
We’ve therefore provided a Solvent Optimizer dataset containing the most used H-free (deuterated) solvents.
If you enter the Target HSP of your chemical (perhaps estimated via Y-MB) and click the 2 button you
instantly find a blend that gives you an optimum match. To be more precise you can click the 3 button to find a
3-solvent blend. If you are worried about cost you can even weight the solvents (high cost = low weight) so that
the optimum blend uses less of the more expensive ones.
Armed with those data it is highly likely that you will find it much easier to get a higher concentration of your
sample – at very little cost and very little trouble. In the following example a tricky [18 8 8] material is likely to
be nicely soluble in the chloroform/DMSO mix – and it is highly unlikely that you would have tried that mix
without help from the Solvent Optimizer.

Figure 24-8 An interesting mix of NMR deutero-solvents

185
Chapter 25 Noses artificial and natural (HSP for Sensors Both Artificial
and Live)
The sense of smell is still something of a mystery. But there are determined efforts to remove that mystery and
there’s a good chance that HSP can play their part.
Professor William Goddard’s group at CalTech provides a good example of how HSP can be used to
investigate both artificial and natural noses. Along the way, the group is also providing new tools for
calculating HSP. We are grateful to Professor Goddard and his colleague Dr Mario Blanco for giving us access
to their data and insights.
The artificial nose
A fruitful approach to producing an artificial nose is to provide a sensor made up of an array of sub-sensors
each of which has a different response to an odorant molecule. Whilst the response from any individual sub-
sensor might not tell you too much about an odorant, the “fingerprint” of responses might be sufficiently unique
to tell you what odorant is there and how much of it is present.
The “JPL Artificial Nose” works in this way. There are seven sub-sensors, each a simple polymer:
Poly(methylmethacrylate) (PMMA), Poly(4-hydroxystyrene) (P4HS), Polyethyleneoxide (PEO), Polyethylene
(PE), Poly(ethylenevinyl acetate) (PEVA), Polysulfone, and Polycaprolactone.
The hypothesis of the Goddard group is that the amount of interaction (and therefore sub-sensor signal) of an
odorant molecule could be predicted on the basis of its HSP distance from each polymer. Their work is
described in their paper (in collaboration with 3M) M. Belmares, M. Blanco, W. A. Goddard III, R. B. Ross, G.
Caldwell, S.-H. Chou, J. Pham, P. M. Olofson, Cristina Thomas, Hildebrand and Hansen Solubility Parameters
from Molecular Dynamics with Applications to Electronic Nose Polymer Sensors, J Comput. Chem. 25: 1814–
1826, 2004. Their analysis is based on their calculated HSP values from molecular dynamics. In this account
we use conventional HSP for the simple reason that the paper’s values for δP and δH are necessarily
constrained by the absence of an agreed methodology for producing these values from the MD data. We are
grateful for their permission to recast the data in our mode and we must stress our enthusiasm for the MD
approach where it is surely only a matter of time before the δP/δH problem is solved.
The aim of the paper was to see if there is a linear relationship between theoretical and experimental response
curve. The logic is that the sensor response depends on swelling of the polymer by the solvent and the closer
the solvent is to the HSP of the polymer the more swelling will occur. The fit was based on 5 parameters: a pre-
exponential term, a term relating to molar volume, then a term each for the absolute differences of the cohesive
energy terms values (polymer – solvent) of D, P and H.
Equ. 25-1 Response = P1 * Exp(-P2*MV)*Exp(P3*(δDp-δDs) + P4*(δPp-δPs) + P5*(δHp-δHs))
We decided to see what happened if we used pure HSP instead. In that case we have just 3 parameters: the pre-
exponential term, the molar volume term and term relating to the standard HSP distance.
Equ. 25-2 Response = P1 * Exp(-P2*MV)*Exp(P3*Sqrt(4*(δDp-δDs)2 + (δPp-δPs)2 + (δHp-δHs)2))
The results are interesting and encouraging. With reasonable values for the δD, δP and δH values of the
polymers, the fits were better (both in terms of slope and R2) than in the original paper. This isn’t as good as it
sounds. The paper used the computed values and produced an honest fit. Because we had no direct knowledge
of the polymers used we could “tweak” the polymer parameters (within reasonable limits) to get a good fit. To a
certain extent we could argue (see the section on Chromatography) that this is a good way to derive HSP for
polymers, but there is rather too much circularity in that argument.
Data from two of the polymers seem to be illuminating:

186
1.200

1.000 y = 0.9563x + 0.0084


R2 = 0.9237

0.800

0.600

0.400

0.200

0.000
0.00 0.20 0.40 0.60 0.80 1.00 1.20

Figure 25-1 Fit from the paper for the Polysulfone data using 5 fitting parameters

Figure 25-2 Fit of the Polysulfone data using the HSP formulation and 3 parameters

187
Figure 25-3 Using file NoseChems and a Polysulfone of [15.4, 4.5, 2.8] There’s a reasonable mixture of overlapping and non-
overlapping solvents, giving an overall wide-ranging response

0.90

0.80 y = 0.9673x + 0.0008


0.70 R2 = 0.9591

0.60

0.50

0.40

0.30

0.20

0.10

0.00
0.00 0.20 0.40 0.60 0.80 1.00

Figure 25-4 For P4HS there is a less uniform response in the 5-parameter fit

188
0.90

0.80 y = 0.954x + 0.0215


R2 = 0.985
0.70

0.60

0.50

0.40

0.30

0.20

0.10

0.00
0.00 0.20 0.40 0.60 0.80 1.00

Figure 25-5 With a slightly better, but still skewed fit from 3 parameters

Figure 25-6 With P4HS [18,8,2] there is very little polymer/solvent overlap, so it’s not surprising that most responses are
clustered near one end of a rather small response curve
So at the very least the “pure” HSP approach looks interesting. And the fact that 3 parameters suffice to fit the
experimental data from 24 solvents (using standard, un-tweaked HSP) with 7 polymers in a complex artificial
nose is at the very least encouraging.
Natural noses
We don’t want to get involved in the major debates on how real noses manage to distinguish between so many
different aromas. But it seems reasonable to most people that unless the aroma molecule has some affinity for a
receptor site then that site won’t be able to detect it. And as soon as the word “affinity” is mentioned, it
becomes natural to ask whether HSP could be a significant part of that affinity, and therefore a significant
predictor of smell. We say “significant” because it is unrealistic to expect that every receptor is simply HSP
generic. We are all familiar with the fact that biological receptors can be exquisitely specific (especially when it
189
comes to optical isomers). So it seems reasonable that there will be elements of specificity in nasal receptors.
But what seems to be clear is that no model based strongly on biological-style specificity has proven to be of
general utility.
So how might one show that HSP can be insightful for understanding olfaction? The hypothesis from Blanco
and Goddard (BG) is elegantly simple. For consistency with the rest of the book (and the eNose example
above) we recast their hypothesis (with their permission) in a slightly different formulation but the effect is the
same. They used Mean Field Theory (MFT) as their descriptor but we can think of it as HSP theory.
The BG-HSP hypothesis
Professor Linda Buck won the Nobel prize for identifying 47 different olfactory receptors. The BG-HSP
hypothesis states that each of these olfactory receptors is defined by a δD, δP, δH and Radius as if it were a
polymer. The response of each receptor to an odorant depends on the HSP distance of the odorant from the
receptor.
Thus the Responsejk of receptor k to odorant j is given by
Equ. 25-3 Responsejk = Sk * Exp(-k*Sqrt(4*(δDk-δDj)2 + (δPk-δPj)2 + (δHk-δHj)2))
which you will recognise as being almost identical to the eNose formula above – without the molar volume
term. The formula can be made even more familiar with one simple change:
Equ. 25-4 Responsejk = Sk * Exp(-Sqrt(4*(δDk-δDj)2 + (δPk-δPj)2 + (δHk-δHj)2)/Rk)

where we have replaced k with the more familiar Radius term Rk – so that the response has decreased by a
factor of 1/e by the time the HSP distance is equal to Rk.
The beauty of this formula is that it can readily be tested and BG have provided the first HSP values for
olfactory receptors.
The response of the S19 receptor to 19 odorants is shown. The Sphere fit gives the HSP for S19 to be ~ [16.9,
8, 9.4].

Figure 25-7 Using file OlfactionS19


The S83 receptor gives: [16.8, 8, 9.3]:

190
Figure 25-8 Using file OlfactionS83
The process can be repeated for the other receptors.
When you try these examples out for yourself you will quickly find that we have shown the best possible
interpretation of these data. The data sets are too small to provide good fits and it would be a massive task to
take on such a vast project.
But as the original BG paper shows, the idea is, at the very least, a very fertile one. If olfaction is, as they
guesstimate, 65% HSP and 35% specific receptor, then there will be plenty of noise in the data, but the HSP
signal should shine through if the hypothesis is correct. And of course, life can be more complicated. Perhaps
(and BG have evidence for this) some receptors have two HSP sites. A single Sphere fit would not do a good
job so a more sophisticated multi-Sphere calculator would be required.
The reason the BG-HSP hypothesis is so important is because if it were shown to be true it would not be yet-
another-correlation but a deep insight into olfactory receptors. Given that HSP are calculable ab-initio from
molecular dynamics, and given that HSP represent fundamental thermodynamics, then olfaction (to, say, 65%)
would become calculable from first principles.
If by the time you read this book the ongoing research has confirmed BG-HSP then we will be pleased that we
spotted the significance of the BG research before it had reached maturity. If it has been disconfirmed then
we’re pleased in another sense. For HSP to be good science it has to withstand the harsh standard of
disconfirmation. If it has proven to have failed on such a big task as olfaction it at least had the merit of offering
a clear prediction which could be refuted, and that’s one of the hallmarks of good science.
The Atlas of Odor Character Profiles
The Atlas, by Andrew Dravnieks, published by ASTM, ISDN 0-8031-0456-1, is a book of tables. For 144
odour chemicals plus a few more odour mixtures it lists to what extent a panel of skilled testers would say that
each chemical smelled like X, where X was a list of 146 different odour sensations such as Fruity, Almond,
Molasses, Yeasty, Incense, Kerosene, Sweaty, Heavy – to take a random cross-section through that list of
sensations.
In the spirit of making refutable predictions, it seemed a good idea to assemble the HSP of all 144 chemicals
then see how these fitted to the odour profiles. Of the 144 chemicals, many were in the standard Hansen table,
but most were not. Thanks to the generosity of SC Johnson, a list of HSPs of 288 odour chemicals (prepared
using calculation/estimation only by Charles Hansen) was made publicly available. This still left quite a few of
the 144 without HSP so Abbott used the DIY-HSP tools from HSPiP to estimate the remaining ones. Both the
SC Johnson list and the 144 list are made available as a contribution to further research on odours and
191
fragrances. The 144 list includes CAS Numbers and Smiles notation to help you make sure which chemical is
being referred to - naming of odorants is rather uncertain. Note that there are some minor errors in the Atlas.
Where possible the table contains revisions to these errors.
Matching chemicals to the different sensations was made possible thanks to further tables in the Atlas. These
listed the 5 highest-scoring chemicals for each sensation. The “minimum HSP sphere” that enclosed these 5
chemicals was then used as an indication of the hypothetical HSP centre/radius for each of these sensations.
Of course there are many problems with this procedure. First, some of the top 5 chemicals were from the
mixtures which have not been included in the HSP list. Second, some of the “top 5” have such low scores as to
make it seem unlikely that these sensations really do have well-defined chemical correlations. Third, the choice
of 5 is rather arbitrary. If the size of responses (where a large number means a strong response) go (43, 41, 38,
34, 11) should that 5th chemical (clearly much less relevant than the other four) be included? Or if the responses
went (43, 41, 38, 34, 30, 29, 28, 11) shouldn’t we include the top 7 chemicals?
But we have to start somewhere. We’re only trying to explore some basic hypotheses. Others can feel free to
refine the process if it seems to be worthwhile.
Although it was a lot of work, the “easy” part of the process was to identify the HSP centre/radius for each
relevant sensation. The raw data are provided for you to save you the tedium of creating it for yourself. There
were 70 aromas with meaningful high scores for which the centres of the minimum spheres were calculated.
Here is a screenshot from HSPiP showing 71 aromas in HSP space:

Figure 25-9 Using file AromaScores. See below for why this dataset is not part of later editions
The hard part is working out whether the data mean anything. The simplest case would be that each sensation
had a unique sensor which had a unique HSP for optimal binding. It’s obvious that aromas cannot work this
way. Some of the sensations must be complex mixes of different sub-sensations. And the most optimistic HSP
case would be that HSP compatibility is necessary but not sufficient – there must be a good HSP match for a
molecule to be happy in the sensor area, but there must be other molecule-specific attributes for the aroma to
register with the sensor.
An alternative would be to follow the process of A.M. Mamlouka, C. Chee-Ruiter, U.G. Hofmann, J.M. Bower,
Quantifying olfactory perception: mapping olfactory perception space by using multidimensional scaling and
self-organizing maps, Neurocomputing 52–54, 2003, 591 – 597. For those who are familiar with
multidimensional scaling and self-organizing maps the data from our explorations are provided. However, the
best that Abbott could do was to prepare a spreadsheet with a 71x71 matrix that calculated the HSP distance
192
between each of the aromas. It was then possible to sort each column to see if the ordering of the aromas made
sense. For example, if the target aroma was “bananas” which is arguably a pleasant aroma, other pleasant
aromas might be expected to be close by and disgusting odours would be far away.
There was, unfortunately, no compelling evidence for this happy outcome. Here is a small section of the matrix
ordered by distance between Banana and other aromas. Some of the fruity odours are gratifyingly close to
Banana, but Urine and Rancid are also fairly close and they are not normally associated with the aroma of
Banana.

Figure 25-10 A portion of the matrix ordered by the HSP distance between Banana and the other 70 odours.
Hansen published a paper in 1997 (Hansen, C.M., Aromastoffers Opløselighedsparametre (in Danish),
Solubility Parameters for Aromas and Scents. Plus Process, 11, 16-17,1997) which anticipated many of these
ideas. Readers may or may not like to know that it is possible to cover the smell of skatole (faeces) with
suitably chosen (i.e. a good HSP match) aromas from hamburgers or bacon.
3rd Edition update
With the benefit of hindsight, some of the ideas above do look naïve, and the quoted datasets are no longer
included. But it still seems to us that the world of fragrances is missing a trick by not taking HSP into account.
At the very least, the packaging industry could get a lot of benefit from the ideas of HSP and diffusion. If there
is a good HSP match of a key fragrance/flavour component, say, Cinnamon (cinnamaldehyde) with a packaging
polymer (such as poly lactic acid, PLA) then it’s highly likely that the polymer will be a poor barrier for it.
Similarly, if a fragrance component (or, more likely, fragrance formulation) has a good match for the HSP of
skin then it’s more likely to penetrate the skin and (most likely) be lost as an odour.
And clearly if a fragrance/flavour component is to be delivered within some polymeric system (e.g. scratchable
spheres) classic HSP calculations will help ensure a balance of good compatibility for creating the system and
poor compatibility to ensure that the fragrance remains locked in to the system till required.
Because Sigma Aldrich have provided a de-facto standard reference for aromas, with helpful designation of the
different types of smell, we are putting into the public domain an HSPiP version of the Sigma Aldrich Flavors
& Fragrances catalogue. This is a somewhat error-prone undertaking as anyone who has ever handled complex
datasets will know from their own experience. The catalogue doesn’t always provide CAS numbers and it
doesn’t supply the Smiles. So at times we used the also excellent GoodScentsCompany website and you might
find some alternative names for the same compound. We also had to decide what to include. Although we could
have included the “W” numbers from the catalogue, there are so many variants of essentially the same
compound that we decided it would not be helpful. Similarly, we could have provided the aroma class, but that
would have created much duplication. From the name and/or the CAS number you should be able to identify
193
most of the chemicals in the catalogue and therefore their aroma class. Finally, of course, we could not include
those aromas that are mixtures and sometimes even the wonderful ChemSpider could not help us identify the
right Smiles for a given compound. Nevertheless, you now have access to the HSP of over 800 aroma
chemicals which you can then cross-reference with the Sigma Aldrich catalogue for your own purposes.
When you load it into HSPiP you will see the usual HSP data. But the horizontal scroll bar will allow you to
scan across for the Smiles, CAS number etc. To search within the data you can try using a name (probably a
truncated version as naming is so variable) or a CAS number.
One way of exploring whether HSP have any relation to aromas is via a Self Organising Map (SOM). Hiroshi
has enjoyed playing with this concept and the following section shows the sort of exploration that can be done.
It’s included as an indication of what might be done if someone wanted to throw some serious resource at the
issue.
10 Fragrant Flowers
A Japanese website http://www001.upp.so-net.ne.jp/iromizu/hana_kaori_for_so-net.html lists the key
ingredients of 10 flowers:

From the HSP of these molecules, and the different fragrances, an SOM can be constructed on a 40x40 matrix:

194
From this partition it’s possible to ask many questions. A typical one is “can we distinguish some key distinct
notes from this?” And here’s one answer. 5 distinct areas stand out: Rose, Orange-colour olive, Lilac, Carnation
and Jasmine stand out.

On the other hand, Ixora, Narcissus, Jasmine are rather similar in SOM space:

195
As we’re not trained in fragrances we can’t comment on the significance of these plots. But it represents
another way of looking at how fragrances and solubilities may be related. If they show no relationship, that’s
useful to know. If there are such relationships (and the physiology suggests that there should be) then HSP
provide an opportunity for data mining in this fascinating area.

196
Chapter 26 Attacking DNA (HSP for DNA , Drugs, and Biological
Membranes Compared)
DNA is at the centre of our lives. In this chapter we discuss how the HSP of DNA can be estimated then how to
apply HSP thinking in saving lives directly (through more efficient DNA diagnostics) and indirectly (by
identifying one reason why cytotoxic chemicals might be cytotoxic).
The HSP of DNA
The HSP for DNA have been estimated in Chapter 16 of the Handbook using data from Ts’o, P.O.P.,
Helmkamp, G.K., and Sander, C., Interaction of nucleosides and related compounds with nucleic acids as
indicated by the change of helix-coil transition temperatures, Proc. Natl. Acad. Sci. U.S.A., 48, 686-698, 1962.
The HSP found are [19, 20, 11] with the same ranking of the magnitude of the effect of the chemicals as
reported in the experimental data.
The values seem believable, but it is noteworthy that the hydrogen bonding parameter is the least of the three.
This is in contrast to the many figures showing the bonding in DNA (bases) [19.75, 12.3, 12.2 average] as being
largely, if not exclusively, hydrogen bonds. As discussed below, the DNA “bases” considered as individual
chemicals are not water soluble. In addition to geometrical fitting, it is thought that this (i.e. the escape from
water) is a major reason for the “binding” of the DNA bases in aqueous media.
Faster DNA diagnostics
There is additional confirmation that the HSP for DNA given above are useful guides in:
Steen H. Matthiesen, Charles M. Hansen, Fast and Non-Toxic In Situ Hybridization without Blocking of
Repetitive Sequences, PLoS ONE: Research Article, published 24 Jul 2012 10.1371
As reported there, HSP methodology has helped to develop a revolutionary improvement in the Fluorescent In
Situ Hybridization method (FISH). Hybridization in this context is the process of joining two complementary
strands of DNA.
The following is reproduced from Wikipedia:
FISH (Fluorescence In Situ Hybridization) is a cytogenetic technique developed by biomedical researchers in
the early 1980s[1] that is used to detect and localize the presence or absence of specific DNA sequences on
chromosomes. FISH uses fluorescent probes that bind to only those parts of the chromosome with which they
show a high degree of sequence complementarity. Fluorescence microscopy can be used to find out where the
fluorescent probe is bound to the chromosomes. FISH is often used for finding specific features in DNA for use
in genetic counseling, medicine, and species identification. FISH can also be used to detect and localize specific
RNA targets (mRNA, lncRNA and miRNA) in cells, circulating tumor cells, and tissue samples. In this context,
it can help define the spatial-temporal patterns of gene expression within cells and tissues.
This new methodology has been dubbed IQFISH by DAKO Denmark A/S, an Agilent Technologies Company,
who started supplying kits for these analyses early in 2012. The test can now be accomplished in one hour
rather than 2 days with complete confidence in a non-toxic analysis that does not use the traditional solvent
formamide. The key is use of alternative solvents with ethylene carbonate (EC) [19.4, 21.7, 5.1] being
preferred. When combined 50/50 with (structured) water [18.1, 17.1, 16.9] the EC/water mixture has HSP
[18.7, 19.3, 11.0]. These values are essentially identical with the HSP for DNA [19, 20, 11]. Neither
(structured) water nor EC have HSP that resemble the HSP of the DNA bases (see below also). There is no
hindrance to reestablishing the usual binding (hybridization) when this is desired.
Here is the abstract of the article:
Formamide is the preferred solvent to lower the melting point and annealing temperature of nucleic acid
strands in in situ hybridization (ISH). A key benefit of formamide is better preservation of morphology due to a
lower incubation temperature. However, in fluorescence in situ hybridization (FISH), against unique DNA
targets in tissue sections, an overnight hybridization is required to obtain sufficient signal intensity. Here, we
identified alternative solvents and developed a new hybridization buffer that reduces the required hybridization
time to one hour (IQFISH method). Remarkably, denaturation and blocking against repetitive DNA sequences
to prevent non-specific binding is not required. Furthermore, the new hybridization buffer is less hazardous

197
than formamide containing buffers. The results demonstrate a significant increased hybridization rate at a
lowered denaturation and hybridization temperature for both DNA and PNA (peptide nucleic acid) probes. We
anticipate that these formamide substituting solvents will become the foundation for changes in the
understanding and performance of denaturation and hybridization of nucleic acids. For example, the process
time for tissue-based ISH for gene aberration tests in cancer diagnostics can be reduced from days to a few
hours. Furthermore, the understanding of the interactions and duplex formation of nucleic acid strands may
benefit from the properties of these solvents.
The following is also taken from the above article (with short additions for clarification) to help explain what
has been done:
A novel hybridization buffer that dramatically reduces the hybridization time is described. This is a buffer that
challenges the dogmas of heat-induced denaturation of double-stranded nucleic acids and of blocking against
repetitive sequences in probes of genomic origin [14] to perform hybridization. These findings will have a
major impact on hybridization based cancer diagnostics and research. In the search for formamide substitutes
that are less toxic and can reduce the hybridization time, the Hansen solubility parameters for DNA [18] were
used as a guidance to identify potential replacement candidates. The solvents: ethylene carbonate, sulfolane,
propylene carbonate, γ-butyrolactone, 2-pyrrolidone and δ-valerolactam were identified. All of these solvents
have reasonably similar HSP, but ethylene carbonate (EC) is preferred because of its water solubility. By
lowering the denaturation temperature from 82°C to 67°C, background staining was reduced. Last, but not
least, EC is non-toxic at the concentration used. A proposed reason for the success of this procedure is that
during the hybridization (restoration of original structure in the DNA) strands can gain easier access and bind
to their complementary strands when there are no effects from the solvent disturbing the base pairing [5], [26].
The new hybridization solvents are strong candidates to replace the use of classic formamide as the preferred
solvent in molecular biology due to their properties to lower the melting temperature, increase the
hybridization rate and decrease health risks. In addition to the results shown in this paper, they also work well
for e.g. LNA (Locked Nucleic Acids) and DNA oligo probes, RNA detection, as denaturants and for stringent
wash (WO 2010/097655; WO 2010/097656; WO 2010/097707). The shortened hybridization time of the IQISH
technology will have a major impact on ISH based cancer diagnostic as the turnaround time from sample to
diagnosis makes a difference for the patient.
Cytotoxicity
An attack on our DNA is an attack on our life or on our quality of life. Such an attack is also required of the
drugs used in chemotherapy. It is therefore rather important that we know if a chemical is likely to interact in
some way with this complex molecule, for example being cytotoxic.
A few moments thought would suggest that HSP could have nothing to say on the subject. Cytotoxicity must be
a hugely complex activity in a complex environment.
However, when a group of well-known cytotoxic chemicals used in chemotherapy all showed HSP values
clustering around a certain value, it seemed a good idea to check whether this was chance or a deep insight. The
gold standard of science is disconfirmation of a hypothesis so it seemed fairly easy to hunt for cytotoxic
molecules with utterly different HSP, thereby refuting the hypothesis.
The fact that we’re writing this chapter means that finding such a refutation has proven harder than we’d
supposed!
The core data came from work designed to find gloves that were safe for handling well-known cytotoxic drugs
and is described in C.M. Hansen, Polymer science applied to biological problems: Prediction of cytotoxic drug
interactions with DNA, European Polymer Journal 44, 2008, 2741–2748. The technique used for estimating
breakthrough times was the based on the same type of correlation as described in the chapter on skin/glove
diffusion and the following results emerged:

Group 1 δD δP δH V Ra (ave.)
Fluorouracil 18.0 11.7 11.6 118.3 1.68
Gemcitabine 19.0 12.6 15.5 260.6 4.12

198
Cyclophosphamide 17.5 11.9 12.6 279.1 2.28
Ifosfamide 17.5 11.9 9.8 261.1 3.37
Methotrexate 18.0 10.2 14.2 378.7 1.99
Etoposide 20.0 7.5 12.5 588.5 4.40
Paclitaxel (Taxol) 18.0 6.6 9.8 853.9 4.50
Average of Group 1 18.3 10.3 12.3 - -
Group 2
Cytarabine 19.0 15.2 20.1 187.1
Carboplatin 27.3 9.0 10.4 185.1

Table 26-1 HSP properties of many cytotoxic drugs. The Ra is the distance to the average
What is interesting is that the 4 base segments included in DNA have the following values:
Segment δD δP δH V
Guanine 20.0 12.7 12.5 126.1
Cytosine 19.5 12.1 9.9 107.8
Adenine 20.0 10.2 13.7 131.5
Thymine 19.5 14.2 12.6 121.7
Average 19.75 12.3 12.2 -

Table 26-2 HSP of DNA bases


At the very least, the “coincidental” similarity of the HSP of the bases and of the cytotoxic drugs was worth
investigating further.
For a drug to be cytotoxic it actually has to reach the DNA. It therefore has to pass through cell walls. The
chapter on Skin has already indicated that passage through (skin) cells requires the following HSP:
δD δP δH
Skin 17.6 12.5 11.0

Table 26-3 HSP of Skin


Again, is this another coincidence?
So let’s look at another set of well-known harmful chemicals:
δD δP δH V Ra (DNA bases)
Average for Group 1 18.3 10.3 12.3 - 2.00
Thalidomide 20.0 11.3 10.2 195.6 2.29
Pyrimidine 20.5 9.4 11.3 78.8 3.39
1,2-Benzoisothiazolin (BIT) 20.0 9.4 9.2 126.0 4.20
Doxorubricin 19.9 8.6 15.1 483.3 4.71
Dioxin 20.0 9.2 7.6 208.2 5.57

Table 26-4 Some well-known harmful chemicals


We now introduce the HSP distance (Ra) from DNA bases as a predictor of cytotoxicity. By the time we reach
dioxin we are at a bigger distance and at a far less potent molecule. Doxorubricin is a potent molecule but its
distance is rather large. However, it is a complex molecule for which the group contribution calculation may

199
not be too accurate and intuition suggests that the δH should be closer to 13 rather than 15.1, leading to a
distance of 3.8. It will be interesting to obtain more accurate values via molecular dynamics or by experiment.
It’s worth attempting another challenge. So let’s look for other cytotoxic papers in the literature. Carr J Smith’s
group at Reynolds Tobacco identified the cytotoxicity of various substituted quinolines. The 4 most potent have
the estimated HSP (using HSPiP’s Stefanis-Panayiotou estimator) shown below. The fit with the hypothesis is
quite acceptable.
Substituent δD δP δH
8-OCOCH3 19.9 7.7 8.6
8-NH2 22 10.4 12
8-OH 20.8 9.8 14.4
8-Cl 21.2 8.6 6.6
8-OCH2Ac 21.9 7 4.7
Average 21.2 8.7 9.3

Table 26-5 HSP of some substituted quinolines


So far, the hypothesis is looking reasonable. But there are plenty of other molecules with HSP in the area of
interest. How toxic are they? By entering the DNA average for the bases into the Polymer table, selecting a
Radius of 4 and clicking the Solvent Match button, the following plot appears if the whole HSPiP Master Data
are loaded:

Figure 26-1 DNA (almost hidden in the cloud of blue) compared to the whole solvent range
Here we hit an immediate problem. Of those molecules with RED < 1 both Caffeine and Vanillin stand out as
chemicals we don’t think of as cytotoxic. However, there is a large body of evidence showing strong
association of caffeine with DNA. For Vanillin there is no major effect, but there is still some doubt in the
literature as to just what is going on (it may convert to its acid form, changing its HSP).
Is this sufficient to refute the hypothesis? Against the naïve claim that HSP match = Cytotoxicity then a
refutation is easily found. But the claim is an “HSP and…” hypothesis. We are making the claim that the HSP
match is a necessary condition for a molecule to be able to get through to, and associate with, DNA. Necessary
and sufficient requires something more than getting through to the DNA. The obvious extra function would be
reactivity - and some anti-cancer drugs are known to be reactive once they associate. Also, the way an

200
associated molecule affects binding during the replication/transcription processes will be a factor in cytotoxicity
as it may change the way that the processes are carried out.
We can use HSP to speculate a little further. Could it be, for example, that the well-known secondary effects of
ethanol are due to its ability to “help” a “bad” molecule to cross biological membranes? Here are two simple
examples.
The following figure shows the HSP sphere for the cytotoxic drugs. The red cubes are for ethanol, dioctyl
phthalate (DOP), and their mixture at 50/50. The 54/46 ethanol/DOP is in blue, being defined as being just
inside the sphere. This shows the distinct possibility for synergism of given chemicals with alcohol to allow
passage of biological membranes. Once inside a cell, such chemicals can physically get in the way of a process.

Figure 26-2 A 54/46 Ethanol/DOP mix is shown just inside (RED=0.992) the cytotoxic sphere
The next figure uses the solvent optimizer with a choice of chemicals to give an essentially perfect HSP match
to the center of the cytotoxic drug HSP sphere. Methyl paraben is already within the sphere from the start with
a distance 2.7 compared to the radius 4.4. This emphasizes that mixtures of chemicals, and perhaps especially
alcohol, can give synergistic effects in biological systems. This is clearly no proof of any effect, but it deserves
thought and perhaps also experiment.

Figure 26-3 Methyl paraben combined with ethanol produces a near-perfect match
It is obvious that we are not experts on cytotoxicity. But what we feel is that the “HSP match is necessary”
hypothesis is, at the very least, worthy of further consideration. Because the biology world have hardly heard of
HSP it’s not surprising that they’ve not tried to take them seriously. We believe that HSP, because of their
thermodynamic grounding, are a worthy alternative to endless QSAR correlations which provide nice numbers

201
but lack the fundamental grounding (and success over a wide field of research endeavours) of HSP. Perhaps this
chapter will persuade those in the biological world that it’s worth a try.

202
Chapter 27 HSP for Pharma and Cosmetic applications
In previous chapters there have been numerous hints about how the Pharma and Cosmetics world can make use
of HSP.
• Knowing the HSP of a pharmaceutical or cosmetic ingredient allows you rationally to think through its
general bio-compatibility with DNA, skin etc.
• Formulating ingredients for controlled compatibility or, if required, incompatibility, is a rational
process. The evaporation modeller in the Optimizer can help you understand how a solvent blend might
change during evaporation – keeping critical components in solution or, if you wish, ensuring that some
components fall out quickly and, for example, concentrate at the surface.
• The analysis of skin permeability based on Jmax instead of permeability coefficient seems to offer a lot
of insight into practical skin formulation issues – in particular making the effects of permeation
enhancers fall into a rational process of thinking about solvency rather than the restrictive and confused
terminology such as “lipophilic v hydrophilic”.
• The Diffusion modeller can be used for many different issues. Permeability through skin is one
example, permeability out of microcapsules is another. The HSP distance is crucial for understanding
the solubility part of the Jmax = Solubility * Diffusion Coefficient/Thickness equation, and the diffusion
coefficient itself can be strongly concentration dependent, another reason for knowing the HSP distance.
From knowledge of the HSP distance it is possible to predict whether a microcapsule is likely to show
Zero Order diffusion (a large distance) or Fickian diffusion (a smaller distance). Which behaviour is
desired depends on the application, but it can be built-in rationally.
As an aid to rational chemical, excipient or solvent use we have included a number of .hsd tables in HSPiP. The
first is a list of HSP for all the GRAS (Generally Regarded As Safe) where we could provide a meaningful
value. Here’s a snapshot from the start of the list within GRAS.hsd:

Figure 27-1 Some GRAS chemicals


A typical use for the table would be to find a rational substitute for a chemical that is not on the GRAS list. For
example, if your chemical were an acid with [16, 3, 8], then Caprylic Acid would be a good starting point for
substitution.
Another important list in Pharma is the Q3C(R3) list with Type 1 and Type 2 “bad” solvents and Type 3 and
Type 4 “good” solvents. Here we’ve made it particularly easy to work out a rational substitution of Types 1 or 2
with Types 3 or 4. We’ve provided Q3C(R3) 1 and 2.hsd which you can load into the main form. This lets you

203
view the key properties of these bad solvents. For example, if you wanted to get rid of chloroform from a
formulation, you have the properties ready for the next phase:

Figure 27-2 Using Q3C(R3) to replace a “bad” solvent


By highlighting Chloroform, then going to Solvent Optimizer into which you have loaded the Type 3 and 4
solvents as Q3C(R3) 3 and 4.sof you already have your target correctly set. Clicking on the “2” button gives
you Cumene/Ethanol as a good HSP match. Of course there is more to matching than just the HSP. The
Optimizer lets you rationally match for other properties such as Relative Evaporation Rate etc.
We’ve done something even more. Thanks to a generous HSPiP user we have a large list of EU cosmetic/food
approved chemicals. Via their CAS numbers Hiroshi was able to generate SMILES for a large number of the
chemicals (many others are ill-defined mixtures). He then used Y-MB to create the Food and Cosmetics.hsd
and .sof files. In searches for alternatives in formulations, this is a most helpful. Load Food and Cosmetics.sof
into the Optimizer and you will find a rich source of alternatives. Of course many of these aren’t solvents so
you have to use caution. But we think it’s important to give you such a large database to explore this rich
domain of “approved” chemicals. Please note, of course, that “approval” can change and many of these
chemicals come with restrictions. We provide it as a courtesy to our users, not as an infallible guide to food and
cosmetics best practice.
Although this is a tiny chapter, we think it’s a pointer to things to come. Solubility and permeability concepts in
Pharma and Cosmetics have, in our opinion, been dominated for too long by one-dimensional concepts such as
LogP, LogD and LogS. How much richer the discourse becomes when scientists can discuss issues in three
dimensions rather than one!
The launch in 2010 of Abbott & Wiechers’ FFE (Formulating For Efficacy) software specifically designed for
the cosmetics and pharma industries and firmly based on HSP has finally brought solubility thinking firmly into
the area. A review by Abbott, An integrated approach to optimizing skin delivery of cosmetic and
pharmaceutical actives, International Journal of Cosmetic Science, 34, (2012), 217-222 summarises the
approach.

204
Chapter 28 Exploring with HSP – (Generating and testing research
hypotheses)
Many users of HSPiP have specific problems they want to solve: dispersing a particular nanoparticle,
substituting a solvent for a particular polymer, finding an environmentally friendlier solvent blend, or perhaps
something more exotic.
But HSP can also be used by the adventurous explorer to map out uncharted territory. All that’s needed is a
bunch of HSP values, a vague hypothesis and some extra support software such as Excel.
Mission impossible – the insoluble polymer
Charles once had to find the best high-temperature solvent for a polymer that was insoluble in just about
everything at room temperature. By chance it was known that one horribly toxic solvent could just about
dissolve the polymer at ~150ºC – which was not a practical temperature. Knowing its HSP it was then possible
to look for other solvents with HSP in that sort of region – just by sorting with RED number. This give a short
list of possible candidates.
These were tested. Not surprisingly, given the single datapoint that started this process, about half the solvents
were worse than the toxic one, but some were better. From this small dataset a better HSP estimate could be
made and a few more solvents could be tested. This then gave a practical solvent that worked at ~120ºC.
That’s a simple example of exploration. The original hypothesis wasn’t brilliant. But it was a start. Without that
hypothesis the solvent screening process would have involved many more experiments and much more time
and expense. Now let’s look at some more complex explorations.
Screening with Distance maps
Suppose, (to take a specific example from the HSP user who inspired this chapter) you are interested in plant
chemicals. You have an idea that solubility plays a key part (“a necessary but not sufficient condition”) in
allowing a chemical to reach its target as, say, an anti-malarial. There are innumerable plant chemicals out there
and if you don’t have access to a pharma-grade high-throughput screening system, how do you narrow down
your choices?
The key is to have a method not of hitting winners (that’s asking too much) but excluding no-hopers. The
problem with screening is that there are far too many possible candidate molecules, so anything which has a
reasonable chance of excluding molecules that won’t work will be of great help.
Your starting point is a (small) list of chemicals that are known to work. If they have very different HSP then
you need not bother to continue. But if (as is often the case) they are clustered near one region of HSP space
then you can create a “target” value from this cluster (e.g. using a Sphere, taking an average, choosing your
favourite), ready for calculating the distances from this target of the pre-screen molecules. An example of this
approach has already been described in the DNA chapter with the cytotoxic chemicals.
So assemble your list of plant chemicals as a list of Names and SMILES and save it as a simple tab-separared
.txt file – Excel does this for you no problem. Then find the File Convert option in Y-MB. If you have a lot of
molecules and many of them are large you may want to go and have a cup of coffee while the computer does all
the work. At the end of the process you have a .hsd file with estimated HSP values for most of the chemicals
you presented. Most? It’s likely that any list of SMILES will have a few problem chemicals which Y-MB can’t
handle. But if you drag the .hsd file into Excel it’s smart enough to recognise that it’s a tab-separated format
dataset and you’ll get a nice table. Search for the word “error” using Excel and you’ll find the failed molecules.
Simply delete their rows – you’ve got more than enough chemicals to screen in any case – or find the correct
SMILES, convert them manually in Y-MB and Paste the values into Excel.
Now create a fresh line at the top which contains the target value which you think (or hypothesise) is a fair
representation of the class. Then it’s easy to calculate the HSP distance of each molecule from your target
D=sqrt(4*(δDt-δDi)2 + (δPt-δPi)2 + (δHt-δHi)2) where t=target and i=the i’th chemical. Excel can then sort by
the distance column and you can decide a cut-off value for screening purposes – rejecting anything with a
distance greater than that value.

205
Figure 28-1 A distance map in Excel, sorted to show the chemicals closest to the target
In this dummy example (of course there are more chemicals not included in the screen shot) I might decide that
anything less than a distance of 4 is acceptable so would only test DDAIP to Octyl Salicylate.
If you want to be more sophisticated you can reason that anything with a MVol > 400 (or whatever value you
choose) will be too slow to penetrate the target or too hydrophobic (see the chapter describing Ruelle’s
solubility calculations) so you can do a sub-sort and remove those molecules that are too large. For this
example I sub-sorted the <4 area by MVol, then resorted the area with MVol<300 to give my final shortlist:

Figure 28-2 A smaller list by excluding high MVol chemicals


If you’ve already got your list of pre-screening molecules in SMILES format then this whole exploration will
have taken less than a few hours. What will you have achieved? At the very least, the molecules that pass your
distance screen will be (because of their HSP match) compatible with the formulation vehicle of your target
active. If you’re an optimist then there’s also a good chance that these molecules will partition into the correct
part of the cell or organism (because of the HSP match) and have a chance of working.
When we’ve tried such explorations we find them to be much more insightful than methods typically used in
pharma: Lipinski’s Rule of 5 or LogP. No simple method will ever be perfect for identifying candidates but we
believe that the HSP distance method has a lot going for it.
Finally, remember that there’s more to efficacy than solubility! It’s up to you to sub-screen the molecules for
the right sort of chemical functionality. An example can be drawn from the chemotherapy drugs mentioned
above. In the specific case of carboplatin, it is the organic segment of the drug that has the proper HSP, and not
the platinum, but the drug can orient such that the platinum is hidden within a kind of micelle, thus allowing the
desired effect.
SOMs

206
Very often the field of exploration isn’t a simple yes/no – good drug/ bad drug. There may be a set of
characteristics which you need to map to find if there is a link with the 3D space of HSP. Those who are skilled
in Principle Component Analysis can readily take a large HSP dataset and try to map the δD, δP, δH and
(usually) MVol values against the factors of interest. But we find that SOMs, Self Organising Maps, provide a
better view of complex terrain. They attempt to put “like-with-like” within a 2D space and the hope is that there
will be clear-cut regions on the map that distinguish one type of behaviour from another. If you are not familiar
with SOMs then the Wikipedia article is a good introduction.
A simple example is the Sphere technique itself. On a SOM a good Sphere fit is equivalent to a (roughly)
circular area of “good” molecules all separated from the rest of space which contains the “bad” molecules. In
Hiroshi’s work on the new fitting regimes (Double-Sphere, Data…) he found that SOMs were very helpful in
revealing problems with the datasets.
Here is an attempt to make the data fit a double sphere SOM. The lines are “guides for the eye” and not part of
the SOM software output:

Figure 28-3 A double-sphere SOM


Within this eBook we’ve already mentioned SOMs in terms of fragrance mapping. Such maps aren’t definitive
about cause and effect, but they are suggestive of hypotheses which can be further explored.
Fitting
Another example from this eBook illustrates another way to explore. We were interested in the partition
coefficient between soil and water, Koc. It seemed reasonable to guess that the HSP distance between the
chemical and some hypothetical soil would correlate with the partition coefficient. But what is the HSP of soil?
The answer was to take a set of known Koc values, make a guess at the HSP of soil, calculate the distances from
each chemical and that soil and, by adding some further coefficients, predict Koc from the distance. The square
of the errors between predicted and actual values could then be summed – showing that the original guess, not
surprisingly, was hopelessly wrong.
That’s when Excel comes in. Ask its Solver to minimise the error sum by varying the coefficients and the HSP
of soil and see what happens. If the fit is still bad then the hypothesis is useless. If the fit is good then you have
an effective working HSP for soil and can play around with some extra parameters (MVol or MWt usually have
a part to play) to find the best fit with the least adjustable values.
If you have more sophisticated fitting algorithms then you can automatically find complex relationships which
will give even better fits.
207
Once you’ve got some good fits it’s time to pause. Do the fits make sense in terms of basic chemistry and
thermodynamics? Have you just found some meaningless relationship by providing too many fitting
parameters, or does the relationship suggest some interesting science? Do the fits from the test data make good
predictions for data not included in the fitting set?
The answers to those questions will vary. Sometimes the fits really are artificial and tell you nothing.
Sometimes they fit so beautifully to standard HSP theory that you can be pretty sure that they are sound. But
remember that the fits aren’t an end in themselves – they are a means to deeper understanding and, perhaps,
useful predictions.
To boldly go
With the tools provided by HSPiP and with the extra tools in Excel (and other programs) it’s not hard to
explore. Often, as with real explorers, the results are failures. But when the risks are small and the rewards are
significant, maybe it’s worth having a go. We’ve used HSPiP’s tools for our own explorations and so made
them as easy as possible to use. We hope you will want to use them for your own journeys into the unknown.
Power Tools and Thrill Seeker Tools
To make it even easier to explore new territory, we have provided Power Tools, with the expectation that their
range and performance will steadily expand. For those who can’t wait to explore, Hiroshi has created Thrill
Seeker tools on his Pirika website. These are not necessarily as debugged, stable, convenient as Power Tools –
and they might even be based on ideas that, in the long run, don’t work out. But some Power Tools were once
Thrill Seeker tools and have benefitted from the insights of those who chose to work with the early versions and
provided helpful feedback about good and bad points.

208
Chapter 29 Liquid extraction – a work in progress
Extracting a chemical from water into a solvent (or the other way round) would be much simpler if you could
easily calculate partition ratios between the two liquids for the desired chemicals (the “good stuff”) and the
undesired chemicals (the “bulk”) that are usually around to make life difficult.
There are two ways to look at the problem. The first has the merits of simplicity, clarity and fitting in with the
general approach of HSPiP. The second is theoretically interesting but shows some limitations to the first
approach which we are doing our best to understand.
The simple approach
Let’s assume that you have a bulk material containing your chemical of interest. All simple processing (e.g.
filtration) has taken place with the bulk and any obvious way to extract the good stuff via a simple process (e.g.
crystallisation from a simple solvent) has been excluded.
So you have two problems. The first is to dissolve the good stuff. The second is to not dissolve the unwanted
stuff.
From the HSP of the good stuff you will quickly identify a set of good solvents from their close HSP match –
the Solvent Optimizer will do that nicely. The real difficulty is with the bulk. Here the recommended way
forward is to do a classic 20-test tube Sphere experiment to get an approximate centre and radius for the bulk.
Then you need to select an extraction solvent that is sufficiently near the edge of the Sphere yet sufficiently
close to the good stuff to be effective.
One subtle variation of this goes back to the old “problem” of Sphere radii. Suppose you want to extract from a
polymer. If you do a Sphere test based on solubility you will get a small radius. If you do the test based on
swellability you will get a large radius. [In practice you would just do one test but score “soluble” as 1 and
“swellable” as 2 and calculate the Sphere with a 1 then a 2 as the criterion]. You can now choose your
extraction solvent as one that is both reasonably good for your target compound and also in the swellability
portion of the Sphere so it has a chance to get inside the polymeric mass to do its extraction.
If you find that there is insufficient distance between the bulk and the good stuff then it’s highly likely that
liquid extraction isn’t a great method for this particular mix.
Of course you might be lucky. If you are doing water/solvent extraction it might be that the bulk material is
specifically biased towards the water, in which case your choice of solvent will be based on HSP distance,
water immiscibility and then cost/safety/environment considerations.
The point of this section is that although HSP may not provide all the answers they do provide a rapid means of
assessing probabilities of success or failure.
The theoretical approch
The calculation of liquid/liquid partition ratios would, from simple theory, depend on MVol and HSP Distance2,
i.e. the chi parameter. So what’s stopping us from providing a liquid extraction modeller to HSPiP?
There are two key problems here. The first is that the water can sometimes contain some solvent and the solvent
can contain some water. The classic example is octanol/water where the water contains essentially no octanol
but the octanol contains 20% (molar) water. The second is that water itself is so odd. As discussed elsewhere in
this eBook, the solubility of many molecules in water is influenced less by HSP distance and more by the
“hydrophobic effect” which is mostly a size-related issue, so that larger molecules are generally much less
soluble in water unless they contain plenty of alcohol, amine or acid functionality or are salts.
Because octanol/water is the most familiar partition coefficient, it’s worth plotting the (non) correlation
between LogKD and HSP distance from octanol:

209
Figure 29-1 No correlation between Octanol distance and Octanol/Water partition
As is well-known, the single best predictor for LogKD is molar volume, and here is the proof:

Figure 29-2 A good (and well-known) correlation of Octanol/Water with MVol


So much is well-known. What is less well-known is that there are some intriguing correlations of water/solvent
partition coefficients with HSP. Here are the data for hexane:

Figure 29-3 Two correlations seem to be present in the Hexane/Water partition coefficient

210
The two correlation lines in there need urgent explanation, as do those very scattered points that are clearly
outside any correlation.
The upper line is for alcohols and amides:

Figure 29-4 The upper line is a good plot for alcohols. The lower line (not shown) is from carboxylic acids, esters and ketones
The lower line comes from carboxylic acids, esters and ketones. The scatter comes from amines (which are
highly chaotic) and then (mostly) from compounds such as vanillin or hydroquinone with hydroxyl groups plus
other functionality. We can perhaps excuse the amines as their state of protonation will make a large difference
in ratios, and small amounts of spare acid or base in the system can make large changes to the ratio. This leaves
us with the hydroxy-containing molecules and the problem of alcohols and amides.
The Ruelle method for calculating the hydrophobic effect seems to offer a good explanation for the
hydroxyl/alcohol/amide exceptions. Their (in)compatibility with water is modified by their ability to hydrogen
bond with the water molecules and therefore reduce the hydrophobic effect. They are therefore more soluble
than their HSP distance would suggest. The problem arises in multi-functional hydroxyl molecules. If, for
example, they have strong intra-molecular hydrogen bonds, they will appear more like non-hydroxy molecules
to the water so will be less soluble than a similar molecule (e.g. an m-hydroxy instead of an o-hydroxy) which
has no intra-molecular hydrogen bonds.
Solvents such as carbon tetrachloride and benzene show similar tendencies to hexane, so predicting liquid-
liquid extraction should be possible if HSP distance and a Ruelle correction term are included in an overall
formula. But chloroform and diethyl ether both show a different set of non-conformities. We are currently
investigating possible root-causes for the problems with these solvents. A potential problem is induced dipoles
as witnesses by variable dipole moments depending on the local environment (see pp. 16-17 in the handbook).
So of the 6 solvents for which we have adequate test data:
• 3 (hexane, carbon tetrachloride and benzene) look as though there’s a possible HSP-based formula if
Ruelle hydrophobic corrections can be included.
• 1 (octanol) is mostly fit with a pure hydrophobic calculation (MVol)
• 2 (chloroform, ether) are not yet explained.
This is not a good starting point for those who want a generalised model for liquid-liquid extraction with other
solvents!
Almost certainly the deviations from HSP predictions would be a lot less if neither liquid is water. At the time
of writing we don’t have a dataset for such systems with which to test the model.
As this is such an important area, the HSPiP development team will continue to explore options to create a
robust modeller to add to the software. If we do, then this chapter will be significantly revised!

211
Chapter 30 The HSP of gel formation
We are all familiar with adding polymers to low viscosity solutions in order to create a gel. Poly(ethyleneoxide)
added to water does just this. HSP considerations can tell you which polymers would be most suitable, and
entanglement theory tells you that you have a trade-off between adding large amounts of low MWt polymer or
small amounts of high MWt polymer.
In the Handbook (pp271-273) a different type of polymer gelator concept is described. An alkyd paint polymer
is nicely soluble in mineral spirits (white spirit) and, because it is not yet cross-linked the viscosity is low – and
this can lead to dripping and spattering on application. The same alkyd (a vegetable oil modified polyester) but
with a grafted Versamid block forms a gel. The gel is easily disrupted by stirring, brushing, or spraying – the
paint is a thixotrope. The gelation does not come from conventional polymer entanglement. Instead, the
Versamid block is insoluble in mineral spirits so it self-associates to form a loose network, strong enough to gel
but easily disrupted by mild shear forces.

Figure 30-1 HSP relations for establishing thixotropy in an alkyd-type paint. The solid circle represents the solubility of the
alkyd (A) and the dotted circle that of the Versamid (B). The Versamid segments associate because they are not soluble in the
mineral spirits. Addition of n-butanol destroys the thixotropic effect, since the solvent then becomes too good for the
associating Versamid segments and they no longer associate, being truly soluble.
To show that the thickening is not due to classic polymer effects, if some n-butanol is added, bringing the
Versamid block into true solution, the viscosity greatly reduces and the gel effect is lost.
As another example, similar HSP concepts have been used to carefully balance the formulation of wet-on-wet
flexographic printing inks to allow gel formation between the applying the successive colours. This
development is still in its early stages, but promises to transform the flexographic printing industry into a
greener one.
There is a third way to make gels which uses low MWt additives. These are called LMWGs (Low Molecular
Weight Gelators) or LMOGs (Low Molecular-Mass Organic Gelators). Just 1% of an organic gelator or 0.1%
of an organic supergelator can create an impressively strong gel. A delightful paper A. Vidyasagar, K. Handore,
and K. M. Sureshan, Soft Optical Devices from Self-Healing Gels Formed by Oil and Sugar-Based
Organogelators, Angewandte Chemie, July 2011, gives you a good idea of the effect:
212
Figure 30-2 Gels formed in pump oil or silicone oil with ~0.3% of organic gelators. Typical dimensions in the images are 2-
5cm. For explanations of shapes and colours, please read the original paper.
They are prepared by heating the gelator in the solvent then cooling to the gel phase. Such gels can be used in
many applications such as optics, printing, cosmetics, coating etc.
The way they work is that the gelators have a tendency to self-aggregate into long chains and networks –
forming a virtual polymer.
So how does one choose the right gelator for their particular system? It seems beyond the state of the art to be
able to predict ab initio that molecule A will gelate system B. Instead one has to look in the literature for
molecules that happen to gelate systems based on B. If you are lucky you will find a molecule that is cheap,
safe, effective and perfect for B. But suppose, for example, that you want to gelate a cosmetic formulation
based on isopropyl myristate (IPM) and you can’t find any literature references to gelators for IPM. How do
you then choose a gelator? You look for solvents that have been successfully gelled by a given gelator and that
are also “close” IPM, but how “close” does it have to be?
The key problem is obvious. By definition the gelators like to self-associate. So a poor solvent for the gelator
will not break up the self-association and the gelator will be simply insoluble. At the other extreme, really good
solvents will be so good that the gelator doesn’t self-associate so you simply have a solution.
The problem is solved, therefore, by finding areas of marginal solubility – not too little and not too much. And
how would you find this? HSP seem to be the natural way to approach the problem.
We are most grateful to Matthieu Raynal and Laurent Bouteiller for giving us permission to use their wonderful
paper Organogel formation rationalized by Hansen solubility parameters, Chem. Commun., 2011, 47, 8271–
8273 to show you how it’s done. The datasets are taken, with their kind permission, from the Supplementary
Data to their paper.
They did a simple HSP scoring of 0 and 1 for insoluble/soluble and, separately, for no-gel/gel. You therefore
get two HSP values, one is the classic “solubility” and the other is the “gelator” value.
Knowing these values for the range of 8 gelators that span a wide range of structures it then becomes possible
to find which one has a gelator sphere suited, for, say, IPM.

213
Figure 30-3 The 8 gelators used in the study. We will focus on 1 & 2 and 7 as they represent opposite ends of the gelator
spectrum - (1) involves H-bonding, (7) involves π-stacking.
For these gelators, the best fits were obtained with GA Classic mode. Here are the results:

214
Figure 30-4 The solubility sphere for Gelator 1, using a GA fit
The gel sphere, not surprisingly, is much smaller and in a very different region:

Figure 30-5 The gel sphere for Gelator 1


The HSP distance from this gelator sphere to IPM ~[16, 4, 4] is greater than the radius of 4 so this gelator will
probably not be useful.
For Gelator 7, which relies on aromatic pairing for gelation rather than –OH bonds, the spheres are very
different.

215
Figure 30-6 The solubility sphere for Gelator 7

Figure 30-7 The gel sphere for Gelator 7


Again, this is unlikely to be a gelator for IPM
Finally, when we try Gelator 2 we get a solubility sphere (not shown) at [18.3, 13.8,10.1] and a gel sphere
which is close enough to IPM that it might well be a reasonable gelator for it:

216
Figure 30-8 The gel sphere for Gelator 2
As the authors point out in their paper, not all the fits are perfect, and some of the datasets need careful
interpretation as gelation behaviour is complex with time-dependent effects that can make what looks like a gel
become a precipitate after a time.
Another suggestion from the authors is to create double-sphere fits. The reason is that although two different
solvents can both be gelators, the type of gel (the network fibre structure) can be different.
The double-sphere option in the GA mode is ideal for this. As you can see in the above fit for Gelator 2, some
of the gel solvents are outside the fitted sphere. Here are the same data fitted with the double-sphere option.

Figure 30-9 A double-sphere “fit” to Gelator 2. Who knows if these values are meaningful?
In this example it is not clear whether the spheres are meaningful. Perhaps it would be better to do a sub-
analysis of the different types of gel (if they exist for this particular molecule) and fit them separately.

217
The point of this brief chapter is not to draw definitive conclusions about using HSPiP with gelator data but to
show that this sort of analysis may well be highly fruitful for further exploration.

218
Chapter 31 Going nano (HSP Characterizations of Nanoparticles)
It’s obligatory to have a nano chapter because nano is new and exciting. In fact the truth is that nano is old and
HSP have been solving nano issues for decades. We know this because the chapter on “insoluble” HSP is all
about carbon black and carbon black has been nano ever since it first appeared as smoke.
But this chapter is a reminder that those working at the cutting edge of science could do well to remember
older, simpler principles. We start with that great symbol of nano-modernity, the carbon nanotube, CNT.
Note. The following section was written before the work of Detriche et al (see below) was published. It is
interesting to compare our predictions with those of the Detriche paper.
Although there are some papers that specifically invoke HSP for understanding the best solvents for dispersing
CNT, we have found the data to be rather unsatisfactory. First, these papers don’t have a sufficiently full range
of solvents to define the sphere in 3D. Second, we think that the data for good dispersion is skewed by high
density or high viscosity liquids. Because the test for a “good” solvent is whether a dispersion remains stable
over time, a high density or viscosity can give a misleading result. Ch.7 of the Handbook introduced the
concept of RST – Relative Sedimentation Time – to help correct for differences in sedimentation due to
density/viscosity:
RST=ts(ρp- ρs)/ η
where ts is the actual sedimentation time, ρp and ρs are the densities of the particles and solvent and η is the
viscosity. The RST values should then be used to decide between “good” and “bad” solvents.
Happily a paper K.D. Ausman, R. Piner, O. Lourie, and R.S. Ruoff, Organic Solvent Dispersions of Single-
Walled Carbon Nanotubes: Toward Solutions of Pristine Nanotubes, J. Phys. Chem. B. 104, (38), 8911-8915,
2000 gives a range of good solvents which, when combined with those known to be bad gives the following
excellent fit:

Figure 31-1 Fit of Ruoff’s CNT data


The values aren’t too far from the less reliable fit which gave a value in the [18, 10, 9] region. The Solvent
Optimizer readily informs us that DMF/Xylene (69/31) or Caprolactone/ Dipropylene Glycol Mono n-Butyl
Ether (67/33) would be mixtures superior to any of the individual solvents used in the above experiment.
The paper by Detriche et al Application of the Hansen Solubility Parameters Theory to Carbon Nanotubes, J.
Nanoscience and Nanotechnology, 8, 1-11, 2008 is an interesting vindication of the comments above. Although
the paper covers many different types of CNTs, the data below are specific to Single Wall Nanotubes, SWNT.

219
First, they confirmed that without thorough centrifugation of dispersed samples results are highly unreliable as
viscosity and density can skew the apparent solubility/dispersability of CNT. Their paper casts some doubt on
some of the “good” solvents used in the fit above, which may partially explain differences between the two
studies in the calculated values for CNTs.
Second, they showed that mixtures of two good solvents can give better solubility/dispersability. So a 50/50
mixture of o-dichlorobenzene/benzaldehyde gave a higher (8x) solubility than either alone.
δD δP δH Ra
SWNT 19.4 6.0 4.5 Radius=3
o-Dichlorobenzene 19.2 6.3 3.3 1.30
Benzaldehyde 19.4 7.4 5.3 1.61
50/50 mix 19.3 6.9 4.3 0.9

Third, and an excellent vindication of HSP principles, is the fact that a 75/25 mixture of two non-solvents gave
a reasonable solubility
δD δP δH Ra
SWNT 19.4 6.0 4.5 Radius=3
Diethyl phthalate 17.6 9.6 4.5 5.1
Methyl naphthalene 20.6 0.8 4.7 5.7
75/25 mix 19.3 6.9 4.3 3.1

With thanks to Dr Detriche who gave us permission to use his very extensive full dataset, here is a revised look
at the HSP of SWNT:

Figure 31-2 Fit of Detriche CNT data

220
Fifth, they recognised a further HSP principle – that larger polymers have smaller radii. They were able to
fractionate CNTs using this principle – the average CNT size in a “good” solvent was considerably smaller than
the size in a “very good” solvent.
Sixth, though it hardly needs pointing out, Hildebrand parameters proved a poor way of predicting solubility
behaviour.
Surfactants are also a popular method for obtaining good dispersions of CNT in water. So far we have not been
able to find a good correlation between the HSP of the surfactants and their dispersion capabilities for the CNT.
The standard hydrophile/hydrophobe chains in these surfactants simply don’t seem to have the capability of
producing HSP with the required high δP and δH.
Two papers from the Coleman group explicitly use HSP to further push the boundaries of CNT and graphene
solubilities. Shane D. Bergin, Zhenyu Sun, David Rickard, Philip V. Streich, James P. Hamilton, and Jonathan
N. Coleman, Multicomponent Solubility Parameters for Single-Walled Carbon Nanotube Solvent Mixtures,
ACNano, 3, 2009, 2340-2350 and Yenny Hernandez, Mustafa Lotya, David Rickard, Shane D. Bergin, and
Jonathan N. Coleman, Measurement of Multicomponent Solubility Parameters for Graphene Facilitates Solvent
Discovery, Langmuir, DOI: 10.1021/la903188a. The HSP analysis (the authors used HSPiP during their
research) is not perfect and doesn’t explain everything, but such huge “molecules” really are pushing the
boundaries of HSP. Nonetheless, it’s clear that HSP do a quite impressive job in helping understand these
complex phenomena.
Similarly, recent work from the Delhalle group throw more insight into CNT surfaces, especially the fact that
many “functionalised” CNT are very poor mixtures of undefined quality. See Detriche, S., Nagy, J.B.,
Mekhalif, Z., Delhalle, J, Surface State of Carbon Nanotubes and Hansen Solubility Parameters, Journal of
Nanoscience and Nanotechnology, 9, 2009 , 6015-6025.
C60
We can’t miss the chance to discuss C60. There have been numerous attempts to understand the slightly odd
solution behaviour of C60. Not many other chemicals are best dissolved in chloro- or phenyl-naphthalene at
room temperature. Many efforts using QSAR techniques sifted from 1000’s of possible “molecular” properties
have produced correlations with R² from 0.6 (poor) to 0.9 (not bad), but with little apparent insight into what is
really going on. Hansen and Smith used basic Sphere techniques in their paper C.M. Hansen, A.L. Smith, Using
Hansen solubility parameters to correlate solubility of C60 fullerene in organic solvents and in polymers,
Carbon, 42 (2004) 1591–1597 to find HSP of [19.7, 2.9, 2.7] and thereby to make predictions of 55 solvents
that might do a reasonable job and then went on to show which polymers (polystyrene is a good example)
would have good compatibility with C60.
We’ve taken advantage of a rather expanded data set to re-do the correlation. Here is the Sphere with 107
solvents instead of the original 87. The definition of good “1” solvents is Log(MoleFraction)>-3.0.

221
Figure 31-3 Fit of Hansen C60 data
The result is [20, 3.6, 2.0], with a radius of 4.6 which is close to the original fit. If the definition of “good”
solvents is changed to include all the “2” solvents, defined as Log(MoleFraction)>-4 the value changes to [1.2,
2.6, 0.5], only a modest change, though with a radius of 8.8.
Because there is such a tradition of doing least-squares fits of the Log(MoleFraction) data, we did the same,
using the formula:
Log(MoleFraction) = K * Sqrt(4*(δDc60-δDs)2 + (δPc60-δPs)2 + (δHc60-δHs)2)

-9 -8 -7 -6 -5 -4 -3 -2 -1 -1

-3

-5

-7
y = 1.062x + 0.1763
R2 = 0.8008
-9

-11

Figure 31-4 Least squares fit of Hansen C60 data


The R² is a respectable 0.8 when the C60 HSP are set to [22.5, 0.6, 2.9], remarkably close to the Sphere fit.
The C60 QSAR allows a fit to Distance with the best being around [20, 4, 2], provided that MVol was included
as one of the parameters.

222
Of course there’s more to solubility of a molecule like C60 than enthalpy. There must surely be some entropic
effects and, presumably, some specific inter-molecular effects. But straightforward HSP do a remarkably good
job at covering this large range of solubilities and, importantly, provide practical predictions on solvents and
polymer compatibility that the working scientist can combine with intuition and experience to help develop
processes for C60. The fit is close to the best published fits via QSAR without the need for those 1000+ factors
available to practitioners of that art. The bottom line is that C60 is in an awkward position in solubility space –
not many liquids have δD values in the 20+ range and even less combine that with low δP and δH values.
Chemicals with such high δD values are (usually) solids. It’s the δD which makes it so hard to find convenient
solvents for processing C60, it’s as simple as that.
And for those who wish to test out a prediction. From published data of cohesive energy density, sulphur comes
out with a value in this high δD range so it’s probably a good solvent if you wish to try it.
Graphene too
The amazing properties of graphene hold great promise for many applications. Geim’s initial method for
producing it using adhesive tape is breathtakingly simple and inspired but not adequate for mass production.
The paper by Coleman and his large team, High-yield production of graphene by liquid-phase exfoliation of
graphite, Nature Nanotechnology, 3, 563-568, 2008, shows that graphene is soluble in a solvent as simple as
Benzyl Benzoate and therefore potentially graphene coatings can be produced direct from solvent. We’ve
plotted their data (with their permission) in HSP terms and obtained the following value for the HSP of
graphene:

223
Figure 31-5 Fit of Coleman’s Graphene data
When [20, 11.2, 7.3] is put into Solvent Optimizer, a near-perfect match is obtained by a 60:40 blend of
Caprolactone and Benzyl Benzoate. It will be interesting to know if this blend is actually better than Benzyl
Benzoate alone.
Nano-clays
Clays are remarkably cheap and, with a bit of exfoliation, are remarkably nano. A lot of people have therefore
spent a lot of time trying to make polymer/clay nanocomposites. With hindsight it is clear that a lot of this work
has been wasted because, first, exfoliating the clays is not easy and, second, it is not obvious which organic
groups would be most compatible with a given polymer.
The best-known method for aiding exfoliation is to create an organoclay via ion exchange to remove the
sodium ions between the plates and replace them with a quaternary ammonium salt, typically containing a
mixture of methyl, benzyl, hydroxyethyl and tallow groups. If this is done well then the resulting clay contains
neither excess sodium nor excess quaternary salt. But doing things well makes the clays more expensive. But
by using impure versions, interpreting the data is very difficult.
Assuming the organoclay is of high quality, what is the best one to use for any given application? Usually the
ideal is total exfoliation of the clay within the polymer. But many users are happy if they have lots of “tactoids”
(nano clusters of clay particles) which are, at the very least, better than mechanically dispersed clay
microparticles.
At this stage it would be good to show that HSP can come to the rescue.
Unfortunately, the data don’t allow the production of good HSP. The file Clay4 fits the data from D.L. Ho and
C.J. Glinka, Effects of Solvent Solubility Parameters on Organoclay Dispersions, Chem. Mater. 2003, 15, 1309-
1312. The clay is dimethyl-ditallow montmorillonite (Cloisite 15A) and gives an HSP set of [18.2, 3.8, 1.7].
Unfortunately, attempts to fit (nominally) the same clay from the data of D. Burgentzlé et al, Solvent-based
nanocomposite coatings I. Dispersion of organophilic montmorillonite in organic solvents, Journal of Colloid
and Interface Science 278 (2004) 26–39, shown in Clay2, gives the impossible values of [16.8, -4.7, -3.3]. The
problem is compounded by the fact that the solvent data contains its own uncertainties. Does one class as
“good” solvents those that swell the clay or those that cause a big increase in the interlayer spacing?
Furthermore, one of the really good solvents in the first paper, chloroform, is a bad solvent in the second.
From the second paper, the dimethyl-benzyl-tallow (Clay1 – Cloisite 10A) gives [20.4, 6.6, 5.9]

224
Figure 31-6 Clay 1 fit using GA mode
and the methyl-di(hydroxyethyl)-tallow (Clay3 – Cloisite 30B) gives [15.8, 15.2, 11.0]. Adding n-alcohol data
from another paper on the 30B gives [16.7, 10.4, 10.2], though there is a contradiction with the data point for
ethanol.
Because organoclay nanocomposites look to be of such great importance it would seem a good idea to re-test
the solvent swelling data on a group of well-defined clays, using a larger range of solvents across HSP space to
gain a better set of values or, conversely, to show that for some reason the HSP approach is not appropriate.
Nevertheless, when one of us (Abbott) tried applying the HSP data to a group of papers on organoclays in
poly(lactic acid) [18.6, 9.9, 6.0], it became obvious that the popular 30B was less likely to be a good match than
the 10A, whilst the also much-used 15A was likely to be unsatisfactory. The revised data on 30B reduced the
degree of mismatch with the poly(lactic acid) – re-emphasising the need for a definitive data set on these clays.
Recent work by a team led by Dr Andrew Burgess (then in ICI, now in Akzo Nobel) gives some visual
elements to this story. We are grateful to Dr Burgess for permission to use their material here. They used a set
of Cloisite clays and attempted to disperse them in a range of solvents. Typical results for four solvents are
shown:

225
Figure 31-7 Some of the data for Cloisite clay dispersions from A. Burgess, D. Kint, F. Salhi, G. Seeley, M. Gio-Batta and S.
Rogers, reproduced with permission.

For example, it’s clear that chloroform is good at dispersing/swelling 10A, 25A and 15A, whilst THF is only
good for 10A and 15A. i-Hexane and acetone don’t do a good job with any of the clays. From their full dataset
they tried an analysis using Hildebrand solubility parameters. The results were unconvincing. The same data
put into HSPiP allowed a more insightful analysis. Here, for example, are the data for the Cloisite 10A:

226
Figure 31-8 A fit for the Cloisite 10A results from Burgess et al.
As the files are provided with HSPiP you can judge for yourself how good or bad the fits are. A larger range of
solvents would, as always, provide greater certainty. But a retrospective analysis of the Burgess’ team’s
attempts to combine the clays with various common acrylates showed that the HSP were a good indication of
the relative ease or difficulty of making stable clay dispersions.
Quantum dots
When a particle of something as ordinary as CdTe becomes smaller than ~10nm then its electronic properties
are governed by the wave function that can fit inside the dot rather than the properties of the material itself. So
CdTe can become green, red or blue depending on the particle size. There are numerous applications for such
quantum dots. But because small particles have large relative surface areas they tend to clump together, losing
their quantum-dot nature and/or their ability to be dispersed in the medium of choice.
Because there are so many different quantum dots, stabilised by a large variety of different methods, there
seems to be no general theme emerging for which HSP give an over-arching insight. However, one data set
kindly provided to us by Michael Schreuder working in Professor Rosenthal’s group at Vanderbilt University
shows a mixture of the expected and the unexpected. CdSe nanocrystals were stabilised with a shell of a
substituted phosphonic acid. Here is a typical example of a fit with 43 solvents for the Butylphosphonic acid
system:

Figure 31-9 The fit to a Quantum Dot


The calculated values [17.0, 4.0, 1.5, 6.6] seem reasonable for a somewhat polar butyl chain. The problem is
that when one goes to the phenyl phosphonate, the values are remarkably similar. The fit of [16, 4.7, 2, 5.3] has
a disturbingly low δD value. The fit for the Octylphosphonic acid version [17, 3.7, 1.5, 6.8] does not show the
expected decrease in δP and δH for the longer alkyl chain. And, surprisingly, the 2-Carboxyethylphosphonic
acid fit [16.4, 4.8, 3.2, 4.7] shows no evidence for the expected higher δP and δH. Even worse, some of the fits
(not included, for reasons we’ll describe in a moment) were very poor quality.
But maybe we are jumping too quickly to conclusions. We’re assuming that the CdSe surface is entirely
covered by a shell of substituted phosphonic acids, with the chains sticking out into the solvent, so the HSP
should be that of the chains. But what if some of the CdSe, or the phosphonate group is accessible to the solvent
– how much would that contribute to the HSP? Conversely, what would happen if there were still an excess of
the phosphonate – that would give strange results. The investigators checked out this last possibility. The
samples that gave poor results were checked using Rutherford Backscattering and were found to contain excess
phosphonate. At the time of writing, the reason for the relatively uniform HSP for the range of phosphonates

227
has not been found, but it is satisfying to note that when some poor fits suggested either that the HSP approach
was wrong or that the samples themselves had issues, it was the latter that was found. This does not prove that
HSP are right, but once again it shows that they can be deeply insightful even down to the level of quantum
dots.

228
Chapter 32 DIY HSP (Methods to Calculate/Estimate Your Own HSP)
Life would be very easy if we had HSP of any chemical of interest to us. But as the number of published HSP is
likely to be less than 10,000 and as there are literally millions of chemicals of interest the chances are small that
you will find the numbers for your specific chemicals, though the 1,200+ chemicals we provide with HSPiP are
a very good start.
So it would be very nice if there were a universally validated method for calculating HSP to a reasonable
degree of accuracy. Unfortunately, some of the methods require knowledge of other values such as enthalpy of
vaporization or dipole moment and you may not know either or both of those.
Below we describe the old techniques such as Hoy and Van Krevelen plus the newer technique of Stefanis-
Panayiotou. However, we strongly recommend that users only use the Y-MB method that has been developed
over the 2nd-5th editions by Dr YAMAMOTO Hiroshi as a core member of the HSPiP team.
The most basic calculation is of δTot. This is simple:
δTot = (Energy of Vaporization/Molar Volume) ½
But where do you find your energy (enthalpy – RT) and your molar volume? There are extensive tables of
enthalpy values available at a price. Any modern molecular mechanics program can do a reasonable job of
calculating molar volume and there are also free on-line tools.
So you might be lucky and be able to calculate δTot.
δP has been shown to be reasonably approximated by the simple Beerbower formula which requires just one
unknown, the dipole moment:
δP=37.4 * Dipole Moment/MVol ½
The more complex Böttcher equation (see equation 10.25 in the Handbook) requires you to know the dielectric
constant and refractive index in addition to the dipole moment. It may arguably give better values if you have
accurate values for all the inputs, but it is unlikely that you have those inputs so you are no better off.
The correlation has been re-done using the updated HSP list which, in turn was updated on the basis of the most
recent databases of dipole moments. There is a necessary circularity to this process but the aim is self-
consistency with all available experimental data so the process is highly constrained. The new fit, based on 633
values is shown in the graph:

229
Dipole Moment Correlation y = 0.8909x + 0.1753
R2 = 0.8196
6

4
DM - Calc

0
0 1 2 3 4 5 6
DM - Exp

Figure 32-1 The Dipole Moment correlation


and the revised formula, which is used in the HSPiP software is:
δP=36.1 * Dipole Moment/MVol ½
The paper by D.M. Koenhen and C. A. Smolders, The Determination of Solubility Parameters of Solvents and
Polymers by Means of Correlations with Other Physical Quantities, Journal Of Applied Polymer Science 1975,
19, 1163-1179 does what the title suggests and finds not only an acceptable equivalent to Beerbower (they had
an alternative power dependency for MVol but our revised data confirmed that 0.5 is optimal) but also a simple
linear relationship between δD and refractive index. The coefficients shown here are our own fit to a more
extensive and revised data set of 540 data points:

RI Correlation y = 0.8351x + 0.238


R2 = 0.8407

1.8

1.7

1.6
RI - Calc

1.5

1.4

1.3

1.2

1.1
1.3 1.4 1.5 1.6 1.7 1.8
RI - Exp

230
Figure 32-2 The RI correlation

δD= (RI - 0.784) / 0.0395


Koenhen and Smolders also showed a strong correlation between δD2+δP2, MVol0.33 and surface tension. Using
498 data points with relevant surface tension data we found a correlation:
SurfTension=0.0146*(2.28*δD2 + δP2 + δH2)*MVol0.2

ST Correlation y = 0.678x + 9.3053


2
R = 0.8016

60

50

40
ST - Calc

30

20

10

0
0 10 20 30 40 50 60 70
ST - Exp

Figure 32-3 The Surface Tension correlation

When it comes to δH there is no obvious short-cut for calculating it from first principles using a few constants.
We therefore have to rely on group contribution methods. And because we can use such methods for those, we
might as well try to use them for δTot, δD and δP as well.
Group contributions
There is a long and distinguished history of breaking molecules down into a number of smaller sub-groups then
calculating a property by adding together numbers for each group, weighted by the numbers of such groups in
the molecule. There is an obvious trade-off in group contributions. It’s possible to define –CH2- as just one
group or as 2 groups (-CH2- in acyclic and in cyclic molecules) or many groups (-CH2- in acyclic, in 3-member
rings, in 4-member rings in 5-member rings etc. etc. etc.). The more subgroups used the more accurate, in
principle, the group contribution but the less likely that there is sufficient statistical data to calculate the fits
with any degree of reliability.
Over the years there has been a convergence on the so-called UNIFAC partition of groups – providing an
adequate balance between over-simplification and over-complication.
So to calculate the group contributions for D, P and H one “simply” divides a set of molecules with known HSP
into their individual groups then does a linear regression fit to the data at hand. In practice this is a lot of work
and only a few such fits exist for HSP.
Because δD comes from Van der Waals forces it is intuitively obvious that group contribution methods should
produce reasonable approximations. It doesn’t matter all that much where a C-Cl bond is as the more important
fact is that there is both a C and a Cl.
231
δP is obviously problematic. A molecule with two polar groups near one end is likely to be more polar than one
where those two groups are at opposite ends and tend to cancel out. It is hard for a group method to capture the
geometrical issues.
Similarly, it is obvious that molecules with two –OH groups in them might differ strikingly in the amount of
hydrogen bonding interactions between molecules depending on how much hydrogen bonding there is within
each molecule. So δH can never be accurately determined from group methods.
So no matter how hard you try, you can’t realistically expect always to get accurate δH and δP values from
group methods.
How does Hansen do it?
The overall goal is to divide the cohesion energy (Energy of Vaporization) into the three parameters discussed
above. One finds or estimates the latent heat of vaporization at 25°C and subtracts RT (395 cal/mol or 1650
J/mol).
The preferred method to find δD is to use one of the figures in Chapter 1 of the Handbook. These give the
dispersion energy of cohesion as a function of the molar volume. There are curves for different reduced
temperatures. Use of reduced temperatures is characteristic of a corresponding states theory, which means that
the HSP are based on corresponding states. The reduced temperature is 298.15/Tc. Tc is the critical temperature
that can be found for many (smaller) molecules, but not for the larger ones. This then requires estimation. The
Tc has been estimated by the Lydersen method as described in the Handbook using group contributions from
the table for this purpose. Tc is found by dividing the temperature at the normal boiling point by the constant
found from the Lydersen group contribution. One can then easily find δD with this energy and the molar
volume. When this preferred procedure is not possible one can compare with similar compounds. Remember
that δD increases with molecular size, especially for aliphatic molecules. This combined with the probable
incompatibility of group contributions with a corresponding states theory makes the accurate estimation of δD
especially difficult, especially for polymers.
δP is usually found with the Beerbower equation given in the above, or else by group contributions as reported
in the Handbook. If a dipole moment can be found for a closely related compound, its δP can be found with its
molar volume, and this can then be used to find a new group contribution value for use with the table in the
Handbook. This procedure is best when the whole procedure of finding HSP values is possible for the related
compound. The ultimate result is two new sets of HSP.
δH in the earliest work was found exclusively by difference. The polar and dispersion cohesion energies were
subtracted from the total cohesion energy to find that left over. This was then used to find δH. When things did
not add up properly comprises were made based on the multitude of experimental data that were generated in
the process of establishing the first values. Up to the point where the Panayiotou procedure came forth, the
usual method of estimating δH was with group contributions as given in the Handbook.
For the sake of historical record, note that the original values reported by Hansen in 1967 were expanded to a
total of about 240 by Beerbower using the Böttcher equation, his own equation, and the group contributions in
the tables in the Handbook. This set was then extended over the years by one method or another by Hansen to
arrive at the values found in the Handbook. The revision process will presumably continue, but the original
values from 1967 seem to be holding up well. The Hoy parameters are not compatible with the Hansen
parameters, particularly with respect to finding a dispersion parameter that is too low. The Van Krevelen
procedure also gave somewhat inconsistent values and did not have a wide selection of groups to use.
Experimental data were found where possible and practical, and adjustments made accordingly, but one must
do this with care, since what looks good in one correlation may totally ruin another.
Much better than nothing
So now you know the bad news about DIY HSP. There is currently no good way to be sure you have calculated
accurate values, for reasons which are fundamental. So, do we abandon hope?
Happily the answer is that by using all the available methods and combining them with your own scientific
understanding, it’s possible to get HSP that are fit for purpose. If the molecule is well outside the sphere it
doesn’t really matter how far outside it is. So it often doesn’t matter if δH is 12 or 14. It suffices for you to
know that it’s not 2 or 4.
232
And if you find that it’s critical to know if δH is 12 or 14 so that you can really refine the radius of the Sphere,
you can resort to good old-fashioned experiment to get the HSP for the one molecule that happens to be of
critical importance.
The 6 ways
In the program we offer you 6 ways to calculate HSP.
1 Via numbers. This lets you input enthalpy, molar volume, refractive index and dipole moment. You therefore
get δTot and δP. If you also enter an estimate for δD the program calculates δH. You can also correct for
temperature of calculation of enthalpy and see an estimation of the surface tension from your calculated
parameters.
2 Stefanis & Panayiotou. The most extensive and accurate published group-contribution method for all 4
values (δTot, δD, δP and δH) came from Panayiotou’s group in Aristotle University in Thessaloniki. The
Stefanis-Panayiotou method (E. Stefanis, C.Panayiotou, Prediction of Hansen Solubility Parameters with a
New Group-Contribution Method, International Journal of Thermophysics, 2008, 29 (2), 568-585) has
established itself as an important method. The extra feature of S-P is that it attempts to distinguish different
forms of similar groups by identifying 2nd-order groups which have their own parameters. If you want a rough
estimate, then keep things simple and ignore the 2nd-order groups. For more accuracy you must include the 2nd-
order groups. It can be difficult to know how to partition your molecules into these UNIFAC groups. Helpfully,
S-P provided an example of each type of 1st- and 2nd-order group to help you break down your molecule in the
correct manner.
A typical example is 1-Butanol which has 1 CH3- group, 3 –CH2- groups and one –OH group. If you enter
these (1st-order) values and press calculate you get values for δTot and then δD, δP, δH of 21.9 and [15.9, 5.9,
13.2] (c.f. [16, 5.7, 15.8]) respectively.
There is a further refinement. If you are confident that the molecule (for whatever reason) will tend to be of low
δP and/or δH, you can click the “Low” option and use group parameters tuned for these respective properties.
To help you with your intuition, if you attempt, for example, to use the Low H option for 1-Butanol you get a
warning because there is not (should not be!) a Low H fitting parameter for this molecule.
Prof Panayiotou went on to develop the very interesting Partial Solubility Parameter (PSP) approach based
(largely) on parameters first calculated using COSMOtherm. The latest version of PSP includes a new
definition of δD by removing some of the polarizable elements and putting them into a new δP. δH is split into
donor/acceptor and the “distance” calculations, a tough problem for Donor/Acceptor are much more complex
whilst at the same time being better rooted in classic thermodynamics. It will be very interesting to see if PSP
can be developed into a fully practical system so that its parameters can be calculated relatively simply for just
about any molecule – by those who have access to COSMOtherm.
3 Van Krevelen is the first to admit that his group method cannot give accurate results, for the reasons
discussed above. His particular contribution to the problem is to introduce a “symmetry” option. If there is one
plane of symmetry then the polar value is halved, with two planes it is quartered and with 3 planes both the
polar and hydrogen bonding values are set to zero. The one-plane choice, for example, would help distinguish
our two cases of C-Cl bonds discussed above.
4 Hoy used a more subtle form of calculation from his chosen groups and includes options similar to
Panayiotou’s secondary groups by taking into account various x-membered rings and some forms of isomerism.
Importantly, Hoy also attempts to make corrections for polymers. It’s intuitively obvious that, for example, the
polar effect of an isolated sub-unit would be rather different from the overall polar effect from the polymer
chain made up from those sub-units.
Hoy also helps with input to the numerical and Van Krevelen calculations by producing an approximate value
for the molar volume. This can’t be as accurate as a proper measurement from density and molecular weight or
from a molecular mechanics program, but it’s a useful aid if you can’t derive it from those sources.
Because Hoy’s method worked so well within the chemical company for which he worked, there was little
reason for him to explain his methodology in greater detail, so much remains unknown. Steven once met a
former colleague of Hoy who offered to arrange a meeting, but it never worked out so we may never fully know
the science behind the Hoy method.
233
Note that because the Hoy definitions of δD, δP and δH are different from the Hansen’s, you should never try to
mix the two schemes by, for example, taking an average of a Hoy and Hansen estimate.
5 Y-MB One of the issues with group methods is that they often can’t satisfactorily predict complex inter-group
interactions. YAMAMOTO Hiroshi therefore adapted his Neural Network (NN) methodology for fitting the full
HSP data set in such a way that inter-group interactions automatically get fitted by the relative strengths of the
neural interconnections. But of course this needed him to have a set of groups. He therefore devised an
automatic Molecule Breaking (MB) program that created sub-groups from any molecules. He used a general
MB technique that allowed him to experiment with which combination of MB and NN gave the best predictive
power for HSP. That’s what you get with HSPiP. And because the MB technique was general, he was able to
take standard molecular inputs (such as Smiles, MolFile (.mol and .mol2), PDB and XYZ) and "break" them so
the user can get completely automatic calculation of individual molecules (plus their formula and MWt) or,
given a table of Smiles chemicals, bulk conversion to a standard .ssd file with a large set of chemicals. [If you
happen to have a set of chemicals in another format, such as Z-matrix, which HSPiP cannot handle, then we
recommend OpenBabel, the Open Source program that provides file format interconversion for just about
anything that’s out there. We used OpenBabel a lot when we were developing the implementation of Hiroshi’s
technology]. Charles and Steve have called the method Y-MB for Yamamoto Molecular Breaking [Hiroshi was
too modest to want such a name] and we believe that Y-MB represents a fundamental change in the way HSP
can be used in the future. Hiroshi’s extensive knowledge of Molecular Orbital (MO) calculations and their
interpretation means that in the future Y-MB might be augmented via MO.
In addition to the HSP values, Y-MB provides estimates of many other important parameters such as MPt, BPt,
vapour pressures, critical constants and Environmental values.
Like all group contribution methods, Y-MB isn’t magic. It can’t accurately predict values for groups or
arrangements of groups that are not in its original database. The more HSP that can be measured independently,
the more Hiroshi can refine the Y-MB technique to give better predictions. As mentioned above, the Y-MB
breaking routine can optionally find the Stefanis-Panayiotou UNIFAC groups.
For the 3rd Edition, Hiroshi carried out a huge analysis of results on a database of many thousands of molecules
including many pharma, cosmetic and fragrance chemicals. From this he was able to refine his list of group
fragments and also test novel NN and Multiple Regression (MR) fits. As a result we now have internal NN and
MR variants for calculating the different parameters of Y-MB. Each has its own strengths for different
properties. For the user the only difference from previous editions is that the estimates are often improved –
particularly for very large molecules where we acknowledged that the original Y-MB had problems.
For the 4th Edition, the calculations were further refined to incorporate key parameters such as Ovality and
Molecular Connectivity Index (MCI) which convey some structural information that simple fragments lack.
The predictions are therefore, in general, more reliable.
For the 5th Edition a new, much more powerful Y-MB was introduced. Behind the scenes it calculates many
more parameters, all of which are available if you choose to use them. The same parameters are also used for
the QSAR capabilities described later. One example of the power of the Y-MB engine is that it uses a 2D
version of the famous QEq (Charge Equalisation) methodology for estimating the charges on each atom. These
charges are used internally to refine the estimations but in the outputs the MinCharge and MaxCharge are
provided – which give you an idea of how extreme (or not) the charge distribution is within your molecule. We
believe that the QEq approach will offer more power to future versions of HSPiP.
Although we have always been able to view and convert Molfiles, we had never been able to create a 3D view
from SMILES. From v5.2.x as an extra bonus, the SMILES is automatically converted to a 3D molecule using
the awesome RDKit from Greg Landrum, for which we express our warmest thanks. You can choose to view
the molecule with or without hydrogen atoms – if you want to swap between the views just click/unclick the
Show Hydrogens in 3D option and click the Calculate button. As a formal note:

RDKIt is licensed under the Creative Commons Attribution-ShareAlike 4.0 License. To view a copy of this
license, visit http://creativecommons.org/licenses/by-sa/4.0/ or send a letter to Creative Commons, 543 Howard
Street, 5th Floor, San Francisco, California, 94105, USA.

234
The intent of this license is similar to that of the RDKit itself. In simple words: “Do whatever you want with it,
but please give us some credit.”
Because we believed that the relatively new InChI (International Chemical Identifier) standard for describing
molecules was going to be of great future importance, we output the “standard” InChI and InChIKey. These
are created with the “No Stereochemistry” option so they are the simplest possible outputs. Importantly, if you
use the first 14 digits of the InChIKey as the search string on places such as ChemSpider (probably the best
one-stop-shop for information on a chemical) then you are guaranteed to get the correct matches. InChIKeys
are unique identifiers created from the InChI so unlike CAS# they are directly traceable to specific molecules
and there is only one InChIKey (well, the first 14 digits) to a molecule. The reason we emphasise the first 14
digits is that they will find all variants of a given molecule, independent of stereochemistry, isotope substitution
etc. Looking back after a number of years of InChI it seems our optimism was misplaced – the world still relies
on the uncertainties of CAS#. Hence the upgrade to v5.1.x provided CAS# as standard in all the datasets.
For a useful quick guide to InChI, visit http://en.wikipedia.org/wiki/International_Chemical_Identifier
6 Polymers are a problem. We have no reliable general method for predicting polymer HSP. This is not
surprising. For example, there is no such thing as “polyethylene”; instead there are many different
“polyethylenes” and it would be surprising if their HSP were all identical. But that doesn’t mean that we should
give up. An intelligent estimate can often provide a lot of insight. Hiroshi had proposed an extension of his Y-
MB technique to include polymers. And by good fortune we found Dr W. Michael Brown’s website at Sandia
National Laboratories:
http://www.cs.sandia.gov/~wmbrown/datasets/poly.htm
With great generosity, Dr Brown gave us permission to use his dataset. Hiroshi then implemented a revised
version increasing the number of polymers from <300 to >600. To make it more consistent with the rest of the
program we’ve used –X bonds as symbols of the polymer chain rather than the pseudo-cyclic “0” used by Dr
Brown.
To calculate the polymer HSP, simply double click (or Alt-Click) on one of the polymers. This puts the Smiles
up into the top box. Then click the Calculate button as normal. You can, of course, enter your own polymer
Smiles manually if you wish.
As the whole area of polymer HSP prediction is so new, the Y-MB values for a single monomer repeat can
often be somewhat unreliable. You can, therefore, set a number of repeating units, say, 4, and the full polymer
Smiles for this 4-mer is created and the Y-MB values calculated. You can use your own judgement as to which
value to use – the 1-mer, 2-mer, 3-mer … There are some complications to this automated process. If, for
example, you had a 2-ring monomer and asked for a 5-mer, you will get a message to say that this is impossible
– the problem is that the first rings would be labelled 1,2, the second 3,4 and the 5th repeat unit would be 9,10 –
and polymer Smiles can only use rings from 1-9.
Although this is hugely helpful, we think there’s even more that can be done with this. With a bit of intelligent
copy/paste you can construct polymer blends. For example, if you take polyethylene, C0C0, and
polycyclohexylethylene, C0C0C2CCCCC2, you can combine them to create the ABAB copolymer
C0CCC0C2CCCCC2, or the AABBAABB copolymer C0CCC CC(C2CCCCC2)CC0C2CCCCC2 etc. It’s a bit
tricky (note the extra parentheses around the middle cyclohexyl group) but it’s pretty powerful. To help you
we’ve added a CP (Co-Polymer) button that you can click when you’ve selected two polymers from the
database. The program automatically creates an AB, AABB or AAABBB polymer according to your choice.
Note that it is possible to make “impossible” polymers this way – the program makes no effort to see if two
monomers could actually be made into a co-polymer.
Again we need to stress that this is all so new that the predicted values should be treated with caution. Above all
we need many more experimental data points for polymers and it seems that IGC offers a lot of hope for the
routine gathering of a lot of relevant data. Armed with more data, the polymer Smiles predictions can be
refined.
We had pointed out to users of the Polymer Smiles method in earlier editions that the limitations were
significant. With the improved Y-MB version we are much happier that Polymer Smiles are more stable and
insightful. They should still be used with caution, but their capabilities are clearly much improved for the 3rd
Edition and some anomalies have been fixed in the 4th Edition.
235
The single most popular request for polymer estimation was the ability to estimate the HSP of some arbitrary
blend of monomers. Intellectually this is much more difficult than it sounds. Take the simple example of a
70:30 blend of A:B. We can think of this as a blend of A-A, B-B and A-B dimers (extending the thought to
trimmers makes things too complex). So “all” we need to do is calculate the ratio of A-A, B-B and A-B dimers,
estimate the HSP of XA-AX, XB-BX and XA-BX and do a weighted average. But what is the distribution of
those dimers? It depends on reactivity ratios, % completion of the reaction and so forth. This requires far too
much information. So the calculation assumes equal reactivity and does a Monte Carlo simulation of a long
polymer chain and counts the ratio of the three dimers. If it happens (at an extreme) that the co-polymer of
70:30 A and B is a string of 70As followed by a string of 30Bs (i.e. a di-block copolymer) the estimate is likely
to be unreliable. However, if you know that it’s a di-block you can just do your own simple average and hope
that the polymer doesn’t phase separate into a dual-sphere system with two totally separate HSP values.
If you want to see the structure of any of the polymers, Ctrl-Shift-click on the polymer in the database and a 3D
representation appears in the Y-MB tab. We created the 3D structures automatically from the polymer Smiles
using the public domain OpenBabel utility.
6.b Y-PB Then in 5.2.x Hiroshi provided the all-new Y-PB Yamamoto Polymer Breaking technique for going
from Polymer SMILES (with X marking the polymerization points) to HSP values and other properties such as
density, refractive index, surface energy, Tg, Cohesive Energy. The work required for this was enormous. First,
Hiroshi had to create a giant database of polymer properties from all available sources. Getting the raw data is
hard enough, but curating it is even harder. To take just one example, one database had a value of density for
PMMA of 0.23. This is obviously wrong, but those who assembled the data did not check for obvious errors, so
Hiroshi had to do the work. Users now have access to 1400+ polymers with a large number of experimental
values provided from his vast polymers database, using “representative” values when reported values cover a
wide range. He then applied his neural network technique to match functional group contributions to the
properties. But clearly an ester group in a polymer main chain is very different from a side chain polyester, so
he had to create two types of functional groups: main- and side-chain. This is easy in some cases (the benzene
in polystyrene is a side chain) but hard in others (is the methyl in polypropylene main- or side-chain?). After
years of testing, Y-PB was launched into the first 5.2.x. We know with certainty that Y-PB will develop over
time with more data and that the quality and range of predicted values will increase as the algorithms are
refined with better data.
Revisions to the HSP table
We’ve used all the above considerations to update the HSP data used in HSPiP. Many of the changes have been
minor, some will be more significant. Any changes will be unwelcome to those who have been using the
Hansen table for years. So it’s worth explaining why we made the changes.
There is a fundamental principle that all worthwhile databases contain errors. The published Hansen table
contained a few typos, and a few errors. But many of the changes have come about because the basic data in
other databases such as DIPPR 801 and Yaws' Handbook of Thermodynamic and Physical Properties of
Chemical Compounds have changed. Thanks to YAMAMOTO Hiroshi we were able to carry out a systematic
comparison of the δTot with the published total solubility (Hildebrand) parameters. We could then see if it was
reasonable to change any values using dipole moment and refractive index data contained in those databases.
The fundamental principle of databases means that those databases also contain errors and conflicts. Wherever
possible we corrected those molecules where there was a large (>1) difference in δTot, but used the principle of
least change if DIPPR and Yaws disagreed, and used the principle of common sense when a value in those
databases simply made no sense.
For the 4th Edition we have reluctantly made an addition to the standard dataset. Over the years the low δP
value of 1,4-Dioxane has worried many people. Surely its δP should be higher than THF! Yet dioxane has no
dipole moment so logically δP should be close to zero. For years this low δP value has been used. But a large
solubility dataset (unfortunately not public domain) provided to the authors seemed to be best fit with a higher
δP value. In the end we decided to include a “High P” alternative for dioxane which happens to be a good
match both for the experimental data and the Y-MB estimate. Users should feel free to explore both options.
Much of the time an individual solvent makes no difference to a fit. If it is very close to or very far from the
centre of a Sphere then changes in its value make no difference to the fit. Only if it is on the edge of a Sphere

236
will its value be of great importance. So even if the new value were generally to win favour, it would not
undermine the majority of historical fits using the old value.
We have continued to work to challenge and revise the HSP database, especially when any fresh data appeared.
We continue to be hopeful that new measurements of HSP (e.g. via IGC) will start to accumulate. See the next
paragraph for how you can help!
DIWF
The alternative to DIY is Do It With Friends. The .hsd format is a simple text format that makes it very easy to
exchange HSP values. If members of the HSPiP user community email to Steve their HSP values for chemicals
not included in the official Hansen list then we can start to share them amongst the community. Although each
individual user might be losing out by giving away some hard-won data, the community as a whole will benefit.
When different users come up with different values, we can choose to quote both or launch a discussion to
decide which is right.
Indeed, it might be time for those with their private collections of HSP to open them up to the world-wide HSP
community. Of course they would lose some commercial/academic advantage by revealing their values. But
they would also gain by having those values corroborated and/or refuted by values from other collections. By
assembling one large “official” HSP table, with differences resolved by expert assessment, many of the glitches
and problems in the literature and in our own practical research enterprises would disappear. Will readers of
this book take up the challenge? We hope so!

237
Chapter 33 Predictions (Many Physical Properties are Correlated with
HSP)
The more powerful the theory, the better its ability to make predictions. This chapter shows that HSP can
provide the basis for some very powerful predictions in important areas. The HSPiP software implements each
of the predictions in turn: HSE, Azeotropes/Vapour pressures, Solubilities.
HSE
The HSE modeller captures in one place all the capabilities we have for making rational choices about
chemicals in terms of Health, Safety and Environment. It lets you enter the SMILES for two chemicals and then
lets you compare a large set of important properties. A typical example is “read across” for REACH and other
chemical safety systems. If you know that a particular chemical is safe or unsafe then a rational starting point
for judging the safety properties of an un-tested material is to “read across” from this chemical to a similar one.
But how similar is similar? Only you as a scientist can judge, but in the HSE comparison you can compare
estimates of
• Phase change properties – melting point, boiling point, vapour pressure
• Solubility properties – solubility in water, Octanol/Water partition coefficient, Soil/Water partition
coefficient, BCF (Bio Concentration Factor) for Fish oil
• VOC properties – RER, vapour pressure, flash point, OH radical reactivity, Carter MIR
• Other properties – Heavy Atom Count, Density, Molecular Weight, Molar Volume
• Numerical comparison – HSP distance, “Functional Group distance”
The functional group distance is an estimate based on the (dis)similarity of the functional groups derived from
the Y-MB analysis. If, for example, the two molecules both have FG#27 (primary alcohol) then their distance is
lower than if one has FG#27 and the other has FG#38 (primary amine) which in turn is larger than between
FG#27 and FG#28 (secondary alcohol). The methodology takes into account the different molecular sizes and
the numbers of functional groups. Clearly a molecule with many functional groups must be quite distant from
one with just a few, even if those few match groups in the larger molecule. To ensure that differences aren’t too
exaggerated, although butanol has 4 carbons and methanol has only 1, each molecule is shown as having just
two functional groups, and the distance between methyl and butyl is not all that large. This calculation is
different from one which would count similar methyl and alcohol groups in both molecules but would have two
un-matched CH2 groups in the butanol.
A note on LogP=LogKow=Octanol/Water partition coefficient
LogP is often seen as a highly important parameter. Although it is important we think that it is very much
misused. In our view, HSP are very often much more insightful than LogP. The main reasons we are sceptical
about LogP are:
• It is a ratio which can hide important details. A LogP of 0 (i.e. P=1) could have solubilities of 100/100
or 0.001/0.001. Chemically the former (high solubility) is likely to be very different from the latter (low
solubility) even though the ratio is the same.
• Chemicals in biological environments don’t have a choice between a “water” environment and an
“octanol” environment. A typical lipid environment might be much closer to [16, 3, 3] than octanol’s
[16, 5, 12] and as we showed in another chapter, a lot of key biological entities such as skin or DNA
binding sites are closer to [17, 10, 10] than to octanol. LogP is far too restricted to be able to give a
reliable guide to where a chemical might be going.
But as a service to HSPiP users we felt it was important to provide the best-possible predictor. Hiroshi’s
www.pirika.com has a long article on his search for the best predictor of LogP. Not surprisingly (see the section
on the hydrophobic effect on solubility below) the best single predictor for LogP is MVol. If you plot the LogP
values of many different classes of molecules (e.g. hydrocarbons, nitro, amide, nitrile, amine) you get a series
of straight lines. So LogP has the same linear dependence on MVol, though with a different offset for each
functional group:

238
Figure 33-1 A typical example of linear correlations between MVol and logP. The slopes are the same, with different offsets.
The original article at www.pirika.com has many more examples.
Armed with this knowledge it is possible to do a more exact prediction of LogP taking into account the offsets
from the different functional groups. This requires the offsets to be additive which, fortunately, they happen to
be. With the functional group correction, the correlation between LogP and MVol is very strong:

Figure 33-2 The full correlation of 5,320 experimental LogP values against MVol with functional group correction using the
Y-MB functional groups.
Within the typical range of 0-5, the predictions span a range of +/- 1 LogP unit. Given the inevitable
uncertainties in the experimental values this is an impressive fit.
Azeotropes and Vapour pressures
If a mixture of two solvents were “ideal” then the partial vapour pressures above the mixture would simply
depend on the saturated vapour pressure and mole fraction of each solvent.
But we know that in most cases the presence of one solvent tends to make the other solvent “uncomfortable”
creating a higher-than-expected vapour pressure. The difference between ideal and real is the Activity
239
Coefficient. So to know everything about the partial vapour pressures of a mixture, the activity coefficients
have to be known.
No perfect way has been found to predict activity coefficients, γ. Simple theory suggests that they can be
calculated directly from HSP using the formula
Equ. 33-1 ln(γ)= ln(φ1/x1) + 1 - φ1/x1 + χ12* φ22
where φ1 and x1 are the volume fraction and molar fraction of solute 1, φ2 is the volume fraction of solvent 2
and the parameter, χ12 is given by:
Equ. 33-2 χ12 = MVol /RT * ((δD2-δD2)2 +0.25* (δP2-δP1)2 +0.25*(δH2-δH1)2)
In the absence of any better formula, this is a good-enough approximation, but checking against a large dataset
of activity coefficients shows that it needs considerable improvement.
At the heart of the problem is that fact that the basic formula does not account for positive interactions between
solvents that create activity coefficients less than 1. And a detailed analysis of the failures from predictions of
the simple formula show that the biggest deviations typically come about amongst solvents with large δP and
δH parameters.
One way to fix this problem is shown by MOSCED – Modified Separation of Cohesive Energy Density. This
splits δH into donor/acceptor terms. This seems a good idea. Unfortunately, MOSCED has not become a
generally acceptable methodology and some of the more recent fittings of the complex parameters mean that
the sum of the cohesive energy terms are often very different from the cohesive energy – in other words,
MOSCED has become more of a fitted parameter technique than one rooted in thermodynamics.
In the absence of any breakthrough theory, for the 5th Edition Hiroshi has done a NN fit to a large database of
Wilson parameters. We have found that the Wilson formulation of activity coefficients is more useful than
relying on “infinite dilution activity coefficients” and Margules, especially as we are often interesting in large
mole fraction solubilities.
For Solute 1 in Solvent 2, the activity coefficient γ1 for mole fraction x1 (and therefore x2=1-x1) is given by the
two Wilson (large) Λ parameters:
Equ. 33-3 ln γ1= -ln(x1+Λ12x2) + x2(Λ12 / (x1+Λ12 x2) - Λ21 / (Λ21x1 + x2))
And of course it’s the other way round for Solute 2 in Solvent 1:
Equ. 33-4 ln γ2= -ln(x2+Λ21x1) + x1(Λ21 / (x2+Λ21 x1) – Λ12 / (Λ12x2 + x1))
The large lambdas, Λ, are in turn derived from the MVols, v1 and v2 and the small lambdas, λ11 and λ12 and λ21
and λ22 via:
Equ. 33-5 Λ12=v2/v1.exp(-(λ12-λ11)/RT)
And
Equ. 33-6 Λ21=v1/v2.exp(-(λ21-λ22)/RT)
For obscure reasons, the large lambdas are small numbers and the small lambdas are large number!
Armed with better predictions across the whole mole fraction range it then becomes simple to calculate the
isothermal vapour pressure curves and only slightly more complex to calculate the vapour pressures at the
(variable) boiling point of the mixtures from which it is possible to identify important azeotropes.
Comparing the azeotrope predictions against experimental data shows one obvious point: the azeotrope
temperatures are quite accurate while the compositions can show large errors. Why is this? Because many
azeotropes are quite close to being ideal so small errors in the calculation can produce large errors in the
composition.
Please note that water is not handled reliably by this methodology. Also, the public domain datasets on F-
containing molecules are rather limited so the predictions of important F-containing azeotropes are not as
reliable as they would be if more data were available for inclusion in the NN fitting. These F-containing
azeotropes are a hot topic among some of the specialist chemical companies (because they offer the possibility
of low ozone destruction and low global warming potential), which make it understandable that only a small
number of datasets are in the public domain.
240
Solubility
It seems odd to say that you cannot directly predict solubility from HSP! But HSP have always been about
relative solubility and have never attempted to issue exact solubility predictions.
However, with some simple equations and some good estimations of key properties, it is possible to predict
solubilities directly.
The equation is simple:
Equ. 33-7 ln(Solubility)= -C + E –A - H
C is the “Crystalline” term, sometimes (confusingly) called the Ideal Solubility. It is the Van ‘t Hoff (or
Prausnitz) formula that depends on the difference between the current temperature, T, and the melting point Tm,
the Gas Constant R and also on the Enthalpy of Fusion ΔF.
Equ. 33-8 C = ΔF/R*(1/Tm – 1/T)
In other words, the higher the melting point and the higher the enthalpy of fusion, the more difficult it is to
transform the solid into the dissolved (liquid) state.
This formula is a simplification which follows convention and ignores some other terms like heat capacities. An
even simpler formula, from Yalkowsky, uses just the melting point:
Equ. 33-9 C = -0.023*(Tm –T)
Recently, Yalkowsky has reviewed the various options for calculating this term, S.H. Yalkowsky, Estimation of
the Ideal Solubility (Crystal-Liquid Fugacity Ratio) of Organic Compounds, J. Pharm. Sci, 2010, 99, 1100-1106
and confirms that -0.023*(Tm-T) is good enough. The paper uses Log10 so the printed coefficient is -0.01.
For calculations where Tm≤T, C is zero.
The E term is (combinatorial) Entropy. This is calculated from volume fractions (Phi) and molar volumes.
Equ. 33-10 E = 0.5*PhiSolvent*(VSolute/VSolvent-1) + 0.5*ln(PhiSolute + PhiSolvent*VSolute/VSolvent)
It’s worth making an important reminder that molar volumes for solids are not based on their molecular weight
and solid density. In the words of Ruelle: “(For a solid) the molar volume to consider is not that of the pure
crystalline substance but the volume of the substance in its hypothetical subcooled liquid state.”
A comes from the activity coefficient. The larger the activity coefficient, the more negative A becomes. As
discussed above, a simple version of A can be calculated directly from HSP, but the more sophisticated Wilson
parameter formulation gives better predictions – though these parameters were optimised for VLE and their
reliability for solubility calculations is less good. Because the Wilson parameter predictions for water are not
reliable, if water is chosen as the solvent then the activity coefficients are set to 1 so you can look at the other
three terms.
H is a Hydrophobic Effect term that is very important for solubilities in water, and somewhat important for
solubilities in low alcohols. The calculation follows the method of Ruelle (see, for example, Paul Ruelle, Ulrich
W. Kesselring, The Hydrophobic Effect. 2. Relative Importance of the Hydrophobic Effect on the Solubility of
Hydrophobes and Pharmaceuticals in H-Bonded Solvents, Journal of Pharmaceutical Sciences, 87, 1998, 998-
1014) and depends on rs*PhiSolvent*VSolute/VSolvent with extra terms depending on how many hydrogen-bond
donors (alcohols, phenols, amines, amides, thiols) are on the solute and whether the solvent is water, a mono-
alcohol or a poly-ol. The value rs is 1 for monoalcohols and 2 for water and, for example, ethylene glycol. It is 0
for all other solvents. If the solvent is water and the solute contains alcohol groups, there are special parameters
depending on whether the alcohols are primary, secondary or tertiary. There is a further refinement (not
included in this version) which discounts some of the solute’s hydrogen bond donors if they are likely to be
internally bonded. The important thing about the Ruelle formula is that solubility in water depends almost
entirely on the size of the solute – bigger molecules are simply less soluble than smaller ones. Their explanation
is more sophisticated than the simple idea that bigger molecules disrupt more hydrogen bonds, but the simple
intuition isn’t a bad approximation. They show that for “simple” molecules (one’s without too many –OH
groups) spanning a huge range of solubilities, a first principles formula based on MVol, with no fitting
parameters, does an excellent job at prediction.

241
The complication is that the E, A and H terms all depend on the volume or molar fraction which is precisely
what you are trying to calculate, so there is an iterative process involved until the equation balances.
Although the output of most interest is the real solubility, it is very instructive to see the effect of the different
terms, so the HSPiP modeller shows the C, E, A and H terms. For all solvents that aren’t water or alcohols H is
zero. For water the H term, not surprisingly, can be very large. But because of water’s small molar volume, the
E term can also be large. Because the A term can also be large, water solubility is hard to judge a priori because
it can involve the (partial) cancellation of large numbers.
A very helpful way to think through solubility issues has been provided P. Bennema, J. van Eupen, B.M.A. van
der Wolf, J.H. Los, H. Meekes, Solubility of molecular crystals: Polymorphism in the light of solubility theory,
International Journal of Pharmaceutics 351 (2008) 74–91. The equations below can be switched on and off in
the Crystalline Solubility Theory modeller and plots can be chosen as x v T (so both are in “normal” units) or as
ln(x) v 1/T which is the van’t Hoff plot which gives a straight line (added as a reference to the plot) for ideal
solubility, making it easier to see the effects of switching on and off the different parameters. The Yalkowsky
approximation is included for reference.
For the ideal solubility case the mole fraction solubility x is given by the equation we have used earlier:
Equ. 33-11 Ln(x) = ΔF/R*(1/Tm – 1/T)
However, this assumes that the heat capacity Cp of the virtual liquid at temperature T is the same as that of the
solid. In general the heat capacity is higher so ΔCp is positive. This happens to increase the solubility,
sometimes to a surprisingly large extent via:
Equ. 33-12 Ln(x) = ΔF/R*(1/Tm – 1/T) + ΔCp/R [Tm/T – ln(Tm/T)-1]
If regular solution theory is used then there is an additional term that depends on ΔHmix, the enthalpy of mixing
and ΔSmix, the enthalpy of mixing. If this is positive (i.e. the solute and solvent do not like to be together) then
the solubility is reduced, if it is higher (there is some positive interaction between them such as donor/acceptor)
then the solubility is increased. The formula including all three terms is then:
Equ. 33-13 Ln(x) = ΔF/R*(1/Tm – 1/T) + ΔCp/R [Tm/T – ln(Tm/T)-1] – (ΔHmix -TΔSmix)/R [(1-x)²/T]
HSPiP allows you to play with these terms. Clearly the dominant effect is still the melting point – the higher it
is (and the higher the enthalpy of fusion) the lower the solubility, but the surprisingly large Cp effect and some
assistance from a negative heat of mixing can at least fight against the low solubility that a high MPt generally
brings.
The fact that x is on both sides of the equation for heat of mixing effects leads to some strange plots for high
values of ΔHmix. The strange plots are not realistic because they happen to represent violations of Gibbs phase
rule. Whether they represent “oiling out” effects is a matter that can be followed up by those who read the paper
referred to above.

242
Chapter 34 Improvements?
We’ve never tried to hide the imperfections of HSP. And we’ve often mentioned that we were working hard to
create improved techniques. This chapter describes some of the outcomes of lots of hard work challenging our
own assumptions.
Sphere fitting
The official Sphere method described in the Handbook has served the HSP community very well for many
years. There have been a few tweaks to it as HSPiP developed and a GA (Genetic Algorithm) method was
added that coped better with some poor data sets. But now there are some logical alternatives to what we’ll now
call the Classic method.
The first is a response to many users’ requests and something we had also wanted to do. It creates a Sphere
based on real data which record, say, solubilities or swellabilities. In this case “more” is “better” – unless you
choose the “good is small” option in which case “more” is “worse”. We had been greatly worried by the fact
that the fitting algorithm would depend heavily on the assumptions behind the fit, and that neither we nor the
user would know what those assumptions should be. But in the end we found that our worries were not
necessary. The GA technique seems to do an excellent job fitting all sorts of data. The centre of the Sphere is
likely to be more accurate than an Inside/Outside fit, but the radius is unknown and has to be judged by
yourself. You can do a simple check on this by selecting a “Split” value that defines “good” above it and “bad”
below it.
The second is also something requested by users: a Double Sphere option that tries to find out if your sample
contains a mixture of materials as in, for example, a di-block copolymer. The first thing to emphasise is caution.
Fitting too little data with too many parameters can lead to too many errors. Finding objective criteria for
finding the best of all possible pairs of Spheres is difficult and the GA works very hard to come up with a
credible answer, but it can only do so much. So don’t get too excited about the two Spheres unless you are
convinced that the data really does support the values. One user emailed very puzzled that the radii of the two
spheres in one fit were each larger than the radius of the single sphere fit. A close look at the data showed an
absence of restraining data in the key directions where the double spheres were placed, so they could expand at
no cost, giving large radii. The fact that the large radii rang alarm bells is good – with all fitting (not just with
HSP!) users should always use their scientific judgement about the value of the fit and, if possible, provide key
extra data (such as some solvents in the areas without data) to challenge the fit.
MVol effects
It’s always been clear that smaller molecules should give higher solubility than their HSP distance might imply.
The MVC (Molar Volume Correction) option was an attempt to correct for this effect and some users
(including ourselves) have found it helpful. In the GA mode there is also a MVol option. The need for it has
been tested by Hiroshi using his SOM (Self Organising Map) technique. With good data sets the SOM readily
splits into two groups representing inside and outside. But with difficult sets, adding MVol as a parameter
clearly helps to create a better split. Therefore MVol should give better Sphere data.
If this is true in general, it raises the question of how the Classic Sphere technique has ever been of much use.
Such a question is even more relevant when we discuss Donor/Acceptor, so we’ll ask the question now, and
provide a surprising answer.
Why has the Classic Sphere technique worked so well for so many years?
We were having a furious debate about Donor/Acceptor and into the mix we threw the question of the MVol
effect. It seemed to us that if these effects were significant then the Classic technique must have been wrong for
all these years. From this we thought that either Donor/Acceptor or MVol effects must always be small, so that
Classic was always right, or that the effects could be significant but somehow didn’t mess up the Classic fit. It
was then we reached our “aha moment”. The majority of “good” solvents tend to sit in the middle of the
Sphere. So if they are made a little better or a little worse by other effects then they would still be inside the
Sphere and would make no difference to the fit. The same applies to the majority of “bad” solvents – if they are
made a little better or worse by the effects then they would still be bad. The result is that the centre of the
Sphere will not be changed much by any of these effects. The effects will change things on the border but that
will mostly affect the radius. But we all know that the radius is not well defined anyway – it depends on the
243
user’s judgement of what is “good” or “bad”. The same polymer could have very different radii if one user was
concerned about solubility and another by swelling.
Turning this on its head, the fact that Classic HSP have worked so well for so many years demonstrates that
Donor/Acceptor and MVol effects must be modest at best. For MVol corrections it turns out that a typical
variation will be in the range of root-2. Theory suggests that MVol affects Distance² and if we say that a typical
solvent has a MVol of 100, other solvents in the tests won’t range much below 50 or above 200. And in the next
chapter we will find a plausible reason why Donor/Acceptor effects are usually irrelevant.

244
Chapter 35 Into the 4th Dimension. Donor/Acceptor

We would like to warmly thank Professor Michael Abraham, University College London, for his generous
assistance with respect to his Abraham parameters.
As we’ve noted in other chapters, there are other approaches to determining solubility. Each has its strengths
and limitations. Here is our view, in alphabetical order, of some of the main approaches.
Abraham parameters. For more than 20 years, Professor Abraham and his team have methodically worked
out a set of 5 parameters that allow users to calculate solubilities and partition coefficients based on linear free
energy relationships. The parameters have been worked out through careful, complementary experimental
processes using solvatochromic shifts, NMR shifts, GC and HPLC retention times. The approach has been
adopted in a number of areas and the large experimental database of parameters is a key aspect of the approach.
We will discuss the parameters later in this chapter.
COSMO-RS. Dr Andreas Klamt and colleagues at COSMOlogic have developed an entirely new way of
working with solubility and partition issues. At the heart of the technique is quantum mechanical calculations of
each molecule (and, sometimes, each major conformer of the molecule). Once the calculated data for the
molecule is known, subsequent calculations of interactions with other molecules are rapid. In principle,
therefore, COSMO-RS can do everything that all methods try to do, but can do it from first principles. There is
no doubt that this approach is very powerful. As the base of validated quantum calculations increases the
usability for everyday problems will increase. For the purposes of this chapter we note that COSMO-RS is able
to generate “COSMOments” and a 5-parameter set which we will discuss below.
MOSCED. MOSCED’s roots are not too distant from HSP. At an early stage it was recognised that increasing
the number and complexity of parameters would provide better fits to the data. In particular, MOSCED
introduced an acid/base split of parameters to recognise specific interactions such as acetone/chloroform.
MOSCED becomes, therefore, a 5-parameter set. Recent MOSCED papers have shown excellent agreement
with activity coefficient data. However there does not seem to be a robust set of MOSCED parameters for
general use and it is unclear to us how the current MOSCED parameters relate to the cohesive energy density
that is at the root of the concept.
PSP. Professor Panayiotou has produced a series of fascinating papers on Partial Solubility Parameters. In early
versions the equivalents of δD and δP remained the same and δH was split into acid/base using COSMOtherm
moments. In later versions, δD was stripped of some of its “polarizable” elements which were placed into and
new type of δP, again using COSMOtherm values to provide the parameters. The problems of what to do with
distance calculations with split δH (see below) was treated with more sophistication. In addition, there is a
brave notion that the sum of the solubility parameters does not (for many molecules) have to equate to the
cohesive energy density. This is liberating but makes it harder to pin down the overall parameter values.At the
time of writing PSP shows great promise but has not yet crystallised into a stable version with parameters that
can readily be obtained for molecules of interest. Further developments in PSP are eagerly awaited.
UNIFAC and its variants. If you are a UNIFAC user then you probably aren’t reading this eBook. UNIFAC’s
strengths are unparalleled in its main domain of use for vapour liquid equilibria in the chemical industry.
However it doesn’t seem to have caught on as a tool for the broad range of applications for which HSP are so
suitable. There is now a huge database of UNIFAC group coefficients and if you have access to the database
you can do many things better than you can with HSP.
If we leave out UNIFAC as a special case, it’s clear that other methods have 5 parameters whilst HSP has only
4. Why 4 – surely HSP has only 3 parameters? In this chapter we’ll emphasise the fact that HSP calculations
regularly use the MVol. MVol has appeared in all HSP data tables and has been quietly working away in the
background. The other 5 parameter sets also include a term that is equivalent to MVol.
An important paper, Andreas M. Zissimos, Michael H. Abraham, Andreas Klamt, Frank Eckert, and John
Wood, A Comparison between the Two General Sets of Linear Free Energy Descriptors of Abraham and Klamt,
J. Chem. Inf. Comput. Sci. 2002, 42, 1320-1331, shows that the Abraham approach and a simplified version of
COSMO-RS (COSMOments) can both be described by a 5-parameter set which can be adequately mapped

245
between the two techniques. This paper, incidentally, is an excellent introduction to both approaches and is
recommended for those who want to understand more about them.
For the purposes of this chapter, the paper can be summarized by saying that the HSP terms of δD, δP and
MVol map onto corresponding terms in both Abraham and COSMOments. But δH is a single term whilst
Abraham has Acid/Base and COSMOments has Hdonor and Hacceptor. The mapping of Abraham Acid/Base onto
COSMOments Hdonor/Hacceptor isn’t perfect – after all the other parameters aren’t perfect maps – but the paper
makes the point that the general idea of a 5-parameter linear free energy space is a core thermodynamic
concept.
Because it is clear that both Abraham and COSMOments work well with extensive experimental databases it
means that HSP has to defend its use of a single δH parameter.
There has been no shortage of suggestions that HSP should divide the δH parameter and there have been a few
attempts to make it happen. But there have been three practical objections. The first is that the 4-parameter HSP
(remember, we are including MVol as a parameter in this chapter) works remarkably well. The second is that
plotting in 4D space (δD, δP, δHD, δHA) isn’t possible in this 3D world. The third is that there’s been no
obvious way to partion δH into the two terms.
But the classic acetone/chloroform case where there is unambiguous evidence of donor/acceptor interactions
between the solvents shows that 4-parameter HSP cannot describe everything. However, as noted in Charles’
history in the next chapter, even this case doesn’t show up as special in general HSP use.
Acid/Base or Donor/Acceptor
Sooner or later we have to decide on terminology. The world is split into those who think that the best term to
describe the two terms is Acid/Base and those who think it should be Donor/Acceptor. As you can see above,
Abraham and Klamt use respectively Acid/Base and Donor/Acceptor.
We decided that Acid/Base is rather too literal so have chosen Donor/Acceptor. Hence we will talk of δHD and
δHA rather than δHA and δHB. The fact that the A of Acceptor is the opposite meaning to the A of Acid is
unfortunate, but there’s nothing we can do about it.
While we are clarifying terminology, let’s restate that we talk about a 4-parameter or 5-parameter set but a 3D
or 4D viewing space. This is because the (scalar) MVol is not shown in the (vector) plots of 3D or 4D space.
Inspired by Abraham
Reading the extensive publications of the Abraham team and examining their large public database of
parameters it is clear that their thoughtful approach to working out the Acid/Base parameters is much to be
admired. The IGC and HPLC techniques for HSP were developed independently of the equivalent Abraham GC
and HPLC work and show that in principle the Abraham parameters could map onto an HSP 5-parameter set or,
to put it another way, HSP could in principle (though this won’t happen in practice) gather Donor/Acceptor
values in a similar fashion.
We therefore decided to create a 5-parameter HSP set using Abraham parameters to help us in one important
step. We decided to split δH using two rules.
Rule 1: δH² = δHD² + δHA²
This rule ensures that everything about HSP stays constant and we can always bring Donor/Acceptor back into
4-parameter space without upsetting 40+ years of work.
Rule 2: For compounds with known Abraham Parameters, δHD:δHA = Abraham Acid:Base
This rule allows us to get started on the whole process of rationally splitting δH. We have no pure scientific
justification for this mapping other than our feeling that the Abraham approach to Acid/Base determination
(e.g. GC/HPLC) fits very well with HSP.
Once we were up and running with a basic set of Donor/Acceptor splits, it was then possible to create Y-MB
methods for splitting molecules for which we had no Abraham parameters.
After that it requires a lot of checking and invoking chemists’ common sense. If the automated process
produced an amine with a large δHD then clearly there was a problem with the process.

246
Out of this work came an important third rule:
Rule 3: If you have no other way to decide, make δHD=0 and δHA=δH.
This surprising rule is less surprising if you glance at any Abraham table. For example, in a table of 500
compounds there are 466 with Base values and 249 with a Acid values. And of those 249, only 96 have a value
bigger than the Base value. So most molecules are Acceptors rather than Donors. Of course Rule 3 should be
used with care but note that it says “if you have no other way to decide”. If a molecule has a carboxylic acid
group then you already know something so Rule 3 doesn’t apply.
For the 5th Edition, Hiroshi finally parted from the Abraham method of calculating donor/acceptor to reduce
various anomalies apparent during extensive tests.
Calculating the distance
The previous section describes in a few words a very large amount of work. This section describes some hard
thinking.
When 5-parameter HSP seemed only an impossible dream we assumed that if we had them it would be easy to
calculate the HSP distance. MOSCED, for example, uses (in our nomenclature) a term (δHD1-δHD2)(δHA1-
δHA2) instead of (δH1-δH2)². This term captures the possibility of a “negative distance” which is what
Donor/Acceptor is supposed to accomplish.
However, we quickly found that we cannot use such a term. If the two donor terms are equal then the first term
is zero so the distance is zero. But this cannot be the case for HSP. For example if δHD1 and δHD2 are both
zero then the second term should be equal to the classic δH distance because each δHA term equals the classic
δH.
We eventually found a distance formula that works well. It gives the sorts of values we intuitively expect in all
the test cases we can find. We would love to be able to tell you that we understand the reason for the formula,
but we admit that we use it because (a) it seems intuitively right and (b) it is the only formula out of many
variants that gave us only values that made intuitive sense. If someone from the HSPiP user community can
prove it or provide a better alternative we would be happy to acknowledge their work in a future edition.
So if we have (δHD1 , δHA1) and (δHD2 , δHA2) then using the nomenclature Min(X,Y) to mean the minimum
of X or Y, we define:
Equ. 35-1 MinX1 = Min(δHD1 , δHA2)
Equ. 35-2 MinX2 = Min(δHD2 , δHA1)
Equ. 35-3 X1 = δHD1 - δHA2
Equ. 35-4 X2 = δHD2 – δHA1
Equ. 35-5 S1 = δHD1 - δHD2
Equ. 35-6 S2 = δHA1 - δHA2
Equ. 35-7 DA1 = Min(δHD1, δHA1)
Equ. 35-8 DA2 = Min(δHD2 - δHA2)
Then
Equ. 35-9 Distance = -Sqrt(MinX12+MinX22) + Min(Sqrt(X12+X22), Sqrt(S12+S22))+Sqrt(DA1²+DA2²)
If both δHD values =0 or both δHA values=0 then this term becomes the classic (δH1-δH2)²
If δHD1 is large and δHA2 is large while δHD2 is small and δHA1 is large then we have a classic donor/acceptor
(chlororform/acetone) pair and the distance is negative, which is precisely what we require. Mathematicians
will note that when we calculate the Distance² (in the general distance formula) Distance² =Sign(Distance)*
Distance².
Most other cases show some reduction in distance from the classic δH distance because of some favourable
donor/acceptor interaction.

247
It’s worth noting that although there are a large number of molecules with a high δHA and a low (or zero) δHD
there are very few molecules (<5%) with a δHD more than twice δHA, and only 2% with more than 3* δHA. So
the likelihood of strong pure Donor:Acceptor effects is very low.
Although the Distance formula is intellectually superior to the well-know Beerbower formula for distance, in all
the tests we have been able to apply, the Beerbower formula gives superior results as well as being much easier
to calculate. So that is now used throughout HSPiP.
Equ. 35-10 Distance = 2(δΗD1- δΗD2) (δΗA1- δΗA2)
Our first 4D Spheres
We were very nervous when we created our first Spheres using standard .hsd files (and therefore known classic
Spheres). What would happen if the results were very different. Would this mean that 40 years of HSP have
been wrong?
We needn’t have worried. It quickly became clear that for most systems the results were not very different.
There are three very important reasons for this. The first is obvious: because δD and δP have not changed, it’s
not possible for the Sphere to move too far. The second reason follows from the facts behind Rule 3 above.
Because a large majority of molecules have a very low fraction of Donor, the majority of interactions are
Acceptor:Acceptor, in other words they are simply classic δH. The third is simple statistics: the number of
systems with large δH values isn’t large and the effects of a Donor/Acceptor split from a small δH value is not
highly significant.
But the fact that they are not very different doesn’t mean that the whole 5-parameter approach is a waste of
time. In some cases the fits improve because there really are specific Donor/Acceptor interactions for a few
solvents and a few polymers, particles etc.
For us, and we think for you, the move into 5-parameter space is a useful option. For most work, the proven 4-
parameter methodology works very well and there is no need to add the complexity of the 5-parameter space.
But given that we can create the Donor/Acceptor split automatically and that the Sphere fitting isn’t too much
slower, it’s easy to explore the option if there are reasons to expect that Donor/Acceptor effects will be large.
And remember that “more” isn’t always “better”. The split adds extra uncertainties to parameters which
themselves have errors. It is possible to make things worse by trying to be too clever. So we recommend that
for routine use, the 4-parameter classic method should be your default. But feel free to enter the 4th dimension
at any time.
In parallel, Hiroshi has been allowing users to try out Donor/Acceptor calculations using the simple distance
calculation (and also the Beerbower 2*(δHD1-δHD2)*( δHA1-δHA2) which also has obvious difficulties). For
some users the results have been useful, so the SphereFit Power Tool includes these options.
Reporting the results
Because we cannot plot a 4D Sphere and because of Rule 3 above, we will continue to plot the classic Sphere.
So δH will always be calculated for you from δHD and δHA via Rule 1 and plotted as normal. But of course we
show you the δHD and δHA values in a table so you can understand what is going on with each of the solvents.
We cannot save 4D data in the old .ssd file format. The new, standard, more versatile .hsd format is equally
happy with 3D and 4D data.
The problem of mixtures
One of the great strengths of HSP is the ability to create arbitrary mixtures and calculate the HSP by the
standard mixing rule. Things get more complicated when solvents are blended with Donor/Acceptor.
As an extreme example (which, see the comments above, is highly unlikely to exist), suppose we have a 50:50
blend of [10,0] and [0,10] – the mixture of an equally strong pure donor and pure acceptor. In δH terms the
classic answer is 5+5=10. One possible intuition is that the donor/acceptor cancel out giving [0,0]. But it seems
unlikely that the resulting blend would have no δH. It also seems unlikely that it would be [10,10] with a
resulting δH of Sqrt(10²+10²)=14.14. After much experimentation we found that the simplest possible mixing
rule, the weighted average of δHD and δHA terms, gave sensible and intuitive results. In this example the
answer is [5,5] with a total δH of 7.07. Most other cases aren’t so extreme and the resulting δH is not much
reduced from the simple weighted average of δH terms.
248
An alternative possible rule is that δH of the mix remains the same as the classic mixing rule, but the ratio of
δHD and δHA is changed depending on the relative values of the components. This is a slightly more complex
rule to implement and in most cases makes little difference, so for the moment we have chosen the simple rule.
The experience of those interacting with Hiroshi on his Pirika website shows that even simple Donor/Acceptor
can be insightful for solvent mixtures. Therefore we have added Donor/Acceptor to the Solvent Optimizer,
using the simplest mixing rule. If you find this option helpful, use it with caution because of the intellectual
limitations. If you don’t want to risk it, just don’t click the Donor/Acceptor option.
Activity coefficients
In other chapters we have pointed out that activity coefficients should be directly calculable from HSP distance.
But there is lots of experimental evidence to show that this is not a reliable method for prediction. Clearly the
impossibility of a negative distance, indicating a Donor/Acceptor interaction, has been a limitation of the 4-
parameter HSP.
We have still not found a way to use a 5-parameter model to give more reliable predictions. The recent Wilson
parameter mode has been an improvement. A new Wilson mode is due for the next edition.
Use with care
We believe that this shift to a 5-parameter HSP is a necessary step in the evolution of HSP. But our evidence is
that the step is in most cases not of great importance. So classic HSP has not suddenly been invalidated.
Therefore we are not shifting everything at once. On the surface, much remains 4-parameter, though behind the
scenes we use 5-parameter when it seems sensible. Unless you select the D/A option you won’t even see the 5-
parameter data.
We have some internal cases where the Donor:Acceptor correction is so large, and the δD and δP parameters
are so close that the distance is negative, though without an especially high solubility. In other words our
formula seems to over-correct in some cases. We will carry on our internal arguments on the issue while this
current edition is out in the open. We very much look forward to your input to the debate.
But the key take-home message here is to “use with care”. It’s possible that we will change the Distance
algorithm and the Mixing algorithm. [This proved to be the case. In v4.2 a new Donor/Acceptor ratio
calculation, bypassing the Abraham method, was introduced and despite our misgivings, the Beerbower
distance estimate seems to be the least bad and has become the standard.] Our splitting ratios may change with
time as we gather more experimental data. So enjoy exploring the 4th dimension, take what value you can from
it, but, at this stage, don’t rely too heavily on its predictions.

249
Chapter 36 QSARs

For as long as they have known each other there has been an argument between Steven and Hiroshi about
“science” versus “fitting”. Hiroshi has successfully using fitting parameters throughout his scientific career both
in industry and via his Pirika site. Steven’s instincts have always been to avoid fitting and find a “scientific”
formula for a given phenomenon. Of course, neither extreme position makes sense. A “scientific” formula is
much better than a fitting formula if the science is known. A fitting formula can be really useful if there is no
available science. And a fitting formula might inspire the development of a scientific formula. The argument
has always been the balance between the two extremes.

But when Hiroshi developed the Y-MB 14Nov engine (November 2014) it was clear that Steven’s reluctance to
add QSARs (Quantitative Structure Activity Relationships) to HSPiP was no longer helpful.

And there was complete agreement that the aim should be to fit to a reasonably small number of parameters
rather than to have a QSAR system with thousands of parameters. The release version actually has rather too
many parameters, but things like number of C, O, N atoms are useful parameters for other reasons and are
automatically included in the Y-MB prediction output. In reality, the reduced list of ~25 parameters which is an
option is probably the one that most of us should use.

Why are we pleased to have fewer parameters? Surely more are better? The opening debate is relevant here.
Yes, we want the benefits of fitting, but we don’t want to lose touch with the science. There are relatively few
truly significant molecular parameters and it is better, in our view, to produce fits based on those core values
rather than having a bunch of parameters whose meanings are obscure to most of us. HSP, Antoine, MPt, BPt,
Density, MWt, MVol for example are going to be useful across a broad range of fits. But even here we have a
problem. If you fit with both Density and MWt the chances are that you will get a better numerical fit than if
you’d fitted with MVol. But MVol contains both MWt and Density (MVol=MWt/Density), so is probably more
scientifically compact. The fact that it gives a poorer fit is simply because fewer parameters generally give a
less good fit for arithmetical rather than scientific reasons.

We’ve written HSPiP QSAR in a way that encourages you to play freely. You can fit manually, you can find
automatic fits, you can try cross terms and you can choose how many terms to include in the automatic and
cross term fits. Most calculations are essentially instant and you get immediate graphical feedback on the
quality (or otherwise) of the fit – with the ability to move your mouse over points of interest (“bad” points are
usually of interest) and find which chemicals they are and what the differencea are between real and predicted
values. Your goal should not be to get the best possible fit. That’s easy - just choose automatic fit with 15
parameters, you’ll get a great fit which is entirely meaningless! The goal is to get an adequate fit with the
fewest possible parameters, because this is likely to be the most scientifically insightful fit. Remember, that
your original data has errors. A highly parameterised fit that accommodates the noise in your data is of no value
to you.

So a fit with a few, scientifically insightful parameters is highly desirable.

Also desirable is a fit with good predictive value – after all, the fitting is usually to help us predict what
molecules will give us desirable properties for our applications. This is another reason to avoid “perfect” fits
with lots of parameters. These can create the problem of “overfitting”, which gain perfection for the training set
at the cost of being wildly wrong for the predicted values. It’s better to have a reasonable fit and reasonable
predictions.

It seems useful to provide (with some edits) the text of the Help file that accompanies the QSAR. Although the
text addresses the mechanics of using the QSAR, it is much more concerned with the issues discussed above.

250
How to use HSPiP-QSAR
Quantitative Structure Activity Relationships, QSARs, are a way of finding which parameters, in what
combination, give a good fit to a set of data. The quality of the fit, the parameters involved and their relative
sizes can provide important insights into the factors behind the data.
Equally, the QSAR can be used to predict the properties of molecules for which there are no experimental data.
Creating a QSAR
All it needs is some molecules, some data, some relevant parameters and a means to fit the parameters. The
fitting is relatively straightforward if linear fitting regimes are used, as in HSPiP-QSAR.
Non-linear fitting directly involves more complex dependencies such as logs and exponentials. It is obviously
more powerful but tends to send QSARs into a realm of over-complexity. As we shall see, things are already
complex enough even when using just linear QSARs.
The full details of the data inputs/outputs are described in the QSAR Help. Note that there is a Log10 option so
that you can see if the fit is better to the log version of your data.
The parameter challenges
There are two challenges with respect to parameters:
1 Finding a consistent set of parameters that might be relevant
2 Choosing which parameters should be included in the fit.
The HSPiP Parameters
In HSPiP-QSAR the parameters challenge is solved through used of the Y-MB (Yamamoto Molecular
Breaking) estimation scheme developed for HSPiP by Dr YAMAMOTO Hiroshi building on the Joback
technique of functional group contributions, though using his own automatic functional group creating scheme
directly from SMILES strings. The Y-MB technique has developed steadily over the various editions of HSPiP
and the quality of its predictions, based on data mined from many resources such as DIPPPER or Yaws, is now
rather high for “reasonable” molecules. The number of parameters generated is “only” 58. Other QSAR
schemes can generate 1000’s of parameters. Aren’t you missing out with only 58? Our view is that 58 is
actually too many – we create them because they are generally useful for other aspects of HSPiP. As mentioned
above, the second parameter challenge is to choose parameters worth fitting. When you have 1000’s that task
becomes unmanageable for our intended user-base, those who are experts in their own field but who aren’t full-
time QSAR users. As we shall see, within HSPiP-QSAR the automatic parameter selection chooses from a
smaller subset of ~23, and even this is rather too large for reasons we will discuss below.
The key is that the fitting parameters should, in principle, capture something uniquely important about each
molecule. Naturally we believe that the HSP are important. Similarly MVol (or MWt + Density), Ovality, BPt,
Antoine Parameters, RI (refractive index), LogS (water solubility) and LogKOW (octanol/water partition)
represent some fundamental parameters that will often be of importance in chemical datasets. But note that BPt
and Antoine Parameters are not independent parameters: BPt can be calculated from Antoine Parameters.
Similarly, RI is strongly related to the δD parameter within HSP. The more parameters you have the less likely
they are to be independent of each other and the less insights/value you will get from a QSAR.
Of course it’s theoretically possible that some key property really depends on one or more of 324 “edge
adjacency indices” or one or more of 213 “2D autocorrelations” and those who are true QSAR experts might
benefit from such a vast choice. But for most of us such remote parameters provide little chance of giving us
key insights into why our molecules behave the way they do, and giving us better ideas of molecules that might
perform even better in our intended application.
To get all the parameters you simply load a data file (simple tab separated or comma separated file created from
a text editor or Excel) containing name, SMILES and data as the three columns. Clicking the Data button
generates values for the whole set. Any missing or wrong SMILES generate an obviously empty line which can
be either fixed with the correct SMILES or deleted for the fitting operation. You can choose to look at all the
parameters, the core set or (once you’ve made a selection) your selected parameters. Feel free to use any option
that makes it easier for you to see what is going on.

251
You don’t have SMILES for your molecules? Your favourite chemical drawing program will certainly provide
them as outputs (often on the Clipboard), freeware such as OpenBabel can convert from other formats (such as
Molfile) into SMILES and on-line resources such as ChemSpider make it very easy to search for a molecule
and get its SMILES code copied onto the Clipboard to paste into your Excel or text file. And for those who
have the full version of HSPiP, there are ~10k SMILES strings that can be found via a quick search via name,
formula or CAS number.
The Parameter Choice Challenge
We always recommend that users start with a manual choice of parameters to get an idea of what factors might
be influencing the behaviour they observe. A typical set would be HSP, MVol, Ovality (shape is amazingly
important in so many behaviours!) and either the Antoines or something like logS or logKOW. Just select them
by typing a 1 into the top row or deselect with a 0. Then, making sure the Manual option is selected, click the
QSAR button
You immediately get a graph showing your fit, a value in the R² box which shows the quality of the fit (1 is
perfect, 0.1 is near worthless) and the best fit equation where you see the fitting coefficients. Moving your
mouse over the datapoints gives you a readout of the name, experimental value and calculated value.
If the fit is great and some of the fitting coefficients are small, try the fit without them. If R² is not much
reduced then the fit is likely to be more meaningful – especially if the noise in your own data is significant so
that a perfect R² is not possible. Get into the habit of playing around casually. The calculations are effectively
instantaneous and bad fits can be as informative as good fits. If you are convinced that logKOW should be
important but it adds no value to the fit then you have learned something useful about your system.
QSARs are basically stupid. They are not magic, they are just numbers. You are smarter than a QSAR
algorithm. So use QSARs as an aid to thinking through issues and as a guide for predicting outcomes. As we
will see, that guide might easily prove to be unreliable if you ask it to take you into places the QSAR has never
seen before.
Now you’ve used your skills and intuitions, it’s good to see if HSPiP-QSAR can do better. So choose a small
number of parameters (say 3), select the Fit to N Parameters option and click QSAR. This automatically
checks all possible combinations of the core subset of parameters and presents you with the best. If you wish,
you can select the entire HSP set (without δH because δHdon and δHacc are more powerful) as one value and
the three Antoine parameters as a single value via the check boxes. This would be sensible if you felt, for
example, that a single Antoine parameter within a fit makes no sense. Now try the same thing with a few more
parameters. Of course more parameters tend to give a better fit, but this might be just arithmetic and nothing to
do with the real science. More parameters also give you less scientific insights. So aim for a good fit with a
small number of parameters.
To avoid the automatic fit using a parameter you know to be unsuitable, put an X into the top of that column –
anything with an X (or x, case doesn’t matter) will be excluded from the automatic fitting.
Again, the aim is not for the black box to give you the truth. The aim is for you to gain insights into what’s
behind your data. Sometimes that automatic method gives us unexpected parameters that suddenly make sense.
Sometimes it just gives scientifically meaningless correlations that happen to be mathematically superior.
Hence the need for the X option to exclude those that seem to make no sense to you. You can always un-X
them if you change your mind.
Cross terms
We know that sometimes parameters can interact to affect an outcome. You can test this possibility if you start
with either a manual or automatic fit then choose a number of cross terms (start small with 1 or 2), select the
Cross Terms option and click QSAR. Sometimes the cross terms are with the same parameter – this implies a
strong (rather than linear) dependence of that parameter. Again it is up to you to work out whether these cross
terms are meaningful. Mathematically they will increase the quality of the fit, but that’s not the same as
increasing the quality of the scientific insights.

252
Making Predictions
Sometimes the fit is all you need, because it then gives you the scientific insights required for your
development needs. Many times, though, you need to predict the outcomes of other molecules or, at least, test
different molecules to see which might have properties in the desired range.
Once you have a fit, simply load another dataset (again a tab or comma separated text file) of name and
SMILES into the Calculated chemicals section and click the Calc button. The Data column now contains the
predicted values. The Clipboard icon allows you to copy/paste into something like Excel for further analysis.
The temptation is to end the explanation at that point. But users have to be clear that QSARs can give stupid
predictions. Suppose, for example, that a QSAR was created based on a set of data from alcohols and then
predictions were required for a set of alkanes. The results are likely to be a disaster. Is that because HSPiP-
QSAR has failed? No! A QSAR simply creates the best fit to the data it is given. If those data are from one part
of chemical space (e.g. the alcohols) and there are no data from another space (e.g. the alkanes) it is impossible
for the QSAR to include terms that (in retrospect) are important for alkane space. It’s not so much Garbage In,
Garbage Out. The alcohol dataset is not “garbage”. It’s “Limited In, Limited Out” – if your input dataset is
limited then so is the output dataset.
So the first rule of QSARs is for the input dataset to cover as much as possible of the desired chemical space for
the final use. So having data from 10 alcohols is far less valuable than having the same number of datapoints
with data from 2 alcohols, 2 ketones, 2 acids, 2 esters and 2 ethers (to use simple oxygen functionalities for
clarity).
The second rule of QSARs is to split your known data (randomly) into two sets – the training set and the test
set. This is trivial to do for HSPiP-QSAR – you just use, say, Excel to sort your data in a variety of ways then
take out random subsets (e.g 50:50 or 75:25) which you save as Set1Train.txt, Set1Test.txt, Set2Train.txt,
Set2Test.txt and so forth. Load each training set, get a QSAR then use it to predict values for each test set. Use
the Clipboard button to bring the predicted values back into your original Excel to compare (e.g. by plotting
real versus predicted). Make sure you copy/paste the formula from each fit (clicking the F clipboard button) so
you have a record of the various QSARs.
Doing this takes very little time and you quickly build up an intuition about which (if any) of the QSARs is
optimal. Perhaps at the end you will combine all the data into one training set and use the QSAR from that large
set for predictions of unknowns. There is no perfect way. If you have a very large dataset of values (and most of
us don’t!) then a few 50:50 or 75:25 training/testing splits will very rapidly lead you to a stable QSAR. And it is
more likely that they cover a large part of chemical space so their predictions will be valid over a wide range of
unknowns. With smaller datasets it is much harder to know if you have a good QSAR. The wise scientist
recognises this and treats predictions with some caution, especially for molecules with functionalities far
outside the training set.
An excellent way to challenge your fit is to deliberately exclude one of your input chemicals. In the HPC
example discussed below the first datapoint, acetic acid, stands out badly in some of the automatic fits. The
question naturally arises, what happens if you exclude acetic acid? You can delete the whole row (click the
“selector” at the left-hand edge, just as in Excel, the press the delete key) or you can just put something non-
numeric such as a “-” sign or an “x” in the data value. Re-running the QSAR either gives you the same basic
result, using the same parameters with a better R² (so acetic acid was a genuine rogue point) or gives a very
different set of parameters with a fit that may be better or worse. This latter case suggests that the QSARs are
rather arbitrary – they just happen to be good arithmetical fits with no chemical insight.
ΔR²
When you do a non-manual fit the program, of course, selects the best fit. If that best fit is much better than the
next-best fit then you know that you can attach some significance to the chosen parameters. Unfortunately, it is
often the case that the difference between best and second-best is in the 4th or 5th decimal place, i.e. there is no
special significance to the parameters that appear. This difference is shown as ΔR². Get into the habit of
glancing at this value. If it’s small then that is a warning to not get too attached to whichever parameters were
chosen as best. Use the “x” trick above to see how the chosen parameters, and ΔR² change to form a judgement
about which parameters are truly significant and which are numerical padding. After a Fit to N Parameters swap

253
to Manual mode and systematically turn on and off each of the parameters in turn. This will allow you to
identify their relative significance.
HSP Distance
The simplest correlation between solubility and chemical properties is via the HSP Distance from the target
molecule. If you have some idea of this target then you can enter its δD, δP and δH values and select the
Distance column via a “1” value. The Distance column is automatically calculated and then treated as yet
another parameter and is, of course, included in the Calculated values. If Distance is not set to 1 then the
calculations are not done (after all they are only relevant to the specific HSP target set) and not included in the
tables.
Temperature effects
If HSP are part of your fit and if the experiments were done at a temperature other than 25°C then it is wise to
use the HSP values at your given temperature. It then becomes tempting to change the MVol, the Density and
perhaps some other values. But things start to become circular. HSP change largely because the density changes
(HSP are related to Cohesive Energy Density) so if you change HSP and Density you are probably causing
more problems than you are solving.
If your chemicals are mostly solids then the temperature HSP effects are much less certain. It is wise, then, to
select the Contains Solids option to ensure that no HSP temperature corrections take place.
User Parameters
There might be experimental values such as T, pH, pressure … that you want to be included in the QSAR.
Simply enter these into the table, reserving the Temperatures fr the User T column which automatically then
changes the HSP values (as discussed above) in each row according to the value in that column.
Clearing all the columns
If you want to start afresh, just click the X clear button and all 1s will be set to 0s.
Saving, Loading, Copying
Remember to save any interesting QSAR – all the data are stored along with the selection state of the columns.
Even if it’s not the best QSAR it can often be insightful to compare some intermediate QSAR to the one you
think is the final one. You can load any QSAR at any time.
The only exception is that cross terms need to be recalculated each time the QSAR file is loaded.
The same happens with fitted data – you can save the fitted values based on the current QSAR then open an
alternative QSAR, re-calculate the fitted values, save those etc. The idea is to encourage you to explore fitting
space – not to regard any specific QSAR as the last word on a topic.
The data in either table can be put onto the Clipboard by clicking the relevant Clipboard button and pastes very
nicely into, for example, Excel for further data analysis.
The camera button takes an image of the whole setup for inclusion in reports etc.
Normalisation and the Meaning of a QSAR
An obvious question is: “Which parameters in the QSAR are most important?” The answer isn’t simple, but the
software makes it as easy as possible to explore. One way is to double-click on the “1” that selects a parameter
so it is switched to 0, then the QSAR is re-run and the quality of the fit observed. That parameter can be double-
clicked back to 1 and the next parameter can be checked. Another way is to look at the weights of each
parameter. But because parameters span a wide range of values, the absolute weights are meaningless. Clicking
the Normalise button finds the min and max in each column and then scales the values from 0 to 1 across that
range. So now the relative weights of the parameters make sense and the software lists them in descending
(absolute) value. It makes intuitive sense that the parameters with the least weight are the least important. And
in many of the QSARs this is the case. If you double-click a low-weight parameter the effect on the fit is
generally less than double-clicking a high-weight parameter.
The sign of the parameter is also of interest. A +ve sign means a +ve correlation and vice versa.

254
Although other methods exist for judging the relative importance of parameters (e.g. determining the relative
errors in a 1-of-x removal scheme) our experience suggests that the added complexity is not worthwhile. Why?
Our experience is that in terms of the R² quality of fit, removing any parameter can have non-intuitive effects.
Sometimes R² goes up owing to its definition on the basis of the number of parameters used to create it.
Sometimes the remaining parameters seem to re-adjust themselves and do a reasonable job even in the absence
of the previous most important parameter. And when you start to use cross-terms you sometimes find that an
“unimportant” parameter suddenly figures in the most important cross-term.
This brings us to the question of what a QSAR means. No one really knows. Let’s take the HSP terms that work
so well in the examples that follow. What does it mean if there is a strong correlation with 1.4δD + 3.2δP +
0.05δH? According to solubility theory there should be an optimum δD, δP and δH so that they should be
neither too big nor too small. But the QSAR implies that bigger is always better. This is the first big problem
with all QSARs – especially those that assemble a range of rather weird parameters and declare that they are the
most significant. The second big problem is that because we have no (or little) physical understanding of what
the QSAR means we have no way to know in advance whether the predictions it makes are of any value.
Suppose you have fit a range of alcohols and ketones to a QSAR. Will the prediction of the properties of some
esters be of any value? With a theory like HSP we can say in advance that the predictions will, broadly, be
valid. With a QSAR we have no independent way of knowing.
When we started to write the QSAR part of HSPiP, Hiroshi was fully convinced of their value and he uses them
in his real job and also in his university courses. Steven has generally been hostile to QSARs for the reasons
described above. Now we have the QSAR functionality, Steven is truly impressed by what it can achieve, and
the many examples that allow interesting predictions speak for themselves. But users should always remain
sceptical. QSARs are only as good as the people who create them. If you expect their predictions to be accurate
no matter how few data points are used and no matter how far outside the original dataset the predicted
molecule lies, then you can expect many disappointments. If you use them as a tool for exploring which factors
have a strong impact on a data series and use that knowledge along with other evidence to help navigate
through a complex optimisation, then you (as Steven learned) have access to a lot of impressive power.

Examples
To get you started we have included a set of examples; sometimes these are just the raw data, others have both
the raw data and our preferred QSAR. You may well disagree with our fits because you might consider other
parameters to be more important. For example, you might prefer to fit to Refractive Index and Density while we
might prefer to fit to δD and MVol. Because there is a strong correlation between RI and δD and between
Density and MVol the fits are often looking at similar effects.
Many of the examples come, with kind permission, from Dr Yamamoto’s famous Chemical Engineering course
at Yokohama National University. They are intended to stimulate debate with the Masters students about which
parameters should or should not be important for various applications.
HPC
As this is the first example, it is spelt out in some detail. The later examples are described more briefly.
The HPC files in the HSPiP Data QSAR folder use a dataset of chi parameters measured for
HydroxyPropylCellulose in a range of solvents.
To get going immediately, load HPC.hsq which is the whole dataset with a potential QSAR based on δD, δP,
δHDon, δHAcc, MVol and Ovality – a perfectly reasonable first guess. The top row selects the parameters, with
a 1 in each of those parameters and 0 in all the others. If you just want to see the selected parameters, click the
Selected Only option.
Making sure that the Manual Parameters option is selected, click on the QSAR button. You see a relatively
poor fit (R²~0.4). Clearly the working hypothesis was not good in this case.
You have a choice. Try a few more parameters. Or stick with the ones you chose and see if they actually work
via interactions. Select the Cross Terms option and choose 2 or 3 cross terms to be added. Now click the
QSAR button. With 3 cross terms the fit is clearly much better. This is a mathematical certainty. The question is
whether the fit is providing you with good scientific insights or (as discussed shortly) good predictions.

255
Finally, use the automated way to find which, say, 3 parameters (from the Core set, click the About button to
find out which they are or click the Core option) give the best fit. Even with 3 parameters the fit gives an R² >
0.8, using δHAcc, Ovality and the Trouton constant. This is very interesting. The Trouton constant (enthalpy of
vapourisation divided by BPt) is a constant for ideal solvents so the value is showing some deviations from
ideality. How those deviations might affect the solvents’ interactions with HPC is a question that might open up
interesting new insights or, as so often with QSARs, may be some numerical arterfact. One way to test this is to
check the predictive value of the QSAR.
As it happens, there is a much quicker way to test the meaningfulness of the correlation with Trouton. Because
it is obvious that the biggest deviation from a nice fit is acetic acid, temporarily exclude it from the fit by
putting a “-” or “x” into its data column. Re-doing the automatic fit generates a very different set of parameters.
Removing others one by one (restoring the previous chemical) will quickly convince you that these automatic
fits are rather meaningless as the chosen parameters change a lot. This is bad news for those who simply want
the QSAR to do all the hard work. But it’s good news for those who want a deeper understanding of their
system. For almost no extra work (removing a few of the input chemicals) it becomes clear that the fits are
meaningless. Better to discover it with so little effort rather than put in a lot of effort to try to understand why,
say, Trouton was of deep significance.
Before doing that, we need to work out how to start this process. For simplicity we started with a pre-made
QSAR.
Load HPC-Raw.txt. This is just 3 columns: Solvent, SMILES, Data. Again this could be comma-separated or
tab-separated. Now click the Data button to estimate all the parameters. Next choose a few parameters that
seem to you as being likely to be relevant, e.g. δD, δP, δH and MVol and put a 1 into their column header.
You’re now ready to click QSAR. With your first QSAR you’ve joined a large community attempting to make
sense of raw data. What happens next is up to you.
Predictions
Suppose we wanted to know how HPC will behave with a set of alcohols. Create a test file with your chosen
alcohols: Name, SMILES. In the examples, this is the file Alcohols.csv (or the identical Alcohols.txt which is
tab separated). Open the file into the lower table.
Now click the Calc button. The program first estimates all the properties then uses your QSAR to estimate the
data values. Because you have two alcohols that are already in your input data (ethanol and propanol) you can
start to compare predictions with experimental. But more importantly you can do a few tests on a few of those
alcohols and see how good (or otherwise) the predictions are.
If the predictions are good then it’s time to dig deeper into the meaning of the QSAR. If the predictions are bad
then it’s time to attempt alternative QSARs with different parameters.
The point is that HSPiP-QSAR makes this all very easy to do. If you hate Trouton appearing in the automatic
fit, put an X in the Trouton column and try the automated fit again. You (not surprisingly) get a worse fit, but
maybe the new fit gives better predictions or suggests other parameters that could be significant. The program
never laughs at your ideas and doesn’t care if you inspired idea gives a worse fit. Any program can get a good
fit. Only a scientist can take a program and get a fit that is both meaningful and predictive.
LogKOW example
A more familiar dataset than the HPC chi parameter is a set of LogKOW parameters. There are some question
marks about the validity of some of the data so if you are genuinely interested in LogKOW prediction please
curate your own version. If you load LogKOW.txt you first have to click the Data button to fill in all the values
(this can take some 10s of seconds). If you then select the Fit to N parameters and choose just 1 parameter, not
surprisingly the most “predictive” parameter is LogKOW. Adding a few more parameters gives a better fit,
though whether Tc and LogS are meaningful in this context is something for you to decide.
The interesting question is “What makes a good predictor for LogKOW?” So if you want to try the automatic fit,
put an X in the LogKOW column so it is never invoked and see what you can find. Not surprisingly, LogS then
becomes a good predictor, but there’s probably some circularity in this, so put an X in that column too. When I
tried that asking for 2 or more parameters, the predictors (e.g. δHacc and Vc) didn’t make much sense. But

256
trying with just one parameter reveals that MVol does an astonishingly good job of prediction. It’s worth
building on this insight and trying things manually.
It has often been pointed out that MVol is a significant predictor of LogKOW – bigger molecules tend to be less
(relatively) soluble in water because (so it is said) they disrupt more of the water structure. What else might
have an influence? Well, HSP would be a good place to look. Adding δD, δP and δH gives a significant
improvement in fit. An alternative idea is that shape plays a role, and adding Ovality to MVol gives results that
are not too bad. Combining HSP, Ovality and MVol gives a respectable fit – each part playing a distinctive
scientific role. If you do an automatic fit (using Xs to omit LogKOW and LogS) to 5 parameters you get a similar
fit: Y=2.88E-1+2.34E-1.δD - 1.06E-1.δP - 1.76E-1.δHAcc + 2.84E-2.MVol - 3.23E0.Ovality, where the δHAcc
has been substituted for δH. So in this case the manual approach and the automatic approach reach essentially
the same conclusion – which is most satisfying.
Oxygen Solubility
A paper, Takashi Sato et al, Solubility of Oxygen in Organic Solvents and Calculation of the Hansen Solubility
Parameters of Oxygen, Ind. Eng. Chem. Res. 2014, 53, 19331−19337, finds a good correlation between HSP
Distance of various solvents from oxygen and its (Log) solubility in those solvents. The raw data from the
paper are provided and if a manual fit to Distance is made the fit is excellent, using their preferred HSP values
for oxygen of 6.7, 0, 3.8 in the Target box. You can get an equally good fit by deselecting Distance and
choosing, instead, δD, δP and δH.
It is instructive to try out the Fit to N Parameters. Some of the fits are truly excellent but almost certainly
meaningless!
Solubility of C60
Because C60 is completely spherical it is hard for spherical solvents to interact strongly. So in addition to the
HSP parameters you must include Ovality. When Ovality=1 the molecule is spherical and a maximally non-
spherical solvent would have Ovality = 2. As is usual, larger molecules tend to be less good than smaller ones,
so MVol also should be included for a good fit.
Solubility of Graphene Oxide
It is nice to see that the solubility of GO can be fitted with some accuracy to HSP. For the similar reasons as for
C60 you need to include parameters such as MVol and Ovality to get an improved fit.
It is also interesting to look at the Reduce Graphene Oxide data. This is exactly the same GO from the same
research group but reduced to make it closer to graphene. If you use the same fitting parameters, do their
relative sizes make sense to you compared to GO?
Absorption onto Activated Carbon
Here δD, MVol and Ovality are important for (by now) obvious reasons.
The specific data are to do with absorbing VOC (Volatile Organic Compounds) so you can use the QSAR to
predict whether activated carbon would be useful for capturing any solvent in which you happen to be
interested.
If you enter the SMILES for formaldehyde (C=O) into the table below and estimate its absorption onto
activated carbon you will see why those suffering from “Sick House Syndrome” cannot be helped with
activated carbon.
Ibuprofen solubility
Clearly HSP are important for dissolving ibuprofen as are MVol and Ovality. But in this example Temperature
effects are included via an extra column (so the Data column has the solubility and T is provided in the User T
column which is one of the last columns in the table) and the fit therefore allows you to predict the solubility in
another solvent and at your desired temperature. This allows you to find the best solvent for crystallisation. You
need a high solubility at high temperature and low solubility at low temperature.
Nitrocellulose

257
The data are the Mark-Houwink “alpha” parameters for nitrocellulose in a range of solvents. When Mark-
Houwink = 0.5 the solvent is poor, when it is 1.0 it is a good solvent and the polymer is fully stretched, giving a
higher viscosity.
Why are these data important? For those who wish to remove nail polish (based on nitrocellulose) it is very
important. You want a large α for good removal. Acetone is OK at 0.78 but is not good for one’s health. Ethyl
acetate is often used as it has an even better α of 0.9 but is also not so green or safe. So which green/safe
solvents might be used? Once you have found a good QSAR you can put in the SMILES of all your favourite
green solvents into the prediction grid to estimate their α values as a first step towards finding a green nail
polish remover.
We have provided GreenSolvents.txt with a selection of common green(ish) solvents. Click the lower Load
Dataset button to select it then click the Calc button to see the predictions.
Dr Yamamoto also has a data set of Mark-Houwink parameters for polyacrylonitrile (PAN) solubility from
which you can find the best green solvent for making carbon fibre!
Microwave heating
We all know that water is heated very rapidly by microwaves. But what about other solvents? The dataset is of
the temperature reached for a given volume of solvent in a microwave oven for 1min. The absolute values
depend, of course, on the volume and the microwave oven but here we are interested in their relative values. Of
course solvents without a dipole moment heat very little and because of the strong correlation between dipole
moment and δP expect to find that it is an important parameter. But once again, Ovality also is important –
more spherical molecules don’t create much heat when they are excited by microwaves.
If you put the SMILES of lactic acid into the test grid below and calculate the temperature rise you will find
that it is very large. This is very useful to know if you want to make PLA (polylactic acid). Microwaves are an
efficient way of starting the polymerisation which generates water which is also rapidly heated by the
microwaves and will evaporate quickly, helping the equilibrium to move to the polymerised form.
Relative Evaporation Rate (RER)
HSPiP’s Y-MB includes a powerful RER predictor. But you can generate your own via this set of RER data.
HSP and MVol are, of course, important. What else, in your opinion, should be important? Well, use the QSAR
to test your ideas. Then check out the predictions by copying SMILES from HSPiP of some other solvents and
compare your predictions to the experimental values in the Solvent Optimizer.
Cellulose Acetate
Cellulose acetate is very commonly used in inkjet printers. Finding the right solvent is therefore very important
to get the correct (low) viscosity without risk of precipitation of the polymer. As before, the Mark-Houwink α is
used and a good QSAR can be obtained with HSP + MVol + Ovality plus one cross-term. You can then put the
SMILES of your favourite green solvents (e.g. the GreenSolvents.txt we have provided) into the prediction grid
to find which will be the best match for your current solvent. Clearly you also need the right RER so you can
optimise by juggling between the two QSARs
Single Wall Carbon Nanotubes
Using the default .hsq file with HSP, MVol and Ovality and one cross-term the fit to solubility data of CNT is
rather good. But what are the HSP of CNT? Click the X button to remove all the fitting parameters, then select
MVol and Ovality and, near the last column of the table, the Distance option. Make a guess of the HSP of CNT
– say, 19, 2, 2 as they must have lots of δD and relatively small amounts of δP and δH. Try the QSAR and look
at the R² value that judges the quality. Now try other guesses for CNT’s HSP and see how much closer you can
get to a good fit.
Decomposition of t-Butyl Peroxide
Everyone knows that this is an important radical polymerisation initiator set off by temperature. What is less
known is that its decomposition rate is strongly dependent on the solvent. What controls the temperature effect?
Just HSP and MVol are enough to give a near-perfect fit. Why is this? The t-BuO radical is obviously highly
polar so it needs high δP and δH to stabilise it.

258
Vitamin C Solubility
The story here is a simple one. The world uses a lot of vitamin C as an antioxidant and needs to make large
amounts of it. To purify it from solvent needs understanding of its solubility. To shift to a green solvent you
need to make sure that it has at least a reasonable solubility values. The QSAR (again it’s HSP plus MVol plus
Ovality and some cross terms) allows you to put your favourite green chemical SMILES (e.g.
GreenSolvents.txt) into the prediction table to see if they can attain the required solubility.
Flash Point
This is rather easy. BPt on its own is quite good, and adding MVol or MWt makes the fit near-perfect. If you
have a larger dataset containing halogenated compounds you will find that the fit to BPt is rather bad. You then
need to add heat of formation to get a good fit.
HPLC of Organic Acids
This dataset contains the HPLC retention times for a set of organic acids. It is easy to find that the standard HSP
+ MVol + Ovality gives a good fit. The fit is somewhat better if the Log10 option is selected.
HPLC of Antioxidants
Just to show that the HPLC principle is general, the fit to retention times of some common antioxidants is
similar.
Oral Absorption of Drugs
This is a very different story. The QSAR using the standard HSP set is not at all good. But using just δH +
MVol + Ovality with 1 cross-term gives a good fit. Clearly absorption in the acidic stomach is dominated by H-
bonding terms.
Skin Penetration
A typical activity in the world of delivery of chemicals through the skin is to predict the (log) permeability
coefficient, LogKp. This is not as useful as it sounds because we are more interested in Jmax, the maximum
flux through the skin which in turn depends on the solubility of the chemical in water. But that’s another story.
Assuming you want to predict LogKp then from the same dataset we have 3 .hsq files.
The first uses HSP + MVol + Ovality. The fit isn’t great. The second adds logKow and the fit looks great. The
third ignores all the HSP and just uses MVol, LogKow and LogS – and provides a great fit.
The point is that you can find these sorts of things for yourself very easily. For example, by using the Fit to 2
parameters you quickly find the MVol and LogKow correlation. This is the classic Potts and Guy correlation
which has proven to be so misleading over the past decades.
Hexane Water Extraction
Many of the chemicals here don’t much like to be in either hexane or water so simple solubility ideas don’t
seem to work. Using the Fit to N option a fit with 5 parameters gave some suggestions and by systematically
turning the 5 parameters on and off it became clear that just 3 parameters are all that is needed: δH, MVol (or
logKow which is mostly the same thing) and, surprisingly, MPt. We don’t know why it is important, but it
certainly helps.
Eicosane Solubility
Not many of us are very interested in dissolving fluorocarbons in eicosane, but if you are then this QSAR
reveals that HSP + MVol + Ovality do a good job, especially with 1 extra cross-term.
Oleic Acid Solubility
Again the standard set do a good job of fitting.
Paracetamol Solubility
It would be nice to report that HSP do a great job of fitting these logSolubility data. But they don’t. From Fit to
3 parameters the answer is density, LogS and RI. We don’t know why this is the case.
Aspirin Solubility

259
The aspirin data are non-log so it is important to select the Log10 option for the fit to HSP. Normally MVol is
required but these data are mole fraction so you do not need to include MVol. However, Ovality is still very
important for a good fit.
Dioxins Toxicity
There is a lot of public anxiety about dioxins, the chlorinated cyclic molecules that can be produced both
naturally and in some industrial processes and poor-quality incinerators. The toxicity of individual dioxins
varies widely and can be predicted rather well using HSP, MVol and Ovality as you will see in the Dioxins
dataset. The Log(TEF) is the log of the Toxic Equivalent Factor where more negative means less toxic.
Diffusion Coefficients in Water
Obviously diffusion coefficient in any solvent is going to depend on shape and size. The data of diffusion
coefficients in water confirms this, with MWt (rather than MVol) being representative of size and Ovality
representing shape. If you add a cross-term then Ovality² is included and the fit is excellent.
Iontophoretic Drug Delivery through Skin
By applying a voltage across the skin various changes take place which can speed up the delivery of drugs
through the skin. The rate of delivery in this TransDermal Delivery System (TDDS) can be modelled
effectively via HSP, MVol and Ovality. Contrary to popular belief, the correlation with LogKow is non-
existent, even if MVol and Ovality are included.
Activity Coefficient in Ethanol
It is no surprise that the activity coefficient (expressed as the Margules infinite dilute parameter, M12)
correlates very will with HSP, MVol and Ovality. It is interesting to try Distance (+ MVol). The fit is OK when
the HSP are set to [15, 9, 19], rather close to the official values of [15.8, 8.8, 19.4], though worryingly the fit
improves if δD is taken as low as possible!
Polymer Values
As you know, Y-MB can also use Polymer SMILES where X signifies the polymer chain. So polyethylene is
XCCX and PMMA is XCC(C(=O)O)(C)X. Many of the calculated values are irrelevant – such as BPt. But
many predictions are OK.
For example, load Polymer-RI.txt, wait some time when you click Data as there are a lot of SMILES to convert,
then select the RI parameter and look at the QSAR results. Given the wide range of polymers, some of them
rather strange, the predictions aren’t too bad.
The prediction of Oxygen permeability is quite impressive using HSP, MVol and Ovality. Although MVol and
Ovality of the monomer units might not strictly make sense, as with so many of these examples, the quality of
the fit is significantly improved when they are used. An alternative approach is to use the HSP Distance as the
single parameter. Earlier in the eBook it was mentioned that O2 has HSP of [14.7, 0, 0]. If you enter these
values and use just the Distance parameter the fit is acceptable. This is a reminder of why PE (Distance rather
small) is such a poor O2 barrier, even though it is highly crystalline, while EVOH, which is not a great
polymer, is a superb O2 barrier as the Distance is so large.
Rat toxicity data
People spend a lot of time trying to predict toxicity based on QSARs of tox data. There are at least two
problems with this. First, the tox action of different chemicals (or chemical classes) might be very different, e.g.
one acting on the liver, the other on the brain, so any QSAR is likely to be misleading. The second is that the
datasets themselves contain lots of errors. When we first tested and EPA dataset of the LD50 of alcohols and
phenols, one molecule, hexadecanol, was far outside the correlation. Checking back on other data it was clear
that the value in the EPA dataset was wrong (for whatever reason) and with the correct value from other
sources the fit was fine.
Another favourite aspect of tox QSARs is to fit to LogKow. As mentioned elsewhere, LogKow is strongly
correlated to MWt or MVol and may not be of great significance in its own right. You can test this out with the
alcohol+phenol dataset, using the trick of clicking on the column header to get an instant graph of experimental

260
value versus parameter value. If you try this with LogKow you find a V-shaped profile, with medium-sized
molecules being more toxic. If you then click on MWt or MVol you see the same profile.
Once again, HSP + MVol + Ovality do a good job of fitting to the data.
Incidentally, both the EPA and REACH refer to Hiroshi’s Pirika site as a source of LogKow estimated values.
You will be pleased to know that the same prediction algorithm is used within HSPiP and the QSAR!
Silk Shrinkage
Silk shrinks when it is placed in organic solvents. What factors control the shrinkage? You can find out for
yourself using this dataset of shrinkage (in units of mg/diameter, whatever that means). The original authors
correlated the results with Hildebrand parameters. So you can see if HSP can do better, via two routes. The first
is to fit to a single Distance parameter. When we tried it we got a good fit ~[14, 12, 16]. The second is to export
to HSD and use the fit to data options in GA mode, and also to create a Sphere using a cut-off value (we used
130) to decide between “good” and “bad”. The results agreed with the QSAR values. You might like to remove
the methanol data point. In the original paper the shrinkage in methanol followed a very untypical curve so the
point might not be valid.
Endocrine Disruptors
The hot topic of endocrine disruptors has created its own class of QSARS. So it was interesting to take a dataset
of 110 steroids and look-alikes (including bisphenol-A) and see what happened. The data are Log(RBA)
(relative binding affinity) and they show that bisphenol-A is 5 orders of magnitude less potent than DES, and
presumably much less potent than many of the steroids in “healthy” foods such as soya and ginseng
We had expected some correlation with HSP values, but in fact there is only one factor which shows any
reasonable correlation – that is MVol. This is readily seen using the “click on the column header” trick which
plots the raw data versus the single parameter. MVol is clearly the best single determinant. LogKow and logS
both show a poor correlation. Using the Fit to N parameter trick does not produce anything significantly better
than MVol.
So there are some datasets that clearly defeat our QSAR. In fact, the data are taken from a paper which uses 3D
structures to be able to predict with some success.

Exporting to HSPiP
Traditionally, HSP’s Sphere has been fitted to a serious of “good” and “bad” solvents scored as 1’s or 0’s. But
HSPiP allows fitting of numeric data to a Sphere, via the Genetic Algorithm (GA) option.
If you want to test your QSAR data against a classic Sphere fit, click the export to HSP Sphere icon. If you
have the full HSPiP+QSAR version then the data automatically open in the main form ready for you to analyse.

261
Chapter 37 SFBox-FE
Many formulations contain a mix of particles, polymers and solvents. Sometimes the polymers are meant to be
free, many times they are meant to be dispersants that sit around the particles protecting them from crashing out
when the particles get too close to each other.
For most of us our knowledge of these systems was built around some general ideas of “steric stabilization”
taken from DLVO theory (in this chapter we are not discussing charged particles and polymers that give charge
stabilization), with some general ideas of what that might mean, mixed with the uncertainty of how best to
choose the right stabilizer and be confident that it won’t cause problems when we change things in a
formulation by, for example, adding another polymer required for different aspects of the formulation.
One rule for these systems established decades ago by Charles was to choose a solvent that was near borderline
for the (simple) stabilizing polymer – otherwise the solvent might remove it from the particle.
Unknown to most of us was the existence of an excellent theory that was originally developed in the 1970s and
continuously refined up to the present, largely via a team at Wageningen U in the Netherlands. The original
developers were Scheutjens and Fleer so we often call it Scheutjens-Fleer (SF) theory. They called it Self-
Consistent Field (SCF) theory. Feel free to use either name, but with the caution that for non-Dutch speakers
it’s rather hard to pronounce Scheutjens (Fleer rhymes with Beer, so that’s easy). If you want a simplistic
approximation, then Skirt-yuns will do.
SF theory is easy to describe (and has been done especially well by Prof Terence Cosgrove of U Bristol who
has been so helpful in the journey to creating the capability described here) but impossibly hard for any of us to
use without some powerful software. Fortunately, Prof Frans Leermakers at Wageningen U has written SFBox
which, as the name implies, is a box into which you post a SF problem and out of which an answer emerges.
How the magic works inside the box need not interest us. Unfortunately, SFBox itself is so amazingly powerful
that it’s near impossible for ordinary formulators to use. Fortunately, again, Prof Leermakers took the time and
trouble to explain how to write a front end called SFBox-FE to become part of HSPiP. His considerable
generosity (and patience!) are warmly acknowledged and SFBox is used with his kind permission.
So what is SF Theory?
Imagine a regular lattice in space. Yes, particles, polymers and solvents don’t arrange themselves in neat
lattices, but it turns out that making such an assumption is perfectly respectable and behind lots of good
theories, including Flory-Huggins and HSP.
Now define one edge of that lattice as the particle. No solvent or polymer can do anything other than sit next to
it. Now place your chosen volume fraction, φ, (everything is done in volume fractions) of polymer anywhere
that seems sensible within the lattice, with the obvious constraint that the individual beads (monomer units) of
the polymer are connected. All the other lattice sites are filled with solvent.
Now calculate the energy of that system. For this you need to know how happy the polymer is to be next to the
particle and/or to be surrounded by solvent molecules. We already know the “solvent molecules” bit. The
Flory-Huggins χ parameter tells us that when it’s 0 the polymer and solvent are completely compatible, when
it’s 0.5 they’re borderline and above 0.5 they are basically incompatible. We can do a similar trick with χS (we
can say that S stands for Silberberg who introduced the idea) which describes the relative affinities of the
polymer and of the solvent for the particle. If the polymer is much happier on the particle than is the solvent
then, for obscure historical reasons, χS = -1.5. If they are equally happy (or unhappy) next to the particle, χS = 0.
From just these two parameters (and, of course, the starting volume fractions of particles and polymer) it’s
possible to work out the energy of the system. Now do some sensible adjustments of the positions and
recalculate the energy. If it reduces then carry on adjusting in that direction, if not, try something different.
Keep going till the energy is minimized.
Via some powerful algorithms, tricks and techniques developed over the years, this minimization process can
take a fraction of a second for the sorts of problems that interest us. The result is a self-consistent field: the
volume fractions are consistent with the energies and the energies are consistent with the volume fractions.
The technique is entirely general. If I want to use a di-block polymer then all I have to do is define the relative
number of A and B units and then χA, χSA, χB, χSB, χA-B (that’s the χ between the two polymer segments) so that
the energies can all be worked out. For a typical modern hyperdispersant which is a comb polymer we just
262
specify the main A chain and attach a chosen number of B comb teeth at various points along the length. We
can even have two free polymers, A and B, where A might represent a typical “stabilizer” and B a different
polymer added for some other reason to the formulation. As we all know, it’s the unexpected behaviour of these
B polymers when we add them to our otherwise stable particle dispersions that can cause problems such as
depletion or bridging flocculation. We can also just graft the polymer onto the particle, giving the ultimate in
stabilization if we are sensible.
What are the polymer chains doing?

We tend to draw polymers on particles as if they are like the ones on the left – nicely sticking out into the
solvent. Of course if the solvent is too good, they are dragged off the particle as on the right. But in fact they are
mostly like the ones in the middle, coiled up like they are in solution. When we do SF calculations we get
values, at each distance from the particle surface, for the relevant fractions of “trains, loops and tails” plus the
amount of free polymer. The term “trains” was introduced by Scheutjens to describe those bits of polymer
running along the surface like trains on a track. It doesn’t seem a helpful term, but we’re stuck with it:

Here's a graph of a typical output of SFBox-FE showing the different portions for this specific
polymer/particle/solvent combination:

This is the Poly200 example file

The x-axis is in “lattice units”. To translate into real units you can choose the option in the software to replace
them with the assumption that each unit is 0.3nm, so in this example the tails extend out to ~3nm and the graph
goes out to 6nm. Notice that the bulk of the polymer (trains and loops) is within the 5 lattice unit regime, just
1.5nm. Our polymer shells are usually much thinner (and denser) than we typically imagine.
263
The first thing you can do is play with the polymer/solvent χ parameter. As you decrease it from its medium
value of 0.25, not a lot happens – the small tails portion moves slightly outwards (of course) and the trains and
loops move inwards. If you increase it towards 0.5 then you see lots more polymer on the particle, which seems
a good thing in simple terms of protecting the particle, and if you go over 0.5 then very little free polymer is
found in the solvent (not surprisingly) and a lot around the polymer. But going to a higher χ value is a disaster.
As we shall see later, we can calculate particle-particle interactions and they become very strong (crashing out)
once χ reaches 0.5, a fact known from basic DLVO theory, but investigated more fully via SF theory.
Next is a comb version. Instead of the single polymer containing 200 units, this comb is an A of 100 units with
4 B comb teeth of 50 units, 300 units in total so 50% larger than the previous example:

This is the Comb300 example file

Now the green represents the B unit, not the tails, and you can see that it’s working as we imagine a comb to be
– a tight A around the particle with plenty of B sticking out. Because the combs are small they don’t stick out
much but, as we shall see, they contain a lot of tails and tails are good for stability.
These are each examples of sample files that come with SFBox-FE and which you can load from the Examples
sub-folder of the SFBox folder within the HSPiP Data folder. By exploring the other examples, and by playing
around with the sliders you’ll quickly build your intuition about what your polymer/particle system is doing and
where the polymer components are residing.
There are lots of options for plotting – for example with or without log axes, with or without auto-scaled axes
etc. Academics love log plots because you see lots of detail which looks exciting and important – till you realise
that it’s volume fractions of 10-4 which are largely irrelevant to formulators.
Particle-particle stability
What we’ve just achieved with SFBox-FE is exceptional – most of us have previously not had a chance to see
what our polymers and dispersants might be doing. But SF theory is sufficiently powerful that we can go to the
next step and see what happens to the system energy (in units of kT where 1kT is the random thermal energy of
your system) as the particles come together. If you load the DiBlock example and click the Interparticle button
then you find an energy landscape like the following. Note that these calculations take a few seconds each so
you don’t have the luxury of sliding sliders to see their live impact on stability. Note too that you may need to
fiddle with some numeric parameters, described later:

264
The energy is defined as 0 when far apart (off the scale to the right of this screen shot) and in this case climbs to
a massive 100,000 kT. There’s no need to worry about the exact numbers – the point is that this system will be
highly stable. More interesting is a comparison of two two-polymer system examples, the 2Poly190 and
2Poly200 example files.
The first (A90, B100) looks like a classic stable system with a huge 100kT barrier at 1.5nm.
The second (A100, B100) simply has an A polymer that’s 10% longer. Our natural instinct is that bigger is
better. But in this case “bigger” means “more loops” and, contrary to our expectations, “more loops” means
“more bridging flocculation”. It turns out that for fundamental thermodynamic reasons, tails are repelling and
loops are attracting. So as we make a polymer longer, the relative amount of tail decreases, and in this case the
system becomes unstable.

If you play with these examples you can also see the obvious fact that changing the solvent (by changing the χ
parameters) has a big effect on the solubility. Indeed, for these specific examples, even a small change will
make a big difference, again reflecting what we find in real life that an apparently stable system becomes
unstable after some seemingly minor change to the formulation.
This links to the Poly200 example – as you increase χ it looks as though you get more protective polymer
around your particle, but too much of anything is bad, so around χ=0.5 the system changes from stable to
unstable. In the screen shots, a 120kT barrier falls to 6kT when χ changes from 0.25 to 0.508. At 0.509 the the
system is unstable:

265
The χ parameters
Because the calculations depend on the χ parameters, there’s a natural fit to HSPiP. The SF community has
always assumed that the χ and χS parameters were somehow known, and maybe they were for the relatively few
simple cases they explored. In any case they tended to use 0 and 0.5 for the χ values and -1.5 or 0 for the χS
parameter. How do we get them for our real-world systems?
The χ values are simple. Historically, if the HSP Distance D between the polymer and the solvent is known,
then we combine it plus the solvent’s MVol via χ = MVolD²/(4RT). But this assumes a radius, R0 of 8. If we
have a different measured value we can plausibly say:
χ = 2MVolD²/(R0RT).
Knowing DS from particle to solvent and the particle’s measured HSP radius RS then we can create a χSol
between particle and solvent via:
χSol =2MVolDS²/(RSRT).
With less certainty, given DP from particle to polymer and the MVol of the polymer’s monomer unit, and using
RP the smaller of the two radii Ro and RS, we can create χPol between polymer and particle via:
χPol =2MVolDP²/(RPRT).
Given that we now have χ values for the solvent with respect to the particle and of the polymer with respect to
the polymer, χS which is the relative attraction of the polymer and of the solvent to the particle can plausibly be
calculated via
χS = 3(χPol-χSol)
giving us -1.5 with a perfect polymer and a borderline solvent and 0 when they are equally attracted. Users will
find this -1.5 to 0 range rather odd. But that’s what Silberberg chose and it’s too late to change it.
The χA-B parameter between the two polymers is calculated via a similar route, using the smaller of the radii in
the calculation and an average of their respective MVols.

Here’s the χ Calculator within SFBox-FE. There is no way that these simple calculations can be accurate, but
they are certainly better than nothing and emphasise the importance of measuring the HSP values of our
polymers and of our particles. It’s also true that the calculations show how hard it can be to balance the
conflicting requirements of different polymer/particle, polymer/solvent and particle/solvent pairs. It is all too

266
easy to set up a bunch of χ and χs values in the model, but it’s very hard to achieve them in the real world
because of near-inevitable conflicts.
The link to NMR
Many users of HSPiP have become familiar with the measurement of the HSP values of particles via the
observation and scoring of particle behaviour in different solvents or getting precise measurements of
sedimentation rates using sophisticated centrifuges. Indeed, HSPiP’s Optimal Binary Fit capability for
determing the best cut-off point between “good” and “bad” measured values was first suggested by the team
using the LUMiSizer centrifuge system for accurate measurements of particle HSP values.
Unknown to most of us was the fact that the NMR relaxation rate of a solvent containing small volume
fractions of particles depends strongly on how well the solvent associates with the particle. Close association
means that the solvent molecule’s degrees of freedom are greatly reduced and this decreases the NMR
relaxation time or, putting it the other way, increases its relaxation rate. Solvents that hardly associate with the
particle have only small changes in their relaxation rates.
This has been known for decades by the academic community and there are plenty of fascinating papers which
also include the effects of polymers on the particle – which are more complex because the polymer chains can
(a) capture the solvent close to the particle and give a large relaxation rate or (b) the flexible tails can interact
with the solvent rather far from the particle, giving minimal change to the relaxation rate.
By no coincidence, this work, much of which was done at U Bristol, has usually been analyzed via SF theory,
partly because of historically strong links between U Bristol and Wageningen U.
A more recent development is of desk-top NMR solvent relaxation rate machines. So instead of tying up a
sophisticated NMR machine more usually used for classic NMR measurements, formulators can make their
samples and measure them within minutes on the desk-top machine. Work in both Japan and the USA has
shown that this is an excellent way to measure HSP values, using the OB-Fit technique to distinguish between
good (high relaxation rate) and bad (low relaxation rate) solvents.
This bringing together of a new(ish) measurement technique in the context of a new capability of analysing the
results in terms of SF theory is another example of how the universality of science makes it so powerful.
Some numeric issues
Like all simulations, things can go wrong. If you get strange results they might be due to the numerics blowing
up. SFBox is amazingly good at avoiding such problems, but they still exist. Two numerical parameters can be
adjusted to see if the problem goes away. The first is Δmax which is the fastest possible change permitted to the
calculations. A large value means that they can go faster, but they might fall down a numerical hole. Smaller
values may be slower but more likely to not head in the wrong direction. A value of 0.5 would be considered
“large” and 0.01 would definitely be very small. The second is Tolerance – which is how accurate the
calculation needs to be before stopping. Normally this can be 1E-7, (which would have a value of 7 in the box)
but this can often go wrong with more complex problems. 1E-11 (11 in the box) is a good compromise.
Hopefully you won’t need to go to 1E-14 accuracy but an option to go to 14 is possible.
A work in progress
The plan to add SFBox-FE to HSPiP was meant to coincide with a set of experiments on representative
formulations to demonstrate the power of the new capability. The major disruptions of 2020 made those plans
impossible, so the approach is not yet formulator-tested. However, given the 30+ years of academic work on SF
theory we can be confident that the general approach is valid and that combining common sense and experience
with the theory is going to be a productive way forward for the formulation community.
A constant, and welcome, stream of questions, observations and suggestions has helped to extend HSPiP’s
capabilities over the years. We will certainly be listening carefully to our users’ experiences with SFBox-FE
and doing our best to upgrade its capabilities.
For those interested in SFBox, the “SFBox Manual” is provided as a pdf in the SFBox folder. It was kindly
provided by Prof Leermakers and gives you a glimpse of the full power of SF theory. Because you have the
template files used for standard cases, if you want to do your own comman line calculations, use one of the
templates as your starting point (remember to save it with a different name, e.g. MyInput.txt), just type
267
SFBox -s MyInput.txt
and, if you are like me, you will read a cascade of helpful error comments (the error trapping in the software is
awesome) that will quickly let you fix the inputs so eventually you get a success message and an output file.
Interpreting the outputs is not so hard if you put the files into Excel so you can see the columns and get an idea
of what is going on.

268
Chapter 38 A Short History of the Hansen Solubility Parameters

Figure 38-1 Where it all began: the initial HSP values for the 88 solvents were determined the hard way on this equipment in
Hansen’s lab. δD is in the direction of the rods which had rings at regular intervals. δD = 14.9 and δP= δH=0 is at the lower
foremost corner where there is a white label for n-hexane [14.9, 0, 0]. Magnets with wires glued to them were used to plot data
for the provisional values for the three parameters using colored beads.

The Main Track


I was born in Louisville, Kentucky. I graduated from the University of Louisville, Speed Scientific School with
a B.Ch.E in 1961. Wanting to continue for a doctorate, I was in the process of working for a Ph.D. at the
University of Wisconsin, Madison, having gotten a Masters degree, but wanting to take a year in Denmark
before having to “settle down” with the advanced degree. My father came from Denmark, arriving in the US in
1929, and my mother’s family came to the US in the late 1800’s. Not really knowing what had been done to
accommodate a useful study, I arrived in Denmark to find that I was able to stay not one year, but two years,
provided I wrote a thesis to obtain a degree then called “teknisk licentiat”. I accepted and delivered the thesis in
exactly 24 months as planned. I knew from earlier correspondence that I could either work on an automatic
process control project or on a question in the coatings industry related to why solvent is retained in polymer
films for years. I chose the latter.
When I was finishing the work for the technical licentiate degree in 1964 [1] there were a couple of Master’s
candidates working as a team on the use of solubility parameters in the coatings industry at the Central
Research Laboratory of the Danish Paint and Varnish Industry. I advised them occasionally and this lead
indirectly to the development of what are now called Hansen solubility parameters. I was formally associated
with the Technical University of Denmark (at that time called Den polytekniske Læreanstalt) where Prof.
Anders Björkman arranged for my stay. The actual work was done at the above laboratory led by Mr. Hans
Kristian Raaschou Nielsen, in a rather small room with a slanting ceiling on the uppermost floor at Odensegade
14, Copenhagen Ø.
As stated above, my licentiate thesis was to explain how solvent could be retained in coatings for many years. It
was thought that this was caused by hydrogen bonding. I showed solvent was retained because of very low
diffusion coefficients. It is especially difficult to get through the surface of a coating where there is essentially
no solvent and diffusion coefficients are very low. The diffusion controlled phase followed a phase where most
269
of the solvent initially present freely evaporated. In the meantime it was necessary to account for the hydrogen
bonding capability of the test solvents, because of what was believed at the time. The work of Harry Burrell [2]
provided the basis for selecting test solvents. He qualitatively ranked a number of solvents according to weak,
moderate, or strong hydrogen bonding. The licentiate thesis did not treat solubility parameters as such, dealing
only with diffusion and film drying, since it was not hydrogen bonding or the solubility parameter that had
anything to do with the problem, other than allowing solution in the first place. There was, however, established
a battery of solvents and knowledge about solubility parameters at the laboratory, and the Master’s candidates
were to further the development of this area.
An article by Blanks and Prausnitz appeared [3] and I advised the students to make use of the new method of
dividing the Hildebrand parameter into two parts, one for dispersion interactions and one for what was called
“polar” interactions. They did not do so, having already gotten into their study and they needed to finish as
planned, being short on time. After I turned in my licentiate thesis for evaluation, I looked at their experimental
data using two dimensional plots of the dispersion parameter versus the new “polar parameter” as described by
Blanks and Prausnitz. I could see there were well-defined regions of solubility on the plots. For some polymers
there were bad solvents within the good region of the 2D plots. For other polymers these were the good
solvents. The other ones had now become bad. The one group was largely alcohols, glycols, and ether alcohols,
with the other being ketones, acetates, etc. It seemed logical to use a third dimension, pushing the bad solvents
into another dimension, and this was the basis for the original terminology “The Three Dimensional Solubility
Parameter” that was used in the original publications in 1967 [4-7]. I followed the rule that the sum of energies
in the (now) three partial parameters had to equal the total reflected by the Hildebrand parameter, recognizing
that Blanks and Prausnitz were correct as far as they had gone. No one up to that point had recognized that the
hydrogen bonding effects were included along with the polar and dispersion effects within the Hildebrand
parameter itself. The Hildebrand parameter is based solely on the total cohesive energy (density) as measured
quantitatively by the latent heat of vaporization (minus RT). Hydrogen bonding was considered too special to
allow such a simple approach as the HSP division of the total cohesion energy into dispersion, polar, and
hydrogen bonding contributions. Efforts prior to Blanks and Prausnitz had used the Hildebrand parameter
together with some more or less empirical hydrogen bonding parameter, for example, in efforts to make useful
solubility plots. Barton’s handbooks review these earlier attempts in an exemplary manner, and as usual I refer
to his handbooks for these developments rather than repeating their content [8,9].
Prior to the public defense of the licentiate thesis, I visited the US, returning to Denmark for the big day. While
in the US I visited the Univ. of Wisconsin to try to establish a continuation of the earlier studies based on the
promising work on solubility parameters that had become obvious to me, at least. Professors Ferry (of WLF
equation fame), DiBenedetto, and Crosby, all would accept me, but only working on projects for which they
already had funding. After return to Denmark for the public defense, Prof. Björkman urged me to stay on to
complete a Danish dr. techn. (similar to D.Sc.). I accepted, and found a room with a relative, rather than in the
student dormitory where I also got indoctrinated into the student life of the time in Denmark. 1967 was a big
year. My father had to come to Denmark twice, once for a wedding and once for the public defense of the dr.
techn. thesis, an event he could not quite believe would happen. He himself was a chemical engineering
graduate from the same school, and knew that not that many got so far. It is my belief that because of the
privileges provided by Prof. Björkman (just do it at your own speed), that I am the youngest (29) to ever have
been awarded this degree. The requirements of the technical doctorate are that one presents and defends his or
her own ideas in a written publication. This must then be defended in a very formal (coat and tails) public event
with official opponents that must not last longer than 6 hours. There was newspaper coverage with an audience
of 125, filling every seat in the auditorium. My official opponents were Prof. Anders Björkman (polymers),
Prof. Bengt Rånby (polymers), and Prof. Jørgen Koefoed (physical chemistry). The event lasted about 4 hours.
As an indication of the iconoclastic nature of this thesis, Prof. Koefoed challenged in advance that I could not
assign the three parameters to formamide, and that the mixture of equal molar amounts of chloroform and
acetone must give deviations. I then proceeded to assign the three parameters to formamide by calculation and
experiment, and tried to experimentally test all of my test solutes in the acetone/chloroform mixture. There
were no errors in the predictions. The thesis was accepted.
I initially had a three dimensional model as shown in the opening figure made with metal rods at equal spacing
supported by clear poly(methyl methacrylate) sides. There were rings on the rods at uniform intervals. The D
parameter was in the direction of the rods, varying from 7 to 10 in the old units (cal/cc)½. Each of what
ultimately became about 90 solvents was represented by a given magnet to which a wire was glued so that
270
given points in the space could be labeled. A small green bead was place on the tip of the wire for a good
solvent and a small red one was used for a bad solvent. One could thus make a 3D solubility plot for each of the
33 solutes. These were mainly polymers chosen to potentially have such widely different solubility properties
as possible. If a given solvent seemed to be giving consistent errors, its P and H parameters were adjusted,
keeping the D parameter constant, and the magnet with wire tip was moved. This trial and error procedure
clearly showed the value of the three dimensional methodology. Tests were made with mixtures of non-
solvents. If such a mixture dissolved a given solute, the solvents had to be on opposite sides of the region of
solubility. It they did not they were on the same side. This method was used to confirm the parameters for as
many of the solvents as was reasonable. I then took a solvent and willfully placed it on the wrong side of the
system and started all over. It became obvious that the system was inverting, so it was concluded that these
numbers were reasonably good, but would probably need revision at some time. Publications were prepared.
The first revision came rather quickly in 1967 from the insight of a colleague at the Danish laboratory, Klemen
Skaarup. He found the Böttcher equation for the polar parameter, did a lot of calculations, and plotting, and the
initial values were revised accordingly. The changes involved in these revisions were not that great as can be
seen from the earlier publications. Mr. Skaarup was also responsible for the first use of the “4” in the key
equation of the methodology, finding this would give spheres rather than spheroids for the solubility regions.
The “4” was generally considered as empirical for many years thereafter.
These “three dimensional” concepts were reported in three articles in the Journal of Paint Technology and in
the dr. techn. thesis, which also included an expanded section on diffusion in polymers and film formation, in
1967 [4-7]. I have reviewed the dr. techn. thesis many times, and have found nothing wrong with it yet. It can
be found as a PDF file on my website www.hansen-solubility.com.
Just prior to the public defense of the dr. techn. thesis I corresponded with Prof. Prausnitz to see whether the
studies could be continued with him. The response was that there was no funding. I then took a job at the PPG
Industries Research and Development Center in the Pittsburgh area. These eight years were very rewarding
with a remarkably inspiring leadership “Making Science Useful” (Dr. Howard Gerhard and Dr. Marco
Wismer). There were many confirmations that the methodology could be used to great advantage in practical
situations. I was popular in the purchasing department during the solvent crisis (oil crisis) where one had to buy
whatever was available on the spot. I could immediately on the phone confirm whether or not a given solvent
could be used and the usual testing was not done. Shiploads of solvent were bought on this basis only.
Dr. Alan Beerbower at Esso (now Exxon) was just waiting for me, as he said it himself, and took up the
developments in the 1967 publications in many areas as can be seen in our article in the Encyclopedia of
Chemical Technology [10] and in his many publications on a variety of topics, often related to surfaces,
lubrication, and surfactant behavior, for example in [11,12]. He developed group contributions, adding to what
was known at that time (citing Fedors), that I used and reported in the handbooks [13,14]. It was Dr. Beerbower
who first used the term Hansen plot as far as I know. Dr. Beerbower authored a brochure for Esso that appeared
in 1970 entitled “Parameters of Solubility”. Here is the cover of that handbook and inside, Beerbower’s
reference to the Hansen principle:

271
Figure 38-2 Perhaps the first reference to Hansen (component) parameters in the literature from Beerbower’s 1970 handbook
and a gratifying confirmation of 97% accuracy for prediction of solubility.

I have put one of his figures in the Handbooks [13,14]. In the Second Edition this is on page 338. This figure
also appeared in Beerbower’s publications but I got it only as a personal communication. Sometime after the
appearance of the article in the Encyclopedia of Chemical Technology [10] in 1971, where the terminology was
not changed, probably because I did not use it, Hansen (solubility/cohesion) parameters replaced the “three
dimensional” terminology on a more general basis. Van Krevelen did not like three dimensional systems, but
did the group contributions for the “solubility parameters” anyway in his “Properties of Polymers” from 1975,
so the change in terminology was not complete at this point in time. Barton’ handbook in 1983 used the
Hansen parameter terminology as cited below. I have never had contact with Van Krevelen. A US Coast Guard
project in 1988-9 studying chemical protective clothing brought me back on track in terms of adding a
significant number of solvents to the database. I was to find solvents for testing that could permeate a PTFE
body suit after having established a correlation for those solvents that had been tested. As it turned out there
were indeed quite a few solvents that permeated the PTFE suit that were characterized by molar volumes less
than about 60 cc/mole and monomers with terminal double bonds that could be somewhat larger [13,14] (see
the figure on page 247 of the second edition of the handbook). I actually initially had a technician looking at the
published Van Krevelen group contribution approach early in this project, before realizing that I had to do it
myself with the Beerbower group contributions that I had gotten as a private communication. The Van
Krevelen and Hoy approaches are now outdated, being surpassed by the work of Stefanis and Panayiotou (See
for example Chapter 3 in the Second edition of the handbook or their other publications. HSP estimates by the
S-P statistical thermodynamics methodology are also included in HSPiP). Even this has been outdated very
recently by the work of Dr. YAMAMOTO Hiroshi in the HSPiP where it is called the Y-MB method for
Yamamoto Molecular Breaking. Both Hiroshi and I independently found that one did much better when using
larger “groups” for the still larger molecules, even to the extent of directly using the existing HSP of
multifunctional molecules as a whole as a single group.
The superiority of modern computers that are capable of working with huge databases to generate correlations
with rapidity and flexibility stands in contrast to what was done earlier. The first calculations for dividing the
latent heats into partial solubility parameters were done using a slide rule. Indeed there were computers that
could have helped with this at the time, but this cost money, and the data were very scattered in the literature.
The first computer program to calculate the HSP spheres from experimental data was probably that at PPG
Industries around 1968. My lab there was set up to routinely determine the experimental data that helped to
optimize solvents and to predict compatibility. Safety and the environment were emphasized. A similar
program was available at the single, central computer of the Scandinavian Paint and Printing Ink Research
Institute, and later on my son, Kristian, wrote the same type of program for use at our home on a Commodore
64. This typically took about 20-30 minutes to calculate the HSP sphere from data on approximately 40
solvents. Much of the data in the handbooks was done on this computer.
272
I left PPG in 1976 to become director of the Scandinavian Paint and Printing Ink Research Institute, being
invited to do so largely at the suggestion of the Swedish participants (Prof. Bengt Rånby, Prof. Sven Brohult).
This was a Danish-Swedish organization at the time, but when I left 10 years later, Finland and Norway were
also part of the Nordic cooperation. These 10 years also led to further progress and development of knowledge
in the area, mostly in the further characterization of materials and from applications in industry. Research as
such was not permitted at my final place of employment, FORCE Technology, so the developments were not as
extensive as what might have been expected. I did manage to write the first edition of the handbook (at home)
[13], and to search for and find what I believe to be theoretical justification for the “4” in the key HSP equation.
The Prigogine corresponding states theory of polymer solutions has the “4” in the first term of the free energy
equation, but only when the geometric mean is used to predict interactions between unlike molecules. Other
averages give quite different results. The HSP approach also uses the corresponding states approach wisely
chosen by Blanks and Prausnitz, comparing data for a given solvent with corresponding states data for its look-
alike hydrocarbon solvent (homomorph). Blanks and Prausnitz inherently also assumed the geometric mean for
the molecular dipole-dipole interactions. To this day there are those who protest inclusion of the hydrogen
bonding as is done in the Hansen methodology. These interactions are considered non-symmetrical with only
symmetrical interactions being describable by the solubility parameter theory. It seems that if dipolar molecular
interactions and the orientation involved are included, there should be no objection to include the hydrogen
bonding molecular orientation. The fact that the dispersion, dipolar, and hydrogen bonding energies sum to the
total cohesion energy for thousands of chemicals is difficult to dispute as well.
One might wonder when usage of the HSP concept first took off. I cannot answer this with any certainty. I have
concentrated on my direct responsibilities in industrial environments, trying to follow the relevant literature as
well as possible. I sense that industrial use has been extensive even very shortly after the 1967 work appeared.
These uses are rarely published. I was shown the number of citations of my publications as a function of year,
and it was clear that something happened around 2000, after the first edition of the handbook appeared. The
academics, who must certainly give the majority of reference citations, first really took interest the past 10
years or so. The key persons involved in the development and spreading of the concept almost all had direct or
close industrial ties including myself, Beerbower, Hoy, Van Krevelen, Abbott, and Yamamoto. The academics
would necessarily include Patterson and Delmas (who showed negative heats of mixing were found as expected
from solubility parameter theory) and Panayiotou and coworkers who put the hydrogen bonding cohesion
energy into a statistical thermodynamics context with success. The following is a typical academic reaction
from the late 1960’s to my early work. This is taken from a series of lecture notes/thesis from Denmark. I
prefer not to name the author here. Quote: The “theory” is applied to a very complicated systems, such as
solutions of macromolecules in polar and hydrogen-bonded solvents and solvent-mixtures. Even though the
method seems to have some technical value, the theoretical basis is extremely weak. It is only to hope that
serious work with the solubility parameter theory is never judged with such empirical methods in mind”. End of
Quote. This sums up the majority of the academics early views on “the three dimensional solubility parameter”,
and there are presumably still many who hold this view or something similar to it judging from the lack of
knowledge in the area that I find during my journal review activity. To my knowledge, with only a few notable
exceptions, there has been only very limited entry into classrooms at universities, although there have been
many Ph.D. thesis that have made use of the concept. The full social and economic potential of this
methodology will not be realized until universities include this in introductory courses. After all, the concept is
very simple and very useful.
The Side Track
For those who want to know a little more of what went on behind the scenes here are some more personal and
informal comments made in response to questions from Prof. Abbott.
The Hoy solubility parameters just sort of appeared some time after I was at PPG. One had to write to Union
Carbide to get a booklet with the tables. The tables were arranged according to alphabetical order, evaporation
rate, total solubility parameter, polar solubility parameter, hydrogen bonding solubility parameter, and boiling
point. The first booklet appeared in 1969. These values were also later revised for some solvents. Quoting from
a letter dated May 23, 1988, from Union Carbide accompanying a booklet dated 1985 - “Enclosed is a recent
copy of the “Hoy Tables of Solubility Parameters” you requested. It is basically the same as the 1975 edition,
but some updating of the data was done in 1981. Ken sends his greetings to you and looks forward to seeing
you in Athens. Signed R.L. Bradshaw.” The Hoy parameters appeared in Barton’s handbook from 1983 [8].
273
They apparently gained wide usage in the USA because there were data for many solvents not in my published
work and perhaps also because of the major influence and support of Union Carbide. Once established in a
given location, there has been a tendency for interest in them to continue. I have never fully understood how
these were calculated. The Hoy dispersion parameter was consistently lower than that found from the
corresponding states approach, and the expansion factor alpha appeared in both the polar and hydrogen bonding
terms, so I felt they were not independent. The dispersion parameter was found by subtracting the polar and
hydrogen bonding contributions from the total. I have always warned not to mix the Hoy parameters with the
original HSP. The Hoy parameters appeared as well in the first edition of Barton’s handbook (1983) with the
title “Hildebrand and Hansen Parameters for Liquids at 25°C, Determined by Hoy as Described in Sections 5.9
and 7.1”. The Hansen parameter terminology was therefore fully introduced at this time. I met Ken Hoy on
many occasions and fully respected his work, also in other areas. I have used the Hoy total parameter on many
occasions, and religiously went through the table in the Barton handbook from 1983 using the Hoy data for
Hildebrand parameters and molar volumes/density for many solvents in a transfer to my own HSP. Only a few
solvents (larger hydrocarbons) were not included in my list.
I gave 5 presentations at Gordon Research Conferences starting in 1967 at the Coatings conference. Here I met
Harry Burrell who gave a talk on hiding without pigment (using light scattering from microvoids), but he had
dropped further solubility parameter work by that time. There was also a talk by Crowley, Teague, and Lowe
from Tennessee Eastman describing their three dimensional approach to polymer solubility which had appeared
the year before. They used the Hildebrand parameter, the dipole moment, and an experimental (empirical)
hydrogen bonding parameter that I think was found from mixing solvents to precipitate polymers, much like
Kauri Gum is precipitated from n-butanol solution to find the KB values. These were not generally used and are
hardly mentioned in the Barton handbooks, but the thinking was in the right direction. I was admittedly a little
disturbed as to where they had gotten their idea, having sent a manuscript to the Journal of Paint Technology
earlier, presumably early in 1966. I withdrew the manuscript for some reason, perhaps for reasons of
knowledge gained in the meantime. I had a feeling the Eastman people had gotten access to this report, but was
assured by Crowley that they had not been aware of it. It was at this Gordon Conference that PPG became
aware of my work, thus leading to employment.
At an Adhesion Gordon Conference, I was confronted in the discussion after the presentation by a comment
from Fred Fowkes, an outstanding surface chemist. He said that I must have invoked Phlogistine theory
(everything is made from earth, fire, water, and air) to assign a hydrogen bonding parameter to toluene. I did
not know what this was at the time (a Google search on the word just confirmed the spelling and meaning), but
I responded that the experimental data clearly indicated that even toluene had some hydrogen bonding
character, although I could not precisely evaluate it. I could see it was less than 2, but greater than 0, so I took
1, not being too far off in any event. The units here are (cal/cc)½. At a Polymers conference my talk led to a
subsequent discussion lasting about 1½ hours. The group was split between the academics, who thought it to be
bunk, and the industrialists, who loved it. I got the traditional Amy Lifshitz award for promoting discussion,
which meant I had to drink a glass of a clear yellow liquid at the Thursday night meeting, having earlier
described the attributes of the most common form of saturated urea/water as used through history for various
purposes as a solvent and swelling agent. An academic exception was Prof. Tobolski who came to me the next
day with support, relating his own problems with the existing Establishment (Flory in particular, who delayed
publication of a paper to get his own in first). As an aside, I might mention that I have been told that there were
three different schools who did not think well of each other at all. There was one school in favor of Hildebrand
(Univ. of California), one school in favor of Prigogine (Univ. of Florida), and one school in favor of Flory
(Stanford). My own personal response to this is that I have never knowledgably had problems with any of them.
What they all lacked was quantification of the hydrogen bonding effects. Another academic, Prof. Donald
Patterson, whom I met on several occasions, was also very supportive and explained things along the way at
key points in time to help me along. A paper with his wife (Delmas) showing negative heats of mixing were not
only found, but they were found as predicted by solubility parameter theory, was a true milestone. This work
was very timely and decisive in changing many minds away from “empirical” to at least “semi-empirical”. A
major objection I often met was how can both negative and positive heats of mixing be accounted for by
solubility parameters? The Pattersons cleared this up as mentioned above. Another major question, also
discussed briefly above, was that Hildebrand assumed the geometric mean rule for calculating the interchange
energy between two different kinds of molecules, and that another rule was probably valid for hydrogen
bonding. My answer to this has been that polar interactions were accepted as following the geometric mean
274
rule. Since these are molecular and involve molecular orientation, I could not see why the molecular hydrogen
bonding interactions should be any different in this respect. In addition all of the many success stories using
HSP, where the geometric mean had been assumed following Hildebrand, have convinced me that this is the
correct way to do it..
My last experience with Gordon Conferences was also something special. Percy Pierce, my very close
colleague at PPG, had invited me from Denmark around 1980, and I was on after the lobster dinner on
Thursday evening. This particularly bad timing did not help, because I showed pictures of brain scans. The
Danish doctors, whom I believed, (but am no longer completely sure of what side effects there may have been
in their patients), claimed to have found and shown brain shrinkage because of solvent exposure. I found out
later that this caused (very) great concern in the coatings industry, but no one could talk about it because of the
rules of the Gordon conferences. Anyway, I have never been to a Gordon Conference since. One might question
why?
I have not attended international conferences outside of the Nordic countries since about 1986. The lack of
salesmanship of this kind probably delayed the acceptance in the academic community. This would have been
done on vacation and at my own expense and just seemed out of the question under the circumstances. The
Danish Establishment has not been particularly supportive in the past few decades. The major grants are
controlled by academics for the sole use of academics. I have only been significantly employed in industrial
environments. There have been some government incentives for cooperation between industry and academia in
an effort to force the cooperation between the more academic endeavors and industry. As an example I will cite
the 5 year grant for cooperation between my employer (FORCE Technology), the Risø National Laboratory
(now a part of the Technical University), and 9 Danish companies. (This was popularly called MONEPOL in
the Danish acronym) The consortium worked on Polymer Degradation. This resulted in about 25 publications
including two Ph.D. theses. The first year was led by my immediate supervisor, who then decided he could not
manage it. I was cautiously asked whether I would take over and did so with great pleasure for the next 3½
years. I could not finish the last half-year, having unwillingly left my job, because of not accepting a forced
reduction in working days.
There have been many papers on solubility parameters, cohesive energy density, cohesive parameters,
interaction parameters, and the like, and Barton did a great service in his thorough collection and reviews of
these. The last Barton handbook appeared in 1991. I have never had the resources and/or time to do this sort of
thing, but there are many significant reports that have appeared in the interim, the results of which should be
collected. I did manage a handbook that appeared in 1999, but time requirements restricted it mainly to what I
had been doing. When I discovered the “4” in the Prigogine theory in 1998, I decided immediately that this had
to be published. At the same time I had written so many journal articles, that I reasoned all I had to do was the
equivalent of writing a few more journal articles and put it into a book. Fortunately CRC thought this was a
good idea as well. Donald Patterson was very helpful at this time, as acknowledged in the handbook.
Having written the handbooks [13,14] has made me more cautious about handbooks. There was indeed no
review, and I could write whatever I wanted to. In the second edition of my Handbook I helped the others who
contributed where I could, but my own writings appeared without review, if that term is used appropriately.
Rest assured that I still stand by what was written.
It is sometimes asked why approaches such as UNIFAC and HSP have never been coordinated. I can only say
that my own attempts to initiate discussions to this end have never been reciprocated. My own ability to
influence matters was usually restricted by the fact that I worked in industry. My attempts to remain in or re-
enter academia and therefore have the time and resources to work on such issues were not well-received. All I
can say is that I have done the best I can with the resources available to me.
One small example of this is that a grant to me at the Technical University of Denmark was stopped after 10
months instead of the 2 years I was told would be the case (writing a book on diffusion). This ultimately led to
employment at FORCE Technology. The main import of my concepts on diffusion in polymers was thusly
delayed for over 20 years. A recent article in the European Polymer Journal (Hansen, C. M., "The significance
of the surface condition in solutions to the diffusion equation: explaining ‘anomalous’ sigmoidal, Case II, and
Super Case II absorption behavior", European Polymer Journal, Vol. 46, 651-662 (2010) summed up what I
would have written in the late 1980’s. Some few additional but significant pieces of information have appeared

275
in the interim, but the main message is the same (the surface condition must be considered to understand the so-
called anomalies).
Finale
In more recent times the second edition of the handbook was written in semi-retirement [14]. It was recognized
again that many others had now done significant work, both academic and industrial, and several of these
contributed to this edition of the handbook in their own areas of expertise (John Durkee, Georgios
Kontogeorgis, Costas Panayiotou, Tim Poulsen, Hanno Priebe, Per Redelius, and Laurie Williams). I was
grateful for these contributions, without which the second edition would not have appeared. Their work
advanced acceptance of the HSP methodology.
I met Prof. Abbott through a Danish company called CPS (Chemical Products and Services). CPS grew based
on the production of more environmentally acceptable cleaners already in the late 1980’s, primarily for the
serigraphic printing industry. I had gotten a special government grant for the fledgling company for the
development of the first series of these cleaners. There were patents with examples based on HSP. This
company was bought by Autotype in England, who were later bought by MacDermid in the USA. Prof. Abbott
led the technical activities at Autotype, and naturally appeared as a member of the board of CPS. On one
occasion I rapidly solved a problem using HSP where Prof. Abbott was having some difficulty. He had not been
looking in the third dimension (the D parameter). This then led to his development of suitable software and
ultimately to where we are with HSPiP in, as I write this, December, 2010.
The most recent and extensive contributions to the HSP theory and its practical applications appear in the
HSPiP (Hansen Solubility Parameters in Practice) eBook and software. This was started at the suggestion of
Prof Steven Abbott, with Dr YAMAMOTO Hiroshi soon joining in. These two have an unbelievable work
ethic. The volume and quality of what has appeared recently, and is still appearing on a regular basis, is
amazing. All of the significant methods for estimating HSP are included for those who may wish to continue
their use. The Stefanis-Panayiotou (S-P) method based on a statistical thermodynamic treatment as described in
the second edition of the Handbook, has already been more or less surpassed in volume and accuracy by the
YAMAMOTO Hiroshi’s molecular breaking method (Y-MB), supported by the extensive data, numerous
comparative correlations, and simple software application (just enter SMILES or MolFiles). I am very thankful
that what was started in the years 1964-1967 will survive and be used for a great many purposes for the benefit
of society and the environment.
References
1) Hansen, C.M., The Free Volume Interpretation of the Drying of Lacquer Films, Institute for the
Chemical Industry, The Technical University of Denmark, Copenhagen, 1964,
2) Burrell, H., A solvent formulating chart, Off. Dig. Fed. Soc. Paint Technol, 29(394), 1159-1173, 1957.
Burrell, H., The use of the solubility parameter concept in the United States, VI Federation
d’Associations de Techniciens des Industries des Peintures, Vernis, Emaux et Encres d’Imprimerie de
l’Europe Continentale, Congress Book, (The FATIPEC Congress book), 21-30, 1962.
3) Blanks, R.F. and Prausnitz, J.M., Thermodynamics of polymer solubility in polar and nonpolar systems,
Ind. Eng. Chem. Fundam., 3(1), 1-8, 1964.
4) Hansen, C.M., The three dimensional solubility parameter – key to paint component affinities I. –
Solvents, plasticizers, polymers and resins, J. Paint Technol., 39(505), 104-117, 1967.
5) Hansen, C.M., The three dimensional solubility parameter – key to paint component affinities II. Dyes,
emulsifiers, mutual solubility and compatibility, and pigments. J. Paint Technol., 39(511), 505-510,
1967.
6) Hansen, C.M., The three dimensional solubility parameter – key to paint component affinities III.
Independent calculation of the parameter components. J. Paint Technol., 39(511), 511-514, 1967.
7) Hansen, C.M., The Three Dimensional Solubility Parameter and Solvent Diffusion Coefficient, Doctoral
Dissertation, The Technical University of Denmark, Danish Technical Press, Copenhagen, 1967. PDF
file can be found on www.hansen-solubility.com.

276
8) Barton, A.F.M., Handbook of Solubility Parameters and Other Cohesion Parameters, CRC Press, Boca
Raton FL, 1983.
9) Barton, A.F.M., Handbook of Solubility Parameters and Other Cohesion Parameters, 2nd ed., CRC
Press, Boca Raton FL, 1991.
10) Hansen, C.M. and Beerbower, A., Solubility Parameters, in Kirk-Othmer Encyclopedia of Chemical
Technology, Suppl. Vol., 2nd ed., Standen, A., Ed., Interscience, New York, 1971, pp 889-910.
11) Beerbower, A., Boundary Lubrication – Scientific and Technical Applications Forecast, AD747336,
Office of the Chief of Research and Development, Department of the Army, Washington, D.C., 1972.
12) Beerbower, A., Surface free energy: a new relationship to bulk energies, J. Colloid Interface Sci., 35,
126-132, 1971.
13) Hansen, C.M. Hansen Solubility Parameters: A User’s Handbook, CRC Press, Boca Raton FL, 1999.
14) Hansen, C.M., Hansen Solubility Parameters: A User’s Handbook, 2nd ed., CRC Press, Boca Raton FL,
2007.

277
Chapter 39 The next steps (What Is Planned and Asked For)
The purpose of this package of eBook, software and datasets is to establish HSP as a routine tool for use in a
wide range of applications from paints to toxicology, from quantum dots to environmental plastics.
We’ve done what we can to make the principles clear and to provide working examples across a whole range of
techniques. We’ve tried to make the software as friendly, powerful and useful as possible – and used every bit
of the functionality for our own purposes in our own work and in writing the book. We’ve even added a set of
Power Tools.
We’ve sometimes had to go outside our own areas of speciality and had to push HSP to the limits of our
understanding.
Our hope was that this would be the beginning of a process. We wanted feedback from readers/users. We
expected criticism for some of the ideas, we expected that users would find bugs or limitations in the software.
We expected that some of the data in our databases would be challenged. We were right! The HSPiP user
community has been wonderful in consistently pushing us to improve the whole package. We remain keen to
receive more ideas to continue the progress.
Our commitment (backed up by the Abbott Guarantee) is to keep upgrading the book, software and data in the
light of our own continuing research and in response to user feedback. The many improvements in the 3rd
Edition have shown how serious is our commitment to this process. Getting to the 4th Edition then to the 5th
Edition has been a lot of work too.
We are acutely aware that we all need more robust ways for pinning down HSP for old and new chemicals.
We’ve done a lot of background work to refine the current tables, and (thanks to the generosity of their
originators) implemented the Y-MB and Stefanis-Panayiotou methods. We also continue to have some active
programs behind the scene which we believe will deliver further improvements. But we hope that readers/users
will also help in the process, by pointing out problems, by bringing their own techniques and ideas to help us
all.
The biggest problem faced in going from the 4th to 5th Editions was the combination of an interface that had
grown messy over the years as we responded to feedback and added more and more functionality, plus the fact
that newer screen resolutions and the sophistication of Windows 8 and 10 were making the interface harder to
sustain. So it was back to basics and a big update of the infrastructure. Finally we no longer offer support for
our Windows XP users (HSPiP now only runs on 64x machines), but we hope that the experience in Win 7, 8,
10 and beyond will be much more satisfactory.
So now you know that we are doing our bit to continue the progress on HSP. Our hope is that you will be happy
to contribute to the journey towards fuller understanding and usefulness.
Please contact us at steven@stevenabbott.co.uk or charles.hansen@get2net.dk or chemneuro@mac.com

Steven Abbott
Charles Hansen
YAMAMOTO Hiroshi
5th Edition November 2015

278
Hansen Solubility Parameters in Practice
5th Edition

About the authors


Professor Steven Abbott is an independent technical software author and consultant in the areas of
coating/printing/formulation and nano-science. He is a Visiting Professor at the School of Mechanical
Engineering, University of Leeds. He has a PhD in Chemistry from Oxford University (but did the work for it
at Harvard University) and has worked extensively in the coating and printing industries. His current research
interests include environmentally safer solvents for the printing industry, bio-mimetic nanosurfaces and
nanoparticle dispersions for high-performance coatings, surfactant theory, adhesion theory, practical skin
permeation science and writing as many technical apps as possible in these areas. He has written 3 books
(Nanocoatings, Adhesion, Surfactants), all linked to free on-line apps, and his 4th book, on Solubility Science is
slowly being written, along with the necessary apps.
Dr Charles M. Hansen is in a state of active semi-retirement working from his residence as consultant and
author, having recently completed a second edition of Hansen Solublilty Parameters: A User’s Handbook, CRC
Press, Boca Raton, 2007. He holds a B.Ch.E from the University of Louisville, an M.S. from the University of
Wisconsin, and lic. techn. and dr. techn. degrees from the Technical University of Denmark. He has worked
extensively with numerous organisations in the coatings, plastics, and related industries with employment by
PPG Industries in the USA, and the Scandinavian Paint and Printing Ink Institute and FORCE Technology, both
in Denmark.
Dr YAMAMOTO Hiroshi was a senior researcher at Asahi Glass Corporation. He has a PhD from Nihon
University “Molecular design of CFC alternatives using Chemo-Informatics” and has been a Visiting Associate
at CalTech. His expertise includes neural networks and data mining for thermodynamic and chemical
properties. Outside work he was “Senior Developer of HSPiP”, “ChemNeuro” and his site, www.pirika.com is
widely used and referenced in the literature for its range of on-line Java-based predictors. His amazing work for
HSPiP was all done in his spare time. Since April 2020 Hiroshi has retired from AGC and is now an
independent researcher, university lecturer, software systems creator (e.g. MAGICIAN - MAterials
Genome/Informatics and Chemo-Informatics Activate Networks) and tirelessly working to improve the power
of HSPiP. He is also HSPiP’s technical support expert for the Asia region.
ISBN 978-0-9551220-2-6

279

You might also like