You are on page 1of 292

TRANSFORMATION OF ALUMINOSILICATE

MATERIALS INTO HIGH VALUE ZEOLITES

Andrea Lucia Hidalgo Rocha


BEng (EnvEng)

Submitted in fulfilment of the requirements for the degree of

Doctor of Philosophy

School of Mechanical, Medical and Process Engineering


Faculty of Engineering

Queensland University of Technology

2023
Keywords

Acid treatment, Artificial Neural Network, Cancrinite, Cation-exchange

capacity, Hydrothermal, Lithium, Machine Learning, Mother liquor, Natural zeolite,

Quench Method, Spodumene, Thermal Activation, Waste recycling, Zeolite synthesis

i
Abstract

Zeolites are essential materials in various industries and applications due to their

unique chemical and physical properties. They find extensive use in areas such as

catalysis, adsorption, detergents, agriculture, construction, renewable fuels, industrial

chemistry, water treatment, and antiseptics. Traditionally they are made from alkaline

solutions of sodium silicate and sodium aluminate in the presence of an organic

template. However, in recent years, there has been growing interest in developing

sustainable methods for synthesizing zeolites using waste materials. While such efforts

hold significant promise, they are also beset by several challenges. In particular,

impurities present in waste materials can exacerbate the difficulties involved in

synthesizing zeolites, and the process of recycling mining waste in particular is

restricted by the energy-intensive nature of existing methods and their limited

scalability for widespread industrial implementation.

The first method discussed in this research is the synthesis of zeolite LTA from waste

spodumene leachate residue (SLR) from the lithium industry. Following the activation

of SLR, two distinct products were generated: a silica-rich mother liquor and

cancrinite. Subsequently, as a next step in the process, zeolite LTA was formed using

the silica-rich mother liquor resulting from the activation of SLR through the

application of alkaline solutions at a temperature of 240 oC. The mother liquor was

doped with sodium aluminate to ensure a Si/Al molar ratio = 1 was present. This

mixture was subjected to hydrothermal synthesis at 80 oC and ultimately produced

high-purity material. It is noted that the various impurities present in the waste

precipitated and as a consequence were mixed with the solid cancrinite.

ii
However, the production of cancrinite posed a challenge as it has little inherent value.

One approach was to dissolve the cancrinite with hydrochloric acid to make active

aluminium and silica species which could be subsequently neutralized with NaOH.

Zeolite LTA was then synthesized by hydrothermal reaction. However, the presence

of sodium chloride in this process was problematic, as it impacted the recycling of the

mother liquor and wash water, and ultimately resulted in the generation of wastewater.

Thus, alternative methods for cancrinite utilization were necessary.

The next method presented in this study, successfully addressed the issues encountered

in the first method. The process involved the conversion of natural zeolite waste

powder into two products, namely zeolite LTA and lanthanum-loaded cancrinite. The

high silica content in natural zeolites promoted the formation of zeolite LTA (i.e.,

mother liquor containing substantially more silicates than when using SLR) and

reduced the amount of cancrinite to be dealt with. Notably, the quality of the zeolite

LTA was not only of high crystallinity but also relatively pure. Doping of the

cancrinite powder with lanthanum species resulted in an acceptable method for

removing phosphate species from water and wastewater resources. Although the initial

performance of the cancrinite-based sorbent was encouraging, it is recommended that

future studies focus on improving the physical properties of the sorbent material.

Despite the success of the previous processes in producing high purity zeolites, there

was still scoped to reduce the activation temperature and to avoid making two

products. The "Quench Method" for zeolite LTA synthesis was discovered. The

iii
approach was to activate the waste in the first stage using high molarity NaOH

solutions (which normally end up making cancrinite and sodalite). Then at a key point

in time, addition of sodium aluminate solution occurred which quenched the activation

process and created conditions suitable for making zeolite LTA. Notably, the addition

of a slight excess of aluminium in the solution further improved the zeolite synthesis

process. The Quench Method has the potential to transform lithium industry waste as

activation occurs at only 100 oC.

To further optimize the proposed method for zeolite synthesis, the use of continuous

reactors is recommended. However, one of the primary challenges in employing such

reactors for this purpose is the issue of viscosity. As the reaction progresses, the

presence of solids in the mixture increases, resulting in increased viscosity, which can

cause problems such as reactor clogging, uneven heat distribution, and difficulties in

controlling the reaction conditions. To address this issue, a combination of a

continuous and a batch reactor has been considered, with the former used solely for

the activation phase to obtain active Si and Al species, while ensuring fast quenching

before reaching the next reactor.

By integrating Industry 4.0 principles, particularly machine learning, with continuous

reactors for zeolite synthesis, several benefits can be achieved. Incorporating machine

learning algorithms can significantly enhance the optimization of zeolite synthesis in

continuous reactors, resulting in reduced production time and costs, improved product

quality, and minimized waste. These algorithms can analyse real-time data collected

from sensors and other sources to dynamically adjust the operating parameters of the

reactor, such as temperature, pressure, and flow rate, thereby maximizing the

iv
efficiency of the synthesis process. Additionally, machine learning algorithms can

predict potential issues and provide early warnings, allowing for timely corrective

actions to be taken, minimizing downtime, and reducing the need for manual

intervention, which is often time-consuming and prone to human error. Therefore, this

study also addressed the need for a clean dataset to facilitate machine learning in

zeolite synthesis.

Developing suitable datasets is particularly challenging when comparing zeolite

synthesis parameters between different laboratories, which can negatively affect the

accuracy of machine learning predictions. To overcome this, researchers must

establish their own database, which involves conducting multiple experiments and is

time-consuming. Zeolite LTA serves as an excellent case study due to its well-

established synthesis process. This study identified critical factors, including limited

use of quantitative X-ray diffraction (XRD) methods in published literature, the lack

of reporting of the mass balance of processes, and the impact of vessel dimensions and

design on the synthesis process. Recommendations include pre-heating of reactant

solutions to minimize variation in reaction temperature and avoiding agitation of the

reactant mixture to improve the quality of zeolite LTA.

In the final part of this thesis, the dataset created was tested using a progressive

machine learning methodology to identify the relationship between zeolite synthesis

descriptors and evaluate the potential for machine learning to predict the quantitative

output of synthesis routes. The hypothesis was that applying statistics and machine

learning principles may enable pre-evaluation and increase zeolite yield and

performance. Chapter 7 shows the application of various machine learning algorithms

v
to zeolite LTA synthesis data, including linear regression, ridge regression, regression

tree, random forest, XGBoost, and artificial neural network models. Using artificial

neural networks resulted in 85 % accuracy in predicting the zeolite synthesis process.

vi
Table of Contents

Keywords ................................................................................................................ 1
Abstract .................................................................................................................. 2
Table of Contents ..................................................................................................... 7
List of Figures .......................................................................................................... 9
List of Tables ......................................................................................................... 13
List of Abbreviations............................................................................................... 15
Acknowledgements ................................................................................................. 17
List of Publications ................................................................................................. 18
Chapter 1: Introduction .......................................................................................1
Background ............................................................................................................. 1
Research Problem ..................................................................................................... 2
Research Aims and Objectives .................................................................................... 3
Thesis Outline .......................................................................................................... 4
Chapter 2: Literature Review...............................................................................7
Lithium in Australia .................................................................................................. 7
Lithium Mining Process............................................................................................. 7
Spodumene Leachate Residue (SLR) ........................................................................... 9
Natural Zeolites mining ........................................................................................... 10
Low silica zeolites from alumino-silicate waste ........................................................... 11
Zeolite Synthesis .................................................................................................... 12
Zeolite synthesis methods ........................................................................................ 15
Industry 4.0............................................................................................................ 20
Machine Learning ................................................................................................... 21
Key insights ........................................................................................................... 25
Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate
Residue via Cancrinite Formation......................................................................26
Abstract ................................................................................................................ 26
Introduction ........................................................................................................... 27
Materials and Methods ............................................................................................ 32
Results and Discussion ............................................................................................ 39
Conclusions ........................................................................................................... 66
Key Insights........................................................................................................... 68
Chapter 4: Transformation of natural zeolite powder to synthetic zeolite LTA
for water softening and lanthanum coated cancrinite for phosphate removal. ..70
Abstract:................................................................................................................ 70

vii
Introduction ........................................................................................................... 71
Materials and Methods............................................................................................. 75
Results and Discussion ............................................................................................ 81
Conclusions ........................................................................................................... 96
Key Insights ........................................................................................................... 97
Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate
Residue using a New Quench Method ............................................................... 98
Abstract................................................................................................................. 98
Introduction ........................................................................................................... 99
Materials and Methods........................................................................................... 104
Results and Discussion .......................................................................................... 109
Conclusions ......................................................................................................... 130
Key Insights ......................................................................................................... 131
Chapter 6: Factors which Influence Development of Robust Data Sets for
Machine Learning and Statistical Analysis of Zeolite Synthesis ..................... 133
Abstract............................................................................................................... 133
Introduction ......................................................................................................... 134
Materials and Methods........................................................................................... 138
Material Characterization ....................................................................................... 139
Solution Analysis .................................................................................................. 140
Results and Discussion .......................................................................................... 141
Conclusions ......................................................................................................... 170
Key Insights ......................................................................................................... 172
Chapter 7: Evaluation and application of machine learning principles to Zeolite
LTA synthesis .................................................................................................. 173
Abstract............................................................................................................... 175
Introduction ......................................................................................................... 176
Materials and Methods........................................................................................... 182
Results and Discussion .......................................................................................... 190
Conclusions ......................................................................................................... 211
Chapter 8: Conclusions and Future Work ...................................................... 213
Conclusions ......................................................................................................... 213
Future Work......................................................................................................... 216
Bibliography .................................................................................................... 218
Appendices ....................................................................................................... 235
Appendix A ......................................................................................................... 235
Appendix B.......................................................................................................... 245
Appendix C.......................................................................................................... 246

viii
List of Figures

Figure 1: Process flow diagram for making both cancrinite and zeolite LTA from
Spodumene Leachate Residue ............................................................. 33
Figure 2: Process flow diagram for the synthesis of zeolite LTA from Spodumene
Leachate Residue via cancrinite formation; Additional NaAlO 2
added to the feed stream to the high temperature hydrothermal
reactor. ............................................................................................ 35
Figure 3. Process flow diagram for the synthesis of zeolite LTA from Spodumene
Leachate Residue via cancrinite formation; Additional NaAlO 2
added to the cancrinite product. ........................................................... 36
Figure 4. Influence of time and NaOH molarity upon cancrinite synthesis from
SLR at 150 and 200 ℃ a) Crystalline cancrinite, b) non-Diffracting
material. Uncertainties were calculated with a 95% confidence
interval. ........................................................................................... 39
Figure 5: SEM images of solid material created by NaOH treatment of SLR at
temperatures of 150 and 240 o C ........................................................... 43
Figure 6: Process flow diagram including stream tables for (1) conversion of SLR
to cancrinite and (2) SLR mother liquor to zeolite LTA ........................... 47
Figure 7: SEM images of the zeolite product after hydrothermal reaction of mother
liquor modified by water and sodium aluminate addition; Reaction
temperature 80 o C; Time = 2 to 5 hours................................................. 50
Figure 8: Process flow for cancrinite synthesis from a starting mixture containing
SLR, NaOH, water and sodium aluminate; one hour at 240 o C in
hydrothermal reactor. ......................................................................... 52
Figure 9. SEM images of solid material made from cancrinite: four hours reaction
time; agitation 100 rpm; reaction temperature 80 ℃; dissolution of
cancrinite with either 2 or 3 M HCl followed by neutralization. ................ 54
Figure 10: Influence of HCl molarity on Al and Si active species formation from
cancrinite/sodium aluminate................................................................ 55
Figure 11.SEM images of zeolite LTA: Influence of different HCl concentrations
on zeolite LTA synthesis at 4 h, 80 ℃ and neutralized with 4 M
NaOH solution: Cancrinite/NaAlO2 dissolved in (a) 2 M HCl and (b)
3M HCl. .......................................................................................... 57
Figure 12: XRD determination of % of crystalline zeolite LTA in solid product as a
function of NaOH molarity used in the neutralization stage: reaction
temperature 80 o C; reaction time 4 h; Si/Al = 0.57 .................................. 57
Figure 13. Influence of NaOH molarity in the neutralization stage prior to the
formation of zeolite LTA. ................................................................... 59
Figure 14. Influence of different HCl concentrations on Zeolite LTA synthesis
using different Si/Al ratios, synthesis at 4 h, 80℃ and neutralized
with 4M NaOH solution. .................................................................... 61
Figure 15. Influence of synthesis time on zeolite LTA synthesis: reaction
temperature 80℃; 0.57 Si/Al ratio; and 100 rpm agitation........................ 62
Figure 16. Process flow diagram including stream tables for conversion of
cancrinite to zeolite LTA .................................................................... 65

ix
Figure 17: Concept for transformation of natural zeolite powder to synthetic zeolite
LTA and lanthanum coated cancrinite. ..................................................77
Figure 18. Active Si in the mother liquor (%) and solids recovery (g) at different
NaOH molarities ...............................................................................82
Figure 19: SEM images of solids formed after natural zeolite activation using 6 M
and 10 M NaOH: reaction temperature = 240 o C; reaction time = 1 h;
10,000x magnification. .......................................................................84
Figure 20. Mass balance for activation of natural zeolite powder using 6 M NaOH
at 240 o C for one hour followed by hydrothermal synthesis of zeolite
LTA at 80 o C for 4 h ..........................................................................88
Figure 21: Influence of synthesis time upon zeolite LTA synthesis: 80 o C, 150 rpm
and SiO2 /Al2 O3 =2.81..........................................................................90
Figure 22: Zeolite LTA SiO2 /Al2 O3 =1.90; Na2 O/SiO2 =5.11; H2 O/Na2O= 4.82, 4-
hour synthesis and 80 o C.....................................................................92
Figure 23: Equilibrium isotherms for phosphate sorption on lanthanum modified
cancrinite: Langmuir (top), Freundlich (middle) and Aranovich-
Donohue models (bottom)...................................................................94
Figure 24: Release of ions into solution as a function of from cancrinite mass
during phosphate equilibrium tests .......................................................95
Figure 25: Illustration of the “Quench” concept for making zeolite from waste
aluminosilicate materials................................................................... 103
Figure 26: SEM imaging and particle size of SLR .................................................... 110
Figure 27: Active Si and Al species in (a) mother liquor and (b) solid as a function
of NaOH molarity and contact time at 25 o C ........................................ 112
Figure 28: Active Si & Al species from SLR as a function of temperature (75℃),
NaOH molarity, and contact time. ...................................................... 113
Figure 29: Active species from SLR as a function of temperature (100 ℃), with
different NaOH molarity and contact time ........................................... 117
Figure 30: XRD analysis of active solid phase formed from reaction of SLR with
NaOH solution at 100 ℃; function of NaOH molarity and contact
time. .............................................................................................. 118
Figure 31. Active species from SLR as a function of temperature (150℃), NaOH
molarity and contact time.................................................................. 120
Figure 32. XRD analysis of active solid phase formed from reaction of SLR with
NaOH solution at 150 ℃; function of NaOH molarity and contact
time. .............................................................................................. 121
Figure 33: a) XRD and b) SEM image of the zeolite product produced by
hydrothermal reaction of non-activated SLR with with NaOH in a
sodium aluminate solution at 80 o C for 4 h: SiO2 /Al2 O3 ratio = 0.95 ........ 122
Figure 34. Influence of SLR activation temperature upon Zeolite LTA synthesis: 20
g SLR, 19.17 g NaOH, 5.33 g NaAlO2 , 80 ℃ synthesis temperature,
3.95 M NaOH ................................................................................. 124
Figure 35. Importance of hydrothermal synthesis reaction time at 80 o C for SLR
activated at 100 o C........................................................................... 125
Figure 36: SEM images of solid zeolite product formed as a function of time after
hydrothermal reaction using 3.95 M NaOH and Si/Al ratio = 0.95........... 126

x
Figure 37. SEM images of solid zeolite product formed after hydrothermal reaction
using 3.95 M NaOH and Si/Al ratio = 0.83 and 3-hour synthesis time ..... 128
Figure 38. Zeolite synthesis at 80 o C as a function of reaction time: SiO2 /Al2 O3
ratio = 0.83..................................................................................... 129
Figure 39: Quantitative XRD patterns for Zeolite LTA: SiO 2 /Al2 O3 = 2; Na2 O/SiO2
= 2.5; H2 O/SiO2 = 15 to 35; Temperature = 80 o C: Note that
trendlines are only a visual aid........................................................... 143
Figure 40: SEM images of zeolite made at 80 o C as a function of time: SiO2 /Al2 O3
= 2; Na2 O/SiO2 = 2.5 and H2 O/Na2 O = 15 or 35 as indicated.................. 147
Figure 41: Mass balance for the synthesis of zeolite L .............................................. 149
Figure 42. Heating profile of vessel in heated incubator at 80 o C. Aluminosilicate
gel conditions: SiO2 /Al2 O3 = 2; Na2 O/SiO2 = 2.5, H2 O/ Na2 O = 42.......... 152
Figure 43. Quantitative XRD patterns for Zeolite LTA: SiO 2 /Al2 O3 = 2; Na2 O/SiO2
= 2.5; H2 O/SiO2 = 15 to 35; Temperature = 80 ℃, pre-heated
samples to 80℃; Reaction Time = 3 h; Reactant mixture stirred at
100 rpm ......................................................................................... 155
Figure 44. Quantitative XRD data for Zeolite LTA: SiO 2 /Al2 O3 = 2; Na2 O/SiO2 =
2.5; H2 O/SiO2 =25 to 30; Reaction Temperature = 80 ℃; pre-heated
samples at 80 ℃; Reaction Time = 3 h; No stirring ............................... 159
Figure 45. Quantitative XRD patterns for Zeolite LTA: SiO 2 /Al2 O3 = 2; Na2 O/SiO2
= 2; H2 O/ SiO2 = 20 to 30; Temperature = 80 ℃ - pre-heated samples
at 80℃........................................................................................... 161
Figure 46. Formation of zeolite LTA crystal over time at 80 o C with preheated
samples; H2 O/Na = 30, Na2 O/SiO2 = 2, SiO2 /Al2 O3 = 2.......................... 165
Figure 47. Particle Sizing for zeolite LTA: SiO 2 /Al2 O3 = 2; Na2 O/SiO2 = 2;
H2 O/Na2 O = 20 to 30; temperature = 80 ℃ - pre-heated samples at
80 ℃............................................................................................. 166
Figure 48: Histogram showing citations of publications with the keywords
"machine learning AND zeolite", "neural network AND zeolite", and
"principal component analysis AND zeolite". ...................................... 177
Figure 49. Visualisation of the different neural network structures (ANN left, DNN
right). ............................................................................................ 179
Figure 50. Knowledge discovery process. ............................................................... 182
Figure 51. Progression and application of machine learning techniques according to
complexity. .................................................................................... 183
Figure 52. Zeolite synthesis products for gel composition of SiO 2 /Al2 O3 = 2,
Na2 O/SiO2 = 2.5 and H2 O/Na2 O = 30; reaction temperature 80 o C............ 190
Figure 53. Zeolite synthesis products as a function of reaction time for gel
composition of SiO2 /Al2 O3 = 2, Na2 O/SiO2 = 2 and H2 O/Na2 O = 20 to
30; reaction temperature 80 o C........................................................... 191
Figure 54. SEM images of zeolite synthesis products as a function of reaction time
for gel composition of SiO2 /Al2 O3 = 2, Na2 O/SiO2 = 2 and H2 O/Na2 O
= 30; reaction temperature 80 o C. ....................................................... 193
Figure 55. Correlation of machine learning variables using hierarchical clustering
methods. ........................................................................................ 195

xi
Figure 56. (a) Variance explained by principal components (b) principal
component coefficients. .................................................................... 196
Figure 57. Biplot comparison of principal components 1, 2, 3 and 4. ........................... 198
Figure 58. Linear and ridge regression feature importance from input synthesis
variables......................................................................................... 200
Figure 59. Regression tree and random forest feature importance for input synthesis
variables......................................................................................... 203
Figure 60. XGBoost testing results using input synthesis variables (results from the
training and testing runs that resulted in the greatest model
accuracy)........................................................................................ 204
Figure 61. ANN model results using input synthesis variables (results from the
training and testing runs that resulted in the greatest model
accuracy)........................................................................................ 207
Figure 62. (a) Comparison of machine learning model results (average R^2 value)
(b) Comparison of approaches to zeolite product analysis through
ANN models (average R^2 value) ...................................................... 209

xii
List of Tables

Table 1. XRD analysis of solid products resultant from SLR activation at 240 o C as
a function of NaOH molarity and reaction time ...................................... 43
Table 2. XRD results after the mother liquor recycling process. ................................... 48
Table 3. XRD data for zeolite LTA synthesis from cancrinite: four hours reaction
time; agitation 100 rpm; reaction temperature 80 ℃; dissolution of
cancrinite with either 2 or 3 M HCl followed by neutralization. ................ 53
Table 4. Quantitative XRD analysis of zeolite LTA: Impact of cancrinite/sodium
aluminate acid treatment conditions ..................................................... 56
Table 5. XRD analysis of zeolite LTA activated at 2M HCl and synthesised at a
reaction temperature 80℃; 0.57 Si/Al ratio; and 100 rpm agitation. ........... 62
Table 6: Mass balance for zeolite LTA formation by conversion of cancrinite ................ 63
Table 7: XRF analysis of natural zeolite powder ........................................................ 75
Table 8. Solid materials obtained after natural zeolite powder activation at 240 o C
and one hour synthesis. ...................................................................... 83
Table 9. Mother liquor composition as a function of storage time and NaOH
molarity: natural zeolite activation at 240 o C for 1 h................................ 86
Table 10: XRF analysis of solid product from activation of natural zeolite with 6 M
NaOH at 240 o C for one hour .............................................................. 89
Table 11. XRD determination of percentage of crystalline zeolite LTA in solid
product as a function of synthesis time: reaction temperature 80 ◦C;
SiO2 /Al2 O3 = 2.................................................................................. 90
Table 12. XRF analysis of zeolite LTA made from natural zeolite powder. .................... 91
Table 13. XRF analysis of SLR............................................................................. 109
Table 14. Quantitative XRD analysis of SLR .......................................................... 110
Table 15: Quantitative XRD data for SLR activated with NaOH at 75 o C..................... 114
Table 16. XRF analysis of solid product from Zeolite LTA synthesis using SLR
activated at 100 o C........................................................................... 127
Table 17: Summary of quantitative XRD analysis of zeolite samples made at 80 o C;
SiO2 /Al2 O3 = 2; Na2 O/SiO2 = 2.5 and H2 O/Na2 O = 15 to 35; stirring
100 rpm ......................................................................................... 144
Table 18. Summary of quantitative XRD patterns for Zeolite LTA: SiO 2 /Al2 O3 = 2;
Na2 O/SiO2 = 2.5; H2 O/ Na2 O = 15 to 35; Temperature = 80 ℃,
Agitation 100 rpm, pre-heated gel at 80℃ before synthesis .................. 156
Table 19. Summary of quantitative XRD patterns for Zeolite LTA with pre-heated
and non-stirring samples: SiO2 /Al2 O3 = 2; Na2 O/SiO2 = 2.5; H2 O/
Na2 O = 20 to 30; Temperature = 80 ℃................................................ 160
Table 20. Summary of quantitative XRD patterns for Zeolite LTA with pre-heated
samples and non-stirring: SiO2 /Al2 O3 = 2; Na2 O/SiO2 = 2; H2 O/ SiO2
= 20 to 30; Temperature = 80 ℃........................................................ 162

xiii
Table 21. Summary of particle sizing analysis for Zeolite LTA with pre-heated
samples and no-stirring: SiO2 /Al2 O3 = 2; Na2 O/SiO2 = 2; H2 O/ SiO2
= 20 to 30; Temperature = 80 ℃........................................................ 167
Table 22. Machine learning input variables. ............................................................ 188
Table 23. Training and testing results for linear and ridge regression. .......................... 199
Table 24. Training and testing results for tree-based models....................................... 201
Table 25. Training and testing results for ANN models (Ackley, Rege, & Saxena,
2003). ............................................................................................ 206

xiv
List of Abbreviations

Al – Aluminium

ANN- Artificial Neural Network

CEC – Calcium Exchange Capacity

Chorine- Cl

DNNs- Deep Neural Networks

HCAS- High Concentration Alkali Solution

HCl- hydrochloric acid

HEU- Heulandite

ICP-OES- Inductive Coupled Plasma-Optical Emission Spectrometer

La- Lanthanum

LOI- Loss on ignition

LTA- Linde type A zeolite

mEq- Milliequivalent

ML- Machine Learning

NaOH- Sodium Hydroxide

NZ- Natural Zeolite

OSDA- Organic Structure Directing Agents

PCA- Principal Component Analysis

QSP- Quasi-Solid Phase Activation Process

xv
SEM- Scanning Electron Microscopy

Si- Silica

SLR- Spodumene Leachate Residue

USY- Ultra-stable Y

XRD- X-ray Diffraction

XRF- X-ray Fluorescence

xvi
Acknowledgements

Firstly. I would like to express my gratitude and appreciation to my supervisory team:

Prof. Graeme Millar, Prof. Richi Nayak and Dr. John Outram for providing help and

support through my PhD journey. I am thankful for the scholarships provided by

Queensland University of Technology. This research was supported by QUT

Scholarship funding. I would like to acknowledge the Central Analytical Research

Facility (CARF) team for their help and the School of Mechanical, Medical and

Process Engineering as well as the Faculty of Engineering.

I would like to extend my gratitude to the excellent lab team I was part of for their

significant help during my candidature, Dr Josefine Probst, Dr Gabriel Colledge, Dr

Cameron Johnston and MSc Richard Freeney.

I would like to acknowledge professional editor, Sue Nielsen, who provided

copyediting and proofreading services, according to the guidelines laid out in the

university-endorsed national “Guidelines for editing research theses”.

I am very grateful for the enormous support of my family, who have always been my

strength and my friends who were with me and provided their invaluable moral support

and company during my PhD.

And last but not least, I would like to thank my fiancé and soon to be husband, Daniel,

who proofread this thesis even when he never understood my research. Thank you, my

love, for always keeping things going.

xvii
List of Publications

B. Conroy, R. Nayak, A.L. Hidalgo, and G.J. Millar. (2022) “Evaluation and

Application of Machine Learning Principles to Zeolite LTA Synthesis.” Microporous

and Mesoporous Materials 335 (Chapter 7).

DOI: https://doi.org/10.1016/j.micromeso.2022.111802

xviii
Chapter 1: Introduction

The mining industry generates significant amounts of waste material daily as part of

the mineral refining process. Disposing of this waste material in landfill sites is not

environmentally, socially, or politically acceptable. Therefore, the transformation and

optimization of waste materials into valuable products is crucial for promoting

sustainability in the mining industry. One promising solution is production of green

zeolites, which can add value to aluminosilicate waste materials.

The aim of this research is to develop an innovative approach to repurpose

aluminosilicate waste materials, such as SLR and natural zeolite powder, into synthetic

zeolites, by optimizing the activation of SLR and synthesizing zeolite LTA. The

incorporation of machine learning algorithms can significantly enhance the

optimization of zeolite synthesis in continuous reactors, leading to reduced production

time and costs, improved product quality, and minimized waste. However, creating a

robust dataset for machine learning in zeolite synthesis presents challenges, and

researchers must establish their own database for zeolite synthesis, despite the time -

consuming nature of this process.

Background

The mining industry is critical for Australia’s economy; in December 2021 the

Australian Department of Industry, Science, Energy and Resources highlighted the

vital role of the mining industry for the recovery of the Australian economy after the

COVID-19 pandemic. This estimated importance will be supported by Australia’s

Chapter 1: Introduction 1
critical mining materials such as iron ore, coal, gold, copper and the rising exports of

lithium and nickel (battery minerals).

Lithium demand has rapidly grown in the last two decades as it has become vital for

the development of industrial products, particularly batteries for electric vehicles,

hybrid vehicles and portable electronic devices (Meng, et al., 2021). Currently,

Australia is the largest producer of lithium in the world with mines located mainly in

Western Australia (Maxwell & Mora, 2019). This rapidly increasing demand for

lithium is producing tonnes of industrial waste material, which has been stored in

landfills.

Zeolite mining in Australia is still relatively small in scale but is expected to grow

(Victoria State Government, 2023). This projected growth is poised to bring about a

significant surge in waste generation.

Research Problem

The Australian mining industry contributes to 75% of the country's exports, which

impacts significantly on Australia's flourishing economy. Nevertheless, an increase in

mining activity also increases in the amount of waste material produced daily. There

is a large benefit in investigating different methods to use these waste materials which

are, as of writing, mostly disposed in landfills. It is estimated that one tonne of lithium

carbonate generates approximately 9-10 tonnes of lithium slag (D.Chen et al., 2012;

Z-h. He, et al., 2017).

Furthermore, natural zeolite deposits generate a waste material (ultrafine powder) after

the crushing step. This by-product represents a loss in production, a liability and a

Chapter 1: Introduction 2
potential hazard (Zijun et al. 2021). These waste materials are mainly composed of Al

and Si, which can be used as raw materials to generate higher value products such as

zeolites.

Currently, there has been limited research on the conversion of mining wastes into

valuable zeolites. This is largely due to the challenges posed by impurities in the

material, which can adversely affect the quality of the final product, as well as by the

high cost of activation. Accordingly, this thesis investigates various methods for

activating and synthesizing zeolites from waste materials, while also assessing the

potential of machine learning approaches for optimizing zeolite synthesis.

Research Aims and Objectives

This project has a twofold aim: firstly, to explore novel and environmentally

sustainable synthesis methods for producing high-quality zeolites from a variety of

waste materials; and secondly, to develop algorithms that leverage machine learning

approaches to enhance our understanding of zeolite synthesis. The project seeks to

contribute to the growing field of sustainable materials science and to advance the

application of artificial intelligence in the zeolite community. The following research

objectives were identified and addressed:

1 To develop and optimise a process to use the waste material and transform it into

synthetic minerals.

2 To create a novel process to transform lithium mining waste into zeolite LTA that

can be scaled up to industry, avoiding a fusion step.

Chapter 1: Introduction 3
3 To explore the potential of using natural zeolites to transform them into synthetic

zeolite LTA using the mother liquor rich in silica after zeolite activation.

4 To study the effect of different variables in zeolite synthesis to develop a robust

data set for machine learning applications.

Thesis Outline

This thesis is presented by monograph, with a total of eight chapters. Chapter 1

introduces the research, emphasising the background, research problems, aim, and

objectives. In Chapter 2, a comprehensive literature review is presented to identify

gaps in the research, such as the limited studies that have considered using Spodumene

Leachate Residue (SLR) and Natural zeolite powder as raw materials for zeolite

synthesis, as well as the challenges of applying machine learning to zeolite science.

Chapter 3 explores the feasibility of synthesising zeolite LTA from SLR via cancrinite

formation, in which the waste material is converted into cancrinite and then

reprocessed under relatively mild conditions. The cancrinite method minimises waste

of resources by recycling the mother liquor after waste material activation and zeolite

synthesis. In Chapter 4, the cancrinite method is further studied by investigating the

activation of the mother liquor using ultra-fine natural zeolite powder instead of SLR,

to maximise the use of different mining waste materials. This chapter also

demonstrates the efficacy of this methodology with different aluminosilicate materials.

Chapter 5 investigates a novel process to transform lithium mining waste into synthetic

zeolite LTA using a new quench method, representing an alternative synthesis method

Chapter 1: Introduction 4
that does not require a fusion step and presents green credentials such as lower

activation temperatures, solvent that can be reused in the process, and minimum water

loss. Chapters 3 to 5 detail green and sustainable alternatives to traditional synthesis

methods.

To enhance the benefits of the methodologies presented in previous chapters, machine

learning is employed to manage the large amount of data. Chapter 6 develops a robust

dataset for zeolite LTA to evaluate machine learning principles for zeolite synthesis.

In Chapter 7, different machine learning algorithms are applied to identify the

relationship between zeolite synthesis variables. Notably, the most accurate algorithm

is the artificial neural network model, with an accuracy of over 80%. However, the

accuracy of the model can be further improved by increasing the network size,

demonstrating that advanced deep learning models should be considered.

Chapter 8 serves as the concluding chapter of this thesis, summarizing the key

contributions and findings. Additionally, it outlines potential directions for future

research. The concluding section offers valuable insights into the main outcomes

derived from the study and provides suggestions for further investigation.

Furthermore, the last chapter features appendices that provide supplementary

information for a comprehensive understanding of the research. Appendix A offers

extra analysis and results, expanding upon the findings presented in Chapter 3. It

delves deeper into the data collected, providing readers with a more in-depth analysis

and supporting evidence. Appendix B showcases a continuous reactor prototype,

which serves as a tangible application of the research conducted in Chapter 5. Lastly,

Appendix C includes supplementary information related to Chapter 7, providing

Chapter 1: Introduction 5
further elucidation and analysis of the discussed topics. It offers additional insights,

supporting the arguments and conclusions presented in the main body of the thesis.

Chapter 1: Introduction 6
Chapter 2: Literature Review

Lithium in Australia

Australia is a significant global producer of lithium, which is experiencing growing

demand due to the increasing need for mobile electronics and electric vehicles.

(Courtney et al., 2017). Lithium and its compounds have historically been used in

application such as batteries, lubricating greases, pharmaceuticals (Choubey et al,

2016; Collins, et al. 2020; G. Han, et al, 2018; Z. Liu, Zhu, et al., 2019; Sun, et al.

2014), glass, ceramics, metallurgy (P. Meshram et al. 2014) and other industries. High

capacity lithium batteries are strong candidates for a feasible solution for energy

storage for the electrical power grid and they are already needed for the accumulation

of green energy (V. Flexer et al., 2018); and commonly used power sources for electric

vehicles (Setoudeh et al. 2020). This means there will most likely be a strong increase

in the demand for lithium in the coming years (Setoudeh et al., 2020).

Lithium Mining Process

Lithium is extracted mainly from hard rock ores (lithium minerals such as spodumene,

petalite and lepidolite) (Han et al., 2018) and aqueous resources such as lithium rich

brines or salt lakes ( Flexer et al., 2018). At present, the majority of lithium production

around the world belongs to: six mineral operations in Australia, two brine operations

each in Argentina and Chile, and one brine and one mineral operation in China (USGS,

2020), placing Australia as the world’s largest producer of lithium (Flexer et al., 2018;

Vikström et al. . 2013). Spodumene is the favoured resource, as lepidolite comprises

of high fluorine content, while petalite, which is generally used for glass, also has high

fluorine content (Vikström et al., 2013).

Chapter 2: Literature Review 7


The continuously increasing demand for lithium and its high economic importance

(Choubey et al., 2016) will require new paths for sustainable exploitation of this

element. Extraction of lithium from spodumene generates waste material called

lithium slag or spodumene leachate residue. There are various processes to extract

lithium from minerals such as the following.

Alkaline Process

For the alkaline process, spodumene or lepidolite ore concentrates are ground and

calcined with limestone between 825-1050 °C. Then, the calcined product is crushed,

milled, and treated with water to yield lithium hydroxide, which can be converted to

lithium chloride by reaction with hydrochloric acid. The lithium recovery by this

method is approximately 85–90 % (P. Meshram et al., 2014).

Acid Process

Yan et al. (2012) reported a novel technique to extract lithium from lepidolite, which

consisted of sulfation roasting of lepidolite followed by water leaching; reaching a

lithium extraction efficiency of 98.9 % and obtaining 99.9% purity of LiCO3. The

optimum parameters were: de-fluorination temperature of 860 °C; de-fluorination time

of 30 min; milling time of 100 min; leaching temperature of 150 °C; leaching time of

60 minutes and liquid-to-solid ratio of 4:1 (Yan et al., 2012).

Due to its exceptional effectiveness, the sulfuric acid method has emerged as the

primary technique for extracting lithium carbonate from spodumene. First, α-

spodumene is transformed into the more reactive β-spodumene by roasting at 1070–

Chapter 2: Literature Review 8


1100 °C (Y. Chen et al., 2011). Then, the calcined material is mixed with sulfuric acid

and roasted at 250 °C followed by a water leaching process to yield a solution of

lithium sulfate. Lithium carbonate can be recovered after the addition of sodium

carbonate to the solution, pH adjustment, purification and evaporation (P. Meshram et

al., 2014). The results show that the purity obtained of lithium carbonate can reach up

to 99.6% (Y. Chen et al., 2011). This method is still the most popular in the industry

due to its high extraction efficiency, relatively low costs, maturity of technology and

low economic cost to start-up (Izidoro et al., 2019).

It is important to mention that in the sulfuric process for each tonne of processed

mineral, 0.95 tons of acid residue is produced (Rosales et al., 2016). Indeed, one tonne

of lithium carbonate generates approximately 9-10 tonnes of lithium slag (D. Chen et

al., 2012; Z-h. He et al., 2017). This situation shows that Li is the only product

recovered from β-spodumene and Si, and Al remains as a waste (Rosales et al., 2016).

Yiren et al., (2019) predicted that the discharge of lithium slag would reach 1,200,000

tonnes/annum. For this reason, it is crucial to develop an efficient way to recycle

lithium slag in order to save the natural resources and protect the environment (Shan

et al., 2018).

Spodumene Leachate Residue (SLR)

The aluminosilicate based waste product from hard rock refining is generically termed

“lithium slag”, and more specifically spodumene leachate residue (SLR), if from a

spodumene source, which has the general chemical formula of AlxSixOx (Zampori et

al., 2012). SLR is the solid waste of spodumene after high temperature roasting and

leaching. Han et al. (2018) reported the composition of the SLR used in their research,

Chapter 2: Literature Review 9


showing that it was mainly composed of SiO 2 and Al2O3 with a total content of 97.91%

for both of them. A large amount of SLR is stored outside in landfills, which poses a

severe threat to the environment ( Chen et al., 2012). SLR is mainly used as cement

clinker, in concrete, as a raw material for ceramic glazed tiles and in activated clays

(G. Han et al., 2018). However, the utilisation rate is about 10% (Beushausen et al.,

2012). It is thus important to start using lithium slag efficiently to reduce/avoid

pollution and waste of resources. Currently, research using SLR as a resource has

focused on recovering lithium (G. Han et al., 2018), even when the main composition

of SLR is aluminosilicates.

Natural Zeolites mining

Natural zeolites are porous minerals, found in geological deposits such as sedimentary

and volcanic rocks (Guarino Bertholini, 2016). There are multiple applications for

these materials such as ion exchange materials for water and wastewater treatment

(Millar, Couperthwaite et al., 2016), ammonium exchange (Millar, Winnett, et al.,

2016), removal of heavy metals from waste water (Rodríguez-Iznaga, et al. , 2018),

removal of bacteria from filters (Ivankovic et al., 2019), and as a resource for

transforming them into higher value products (Wruck et al., 2021).

Mining a natural zeolite deposit usually involves crushing, grinding and fractionation

(screening into size classes or hydro-classification) (Yusupov et al. , 2000). However,

after the crushing step, a non-marketable natural zeolite ultrafine powder is created

(<50 µm). This material currently represents a loss in the process and an economic

liability due the costs of storage and disposal as well as a potential health hazard due

the small particles in the air (Zijun et al., 2021).

Chapter 2: Literature Review 10


Ultrafine powder from natural zeolite is considered a waste material and it is important

to find efficient methods to use it. Natural zeolite is high in Si and Al, materials that

can be used for industry purposes.

Low silica zeolites from alumino-silicate waste

The literature on the use of alumino-silicate waste has a primary focus on the

production of low silica zeolites, such as Zeolite LTA or faujasite (Poole, Prijatama,

& Rice, 2000). For example, Mahima Kumar et al. (2020), synthesized LTA Zeolite

from raw fly ash (without calcination or acid treatment) by using fusion with alkali

and hydrothermal synthesis. Moreover, Izidoro et al. (2019), used the waste of iron

mine tailing dam as raw material to produce Zeolite LTA; this synthesis was carried

out with a fusion step followed by a conventional hydrothermal reaction.

Zeolite LTA
Zeolite LTA is characterised by the formula Na +12(H2O)27)8 (Al12Si12O48)8 which

corresponds to its most common hydrated sodium form (Drioli & Giorno, 2016). It has

a cubic crystal structure with a lattice parameter of 12.32 Å and 3-dimensional network

involving spherical cavities of 4.2 Å in diameter. Zeolite LTA is well known in the

industry due to use in laundry detergents (S. U. Meshram et al. 2014). It can be used

as a substitute of sodium tripolyphosphate (STP), which has been banned in many

countries (Ayele, Pérez-Pariente, Chebude, & Díaz, 2016a). STP has been identified

as the main cause of eutrophication and this problem can be solved by replacing the

phosphate based compounds for a greener option such as zeolite LTA (Rayalu et al.,

2001). Zeolite LTA has also been used in different industrial processes such as

dehydration of bio-ethanol (Sato et al., 2008), heavy metals removal (Jamil et al. 2010)

Chapter 2: Literature Review 11


and desalination of radioactive solutions (Malekpour et al. 2008). For this reason,

zeolite LTA is by far the most commonly produced synthetic zeolite, making up 73%

of global production (Ayele, Pérez-Pariente, Chebude, & Díaz, 2016a).

Zeolite LTA can also be used in the healthcare industry. When it is exchanged with

silver ions it can be used as an antimicrobial material (Collins et al., 2020; Y. Zhang

et al., 2009). It has also been proven that the use of a silver zeolite-impregnated

umbilical catheter is effective in decreasing the risk of developing a catheter-related

bloodstream infection (CRBSI) in preterm infants (Bertini et al., 2013) .

Synthesis of zeolite LTA has been successfully performed using different low cost

natural/waste materials such as bentonite (H. Ma et al., 2010), kaolinite (Johnson &

Arshad, 2014; Rios et al.,2009), fly ash (Rayalu et al., 2001), sugarcane bagasse ash

(Moisés et al., 2013), clay (García et al., 2015) and lithium slag (B. Wang et al., 2022;

Huang et al., 2023).

Zeolite Synthesis

Waste aluminosilicate and naturally occurring aluminosilicates, such as clay, can be

transformed into higher value products such as zeolites. Zeolites are typically

microporous or mesoporous with the general composition: Mx/n (AlO2) x (SiO2) y ∙ zH2O

where M is the cation that compensates for the negatively charged framework, n is the

cation valency, y/x is the Si/Al ratio, and z is the water content. Zeolites are employed

commercially for application such as catalysis, gas separation, and sorbents (Müller et

al., 2015; Whiting et al., 2019), adsorption of greenhouse gasses (Müller et al., 2015;

Chapter 2: Literature Review 12


Xu et al., 2019), ion exchange (Müller et al., 2015; Zhang et al., 2015) and as molecular

sieves for separation of compounds (Fakin, Ristić, Mavrodinova, & Zabukovec Logar,

2015; Klumpp, Zeng, Al-Thabaiti, Weber, & Schwieger, 2016). In recent years, a need

has developed to create green synthesis methods for zeolites ( Liu et al., 2014; Pan,

Wu, & Yip, 2019). Hence, efforts have been made to reduce the use of organic

structure directing agents (OSDA’s) which are potentially toxic. The synthesis of

zeolites under organotemplate-free conditions has been proven successful, avoiding

the combustion of OSDAs and the generation of liquid wastes and harmful gases;

which means this practice can be consider as environmental friendly synthesis (Y.

Wang, Wu, Meng, & Xiao, 2017). Y. Ji, Wang, Xie, & Xiao (2015) demonstrated that

the organotemplate-free synthesis was possible using a seed-directed synthesis where

it was suggested as a third zeolite structure-directing agent.

Avoid solvent consumption.


It has been demonstrated that zeolite synthesis under solvent-free conditions can be

successfully performed by mixing, grinding, and heating starting solid materials. Wu

et al. (Wu et al., 2018), indicated exceptional advantages of this method, including the

development of hierarchical micro-, meso- and macrostructures, fast synthesis at

higher temperatures, and construction of a novel catalytic system for encapsulation of

metal nanoparticles and metal oxide particles within zeolite crystals synergistically.

The combination of organotemplate-free synthesis with a solvent-free methodology

has the potential to develop a more environmentally friendly synthesis, however

minimal research has been conducted in this field.

Chapter 2: Literature Review 13


Utilise waste alumina-silicate resources.
Industrial wastes with Si and Al as main components can be used as raw material for

zeolite synthesis. Kuroki, et al., (2019), successfully synthesised zeolite X and zeolite

LTA from crushed stone and aluminium ash by two hydrothermal treatments at low

temperature (>150℃). Selective synthesis of zeolite X and A was accomplished by

controlling the acid treatment conditions. As mentioned previously, the current

approaches with waste resources have been increasing and more waste material is

being used for the synthesis of zeolites, in order to reduce contamination, costs and

save natural resources. However, there is a lack of studies on what can be used as

starting material for zeolite synthesis. SLR is a waste alumina-silicate material and

therefore it does not easily form into zeolites (T. Li et al., 2012). Therefore, an

activation strategy may be required; with one of the simplest approaches being heating

at elevated temperatures. Heat treatment in order to improve the properties of SLR has

been studied in relation to construction industry utilisation (Z. Liu, Wang et al., 2019).

Minimise energy consumption.

Majdinasab et al. (2019) proved that the conversion of waste glass into zeolites can

be achieved using a hydrothermal reaction and microwave heating for an effective and

rapid conversion of the waste glass to zeolites with a relative crystallinity of 60%. Due

to the high-energy efficiency, superheated conditions can considerably reduce the

crystallisation time from numerous hours to minutes.

All these applications show the need to develop new techniques that allow industry to

use their waste materials in a more environmentally friendly way. Zeolite market

demand is growing and so is the demand for synthetic zeolites. This project will focus

Chapter 2: Literature Review 14


on the use of SLR for zeolite synthesis using the “one-pot” technique, which will allow

not only the use waste materials, but also minimize the consumption of resources.

Zeolite synthesis methods

High Concentration Alkali Solution (HCAS) Activation


T. Li et al. (2012), explained how to activate natural alumina-silicate minerals without

the involvement of inorganic chemicals (aluminium and silicon) in order to transform

them into zeolite Y. To perform this method, the selection of at least two suitable

natural mineral sources was required; in this case kaolinite (which contain high levels

of alumina and silica), and diatomite (rich in silica) were chosen. The activation of

the minerals, where the inert alumina-silicate (kaolinite) was treated with a high

concentration alkali solution (HCAS) activation method, while the mineral in the non-

crystalline state (diatomite) was activated by the normal thermal activation method.

This research proved that the purity of the synthetic zeolite Y was comparable to a

commercial sample and highlighted that all the impurities originally existing in the

material remained in the mother liquid. It also revealed that even at 200℃ (normal

thermal activation performs at 800 ℃) HCAS activation can successfully destroy the

structure of kaolin and other alumina-silicate minerals and generate higher contents of

active alumina and silica which created higher reactivity for zeolite synthesis. The

material balance assessment analysis illustrated that 97.4% and 64.2% of kaolinite and

diatomite respectively were transformed into zeolite. Finally, it was proven that this

method compared to the conventional synthesis has numerous advantages such as

reduction of solid and water discharges, CO2 emissions, and energy consumption (T.Li

et al., 2012).

Chapter 2: Literature Review 15


Quasi-Solid Phase Activation Process (QSP)
Yang et al. (2017) described how to activate kaolin under relatively mild conditions

in order to make it amenable for conversion into zeolite LTA and zeolite Y. Initially,

kaolin was kneaded with some sodium hydroxide and then water was added.

Subsequently, this material was extruded into shapes 1.5 mm in diameter and 2 to 3

cm in length. The extrudate was finally heated to between 60 and 100 ℃ for 30

minutes in a belt type calcination oven. This new methodology was termed Quasi-

Solid Phase Activation Process (QSP). The principal advantages relative to the HCAS

approach were:

 Lower activation temperature.

 Less water required.

 Continuous processing (as opposed to batch for HCAS)

 Reduced energy consumption.

The superior effectiveness of the QSP approach to the HCAS methodology was

illustrated for hydrothermal synthesis of both zeolite LTA and zeolite Y.

Alkali fusion synthesis


Ayele et al. (2016b) applied alkali fusion synthesis for the activation of raw kaolin by

combining 1.25g of raw kaolin with 1.5g of NaOH for 30 minutes, afterwards the mix

was calcined at 600℃ for 1 hour in a furnace. Once the fused product was ready, it

was ground in a mortar and mixed with 12.5mL of ultrapure distilled water at 50℃ for

1 hour. For crystallisation, the samples were heated using water baths under static

conditions at 100℃ for three hours. They achieved an optimum crystallinity of 84%.

Chapter 2: Literature Review 16


This research also compared alkali fusion synthesis against the conventional

hydrothermal method and proved that alkali fusion was the best method in terms of

time, energy cost, and product quality and it also allowed the use of low-grade virgin

kaolin without purification, whereas the conventional method needs a purified raw

kaolin to achieve a good quality zeolite.

Zijun et al. (2021) used ultrafine natural zeolite powder to make zeolite W by thermally

activating the material, followed by a fusion step to combine the natural zeolite with

potassium hydroxide. After that, the mixture was combined with aluminium

hydroxide, potassium hydroxide and water and synthetized at 150 oC and 16 hours.

The material made showed a cation exchange capacity of 324meq/100 g indicating

potential for ammonium uptake.

Wruck et al. (2021) investigated the impact of a pre-activation of natural zeolite upon

zeolite LTA synthesis, using natural zeolite samples from three different deposits.

However, the study showed that all the samples lost their cation exchange capacity

after being heated at elevated temperatures, so an alternative activation was used

(alkali fusion). Fusion activation resulted in a significant increase in zeolite LTA, and

it was reported that the product resulting from hydrothermal synthesis was formed of

85% crystalline zeolite LTA.

One-pot synthesis
‘One pot’ synthesis method for waste-to-zeolite processes is not only simple but a

greener option from the scientific point of view since it effectively eliminates the

Chapter 2: Literature Review 17


fusion step (X. Ji et al., 2018). The dissolution of silica in sodium hydroxide is widely

reported, but the detailed mechanism is poorly understood. Dissolution of amorphous

silica can be expressed as:

𝒙(𝑺𝒊𝑶𝟐) + 𝟐𝑯𝟐 𝑶 ↔ (𝒙 − 𝟏)(𝑺𝒊𝑶𝟐) + 𝑺𝒊(𝑶𝑯)𝟒

Equation 1. Dissolution of amorphous silica

There is some literature pertaining to the use of sodium hydroxide leaching for Li

extraction However, little focus has been given to the structure changes of the solid

residues and, indeed, the kinetics of spodumene dissolution, even though this method

is more convenient, simple and environmentally benign than the conventional one (X.

Ji et al., 2018), further investigation needs to be done to explain the mechanisms of the

crystal phase and understand the optimum conditions for zeolite synthesis.

Green methods
The use of green techniques for zeolite synthesis is essential for promoting

sustainability and reducing environmental impact. However, there are still some

challenges that need to be addressed. For instance, the optimization of resource

utilization is crucial to minimize the environmental impact of zeolite synthesis. This

can be achieved by developing methods that reduce the consumption of energy, water,

and raw materials.

The conventional hydrothermal synthesis of zeolites is an exigent challenge from the

green chemistry standpoint because it depends on the use of organic templates which

Chapter 2: Literature Review 18


are removed at the end of the process by calcination, or extracted using organic

solvents which produce several risks to the environment including the emission of

green gases during the combustion of toxic organic templates (Pan et al., 2019).

According to Pan et al. (2019), the green synthesis of zeolites can be classified in four

main categories:

 Synthetic methods which avoid the use of a template or use a recyclable or

renewable template,

 Synthetic methods that use sustainable resources,

 Solvent-free methods

 Facile synthesis methods.

As demonstrated through the aforementioned methods, researchers have attempted to

develop novel techniques to promote environmentally friendly zeolite synthesis. The

integration of various green methods requires further investigation to optimize

resource utilization and minimize environmental impact. One of the green techniques

that not yet mentioned is the application of ultrasound pre-treatment for zeolite

synthesis (Reinoso et al., 2018). Reinoso et al. (2018) synthetized faujasite without

using any OSDAs, at low temperature (30 °C for the aging pre-treatment and 100 °C

for the hydrothermal treatment) and applying ultrasound-assisted aging. This was

accomplished in two stages, aging step, and hydrothermal treatment, obtaining a pure,

well-crystalised faujasite with small crystal size (38 nm).

Despite the progress made in the development of green methods for zeolite synthesis,

there are still some remaining gaps that need to be addressed. For instance, there is a

need to optimize the efficiency of these methods to make them more cost-effective and

Chapter 2: Literature Review 19


practical for industrial-scale production. Additionally, there is a need to develop more

comprehensive and standardized methods for evaluating the environmental impact of

zeolite synthesis. This will help to ensure that the methods developed are truly

sustainable and environmentally friendly.

Industry 4.0

The concept of Industry 4.0 involves the technological innovation of automation,

control and information technology applied to manufacturing process (Machado &

Paulo, 2020). It is focused on smart procedures, processes, and products (smart

production). This means there is direct communication between equipment, labour,

and resources (Heng, 2020).

The Industry 4.0 principles are considered a new industrial stage where many

emerging technologies are converging to enable digital solutions (Frank et al 2019). It

depends on the implementation of digital technologies to collect and analyse real-time

data to provide useful information to the manufacturing system (Frank et al., 2019).

According to Frank et al. (2019), Industry 4.0 technologies can be distributed into two

layers: “front-end technologies” which are focused on operational and market needs,

and the “base technologies” that supply connectivity and intelligence for the first layer

to then be a complete integrated manufacturing system.

The base technologies support all the “Smart”. They are constituted by the information

and communication technologies (ICT). The combination of IoT (Internet of Things)

Chapter 2: Literature Review 20


and Cloud allows various equipment to be connected, gathering a massive amount of

data (big data storage) (Lu, 2017). Big data combined with analytics, such as machine

learning, is considered a strong alternative or contender for shaping the future and

playing a crucial role in driving the fourth industrial revolution (Frank et al., 2019) .

In 2017 a cooperation agreement with Germany’s Platform Industry 4.0 was signed to

provide a boost to Australia’s economic competitiveness (Australian-Government ,

2020). This agreement explores the framework and principles for adopting Industry

4.0 in the country and prepares for the transition to smart factories. One of the priority

growth sectors where these new technologies are/will be likely to be implemented is

the mining industry. Currently, the government has invested approximately $2.4

billion in growing Australia’s research to support a stronger and smarter economy and

$29.9 million to develop artificial intelligent and machine learning (Australian-

Government, 2020).

It is important to recognise that the lithium mining sector in Australia has not been

focused on the efficient management of the waste material it produces, but Industry

4.0 technologies could be applied to transform these materials.

Machine Learning

Machine learning is a quickly growing technical discipline combining statistics,

computer science and a variety of fields concerned with decision making under

uncertainty and continuing automatic improvements (Jordan & Mitchell, 2020).

Unlike traditional computational approaches where the computer operates under a

Chapter 2: Literature Review 21


hard-coded algorithm written by a specialist (Butler et al., 2018), ML gives us the

opportunity to develop algorithms which are able to generate predictions without being

explicitly programmed (Sen et al., 2018) (output) by assessing a portion of given data

(input) (Butler et al, 2018). Additionally, ML algorithms are able to learn from their

inputs to customize their services to the needs required (Jordan & Mitchell, 2020). ML

gives us the opportunity to accelerate the process of materials discovery and

optimisation (Evans & Coudert, 2017), but before setting up the algorithms it is

important to make sure how the dataset is going to be managed in every stage of the

process (i.e., data collection, data representation, data procession, choice of a learning

algorithm, and model optimisation).

Current applications of Machine Learning in zeolite synthesis.

Preparation of zeolites is usually defined by several connected variables which are still

reliant on trial-and-error methods. In the past decade ML methods have evolved to

solve complex problems including nonlinear or significantly combinatorial processes

that traditional methodologies cannot solve ( Moliner et al,, 2019). ML methods have

already proved to be useful tools in different fields such as chemistry, biochemicals,

drug design, computational biology (Carr, Lach-hab, Yang, Vaisman, & Blaisten-

Barojas, 2009), water desalination (Al-Shayji & Liu, 2002), and reverse and forward

osmosis (Cabrera et al., 2017; Jawad et al., 2020).

Considering the extensive and interconnected areas of zeolite synthesis, the use of ML

is exceptionally timely for enhancing zeolite synthesis. The complex space of zeolite

synthesis requires drawing inferences from partial and imperfect information, for

which ML methods are very well-suited, to replace the intuition-based methods used

for experimentation (Moliner, Y. Roman-Leshkov, et al., 2019).

Chapter 2: Literature Review 22


ML clustering analysis has proved that there are three groups of zeolites with the HEU

(heulandite) topology, S. Yang et al. (2010) used machine learning clustering analysis

of crystallography data of zeolite crystals. Before this study it was understood that the

heulandite family was divided into two mineral groups: heulandite (HEU-h) and

clinoptilolite (HEU-c). Their research proposed a new group based on the expectation–

maximization algorithm, HEU-m. It confirmed the existence of an intermediate group

between HEU-h and HEU-c.

Jensen et al. (2019) applied a machine learning approach to zeolite synthesis using

automatic literature data extraction for germanium-containing zeolites. To achieve

this, they trained two algorithms (random forest regression and decision tree model)

using data from a learning Python library, sci-kit. As a result, they developed an

automatic data extraction pipeline that localises, extracts, and formats synthesis data

for germanium containing zeolite predicting resulting zeolites topology.

Moliner et al. (2019) applied ML techniques to zeolite synthesis developing new

useful tools in significant machine learning areas such as data mining, generating large

amounts of data using High-Throughput (HT) platforms for synthesis and

characterisation. The enhancement of HT zeolite synthesis systems increased the

number of experiments and variables that can be analysed simultaneously and at the

same time all this information formed part of the development of data mining

techniques making it able to analyse the large quantity of data generated every time.

They used artificial neural networks (ANNs) to predict the internal relationships

between different synthesis variables and once this technique was properly trained the

Chapter 2: Literature Review 23


genetic algorithms (Gas) were used to successfully guide experimental design,

improving catalytic behaviour of Ti-silicates. Clustering using k-Means helped the

data classification in this research; the analysis used diffraction patterns to distribute

the data into three groups, proving that this technique has been found useful to identify

pure phases in mixed systems. Finally, this research illustrates the need for new ML

friendly representation of crystal structures, because the standard representation has

been developed for human understanding which will not be optimal for ML algorithms

(Moliner, et al., 2019).

This literature review shows us that analysing our data through machine learning

approaches can help future investigations to develop and improve zeolite synthesis

routes. It has also presented a new way to use our “unsuccessful” data, feeding the ML

algorithms with them to achieve the output expected. To date, there is no research

related to the use of machine learning focused on mining waste reuse. This represents

a great opportunity in the field of zeolites but also to help with the reduction of mining

waste material that, as previously mentioned, is causing environmental problems.

The use of machine learning in the zeolites field is challenging; one of the main

obstacles that this project has faced is the amount of data generated during the first

years of the research, since the efficacy of ML is constrained by the quantity and

quality of data (Moliner et al., 2019). However, possible approaches to zeolite

synthesis can be tested with just small data sets. Moreover, the lack of research into

zeolite synthesis from waste materials and the incomplete understanding of the

synthesis/crystallisation is a challenge that the project addresses. However, using a

Chapter 2: Literature Review 24


continuous system with real time measurements will help to improve the

understanding of the process.

Key insights

This chapter presented a comprehensive literature review of research on the

significance of recycling mining waste in Australia. With mining extraction generating

significant amounts of waste material, and decreasing landfill capacity, the need for

new methods has become increasingly crucial. Previous studies indicate that only 10%

of mining waste is reused, and hence recycling this material into high value products

such as zeolites can not only mitigate environmental problems but also provide

financially viable options for the mining industry.

Lithium slag and natural zeolites are significant waste products that predominantly

contain silica and aluminium, offering potential for recycling. However, current

methods are energy-intensive and pose significant scaling-up challenges for industrial

implementation. Therefore, there is a pressing need to develop new methodologies that

enable efficient transformation using green methods and that are easier to scale up for

industry.

In the next chapter, an activation technique will be developed to transform mining

waste materials into zeolitic products, utilizing the multi-step process to synthesize

zeolite LTA from mother liquor and solid product.

Chapter 2: Literature Review 25


Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate

Residue via Cancrinite Formation

Abstract

Waste materials such as Spodumene Leachate Residue (SLR) from lithium ores are increasing

in volume due to higher demand for lithium batteries. Transformation of SLR into higher value

zeolite LTA is of interest if the critical problem of low cost SLR activation can be solved.

Therefore, this study made zeolite LTA via a cancrinite phase. Preferred conditions for making

cancrinite involved heating SLR at 240 oC for one hour in 7 M sodium hydroxide (NaOH). In

one approach, the silicate rich mother liquor was mixed with sodium aluminate (NaAlO2) and

water to adjust NaOH molarity to form zeolite LTA (68 wt %). Alternatively, a process

wherein NaAlO2 was added to adjust the silica to alumina ratio to 2 in the starting mixture was

found to produce low quality cancrinite (26.5 wt% cancrinite, 13 wt% sodalite and 59 wt%

amorphous). Improvements were observed when a mixture of cancrinite and NaAlO 2

underwent acid dissolution to extract available Si and Al, followed by neutralization and

subsequent hydrothermal reaction. Notably 2 M hydrochloric acid (HCl) was preferred for the

acid dissolution step and neutralization of the acidic solution was best with 4 M NaOH. The

zeolite LTA content was highest (63.1 wt%) when aluminium content in the reactant mixture

was greatest. The inherent excess of aluminium present in the mother liquor and wash water

may be recycled along with residual NaOH. However, the presence of excess sodium chloride

may inhibit recycling of mother liquor/wash water and thus the economics may not be

attractive.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 26
`
Key Words: spodumene; zeolite; cancrinite; lithium; waste

Introduction

Lithium demand is projected to increase substantially in the next few years due primarily to

the growth in demand for batteries for electric vehicles and renewable energy storage

(Dessemond et al., 2019). Notably, lithium is mainly sourced from either hard rock deposits

such as spodumene (LiAlSi2O6) or lithium brines ( Flexer et al., 2018). Currently, hard rock is

the largest source of lithium for commercial use, with Australia supplying >50 % of market

needs (Hao et al., 2017). In Australia, the lithium sources are located in Western Australia and

mined by various companies (Karrech et al., 2020).

Despite the economic and social benefits of lithium mining there exists a challenge in terms of

the waste produced when lithium is extracted from ore bodies. Typically, lithium ores such as

α-spodumene are initially roasted at high temperatures to make the more reactive β-spodumene

phase and then leached using sulfuric acid (P. Meshram et al., 2014). The residual

aluminosilicate waste material is termed lithium slag or spodumene leachate residue (SLR)

(HAlSi2O6). This material is produced at a ratio of approximately 10 tonnes of lithium residue

for every one tonne of lithium carbonate made (Z-h. He et al). At present, only a small fraction

of SLR finds use in applications such as ceramic glazed tiles, concrete, or cement clinker (G.

Han et al., 2018). Hence, there is a need to find innovative means to recycle this waste stream

into valuable products.

As SLR is mainly comprised of aluminosilicate materials, it is logical to examine whether this

waste can be used as a resource for making higher value products such as zeolites (X. Wang et

al., 2021). Synthetic zeolites are mainly aluminosilicate-based minerals and employed in

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 27
`
applications including catalysis (Weitkamp, 2000), molecular sieving (Davis, 1991), gas

separation (Kosinov et al., 2016), green chemistry (H. Li et al., 2016), water/wastewater

treatment (Koshy & Singh, 2016), and laundry detergents (Fruijtier-Pölloth, 2009).

Commercially, zeolite LTA (Na 12Al12SiO48.28H2O) is manufactured in the greatest volume as

it is primarily employed in detergent formulations to soften water (El-Nahas et al., 2020) and

for drying of natural gas (Shirazian & Ashrafizadeh, 2015).

Traditionally, many zeolites are made via an aluminosilicate gel which is formed by

appropriate mixing of alkaline solutions of sodium silicate and sodium aluminate followed by

hydrothermal synthesis for several minutes to hours ( Cundy & Cox, 2005). The recovered

“wet cake” is subsequently washed and dried in the oven ( Probst et al., 2021). In contrast,

making zeolite LTA from wastes such as SLR is more demanding as an additional

aluminosilicate depolymerisation step is required to enhance the reactivity of the waste material

(Collins et al., 2020). Activation strategies include thermal treatment at elevated temperatures

(H. Liu et al., 2014), solid alkali fusion (Ayele et al., 2016b), high concentration alkali solutions

(HCAS) (T. Li,et al., 2012) and the quasi-solid phase activation process (J. Yang et al., 2017).

Activation of clay materials including kaolin, rectorite, illite, and montmorillonite has been

accomplished by the HCAS method (T. Li et al., 2012; Liu et al., 2015; Yue et al., 2014; Y.

Yue et al., 2015). This approach uses high molarity solutions of sodium hydroxide (for

example 15 M) at temperatures in the range 200 to 250 oC for several hours to produce zeolite

P and zeolite Y. However, the HCAS strategy is not industrially viable due to problems such

as evaporation of large amounts of water during the heating stage and consequently excessive

energy consumption (H. Liu et al., 2015). An alternative activation approach was developed

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 28
`
by J. Yang et al. (2017) and termed Quasi-Solid Phase Activation Process (QSP). In this case,

kaolin was kneaded with sodium hydroxide pellets and minimal amounts of water; and then

extruded. The extrudates were subsequently heated in the range 60 to 100 oC for 30 minutes

in a belt type calcination oven. Yue et al. (2020) demonstrated, using the QSP approach, that

>95 % depolymerisation of the rectorite was obtained by heating at 170 oC for 1 h.

Subsequently ZSM-5 zeolite was made after reaction at 70 oC for three hours. The principal

advantages relative to the HCAS approach were: (1) lower activation temperature; (2) less

water required; (3) continuous processing (as opposed to batch for HCAS); and (4) reduced

energy consumption. However, the QSP process has the disadvantage that it needs to evaporate

water at significant cost from recycled sodium hydroxide mother liquor to form the solid

sodium hydroxide used in the initial kneading phase. If excess sodium hydroxide in the mother

liquor is not reused, then the economics can become unfavourable (Ghrear et al., 2020).

With regard to SLR activation and the ensuing synthesis of value added products, the formation

of FAU/LTA zeolites has been reported by . Lin, Zhuang, Cui, Wang, and Yao (2015). The

basic strategy was to initially activate SLR by heating with NaOH at a moderate temperature

followed by mixing with water and heating at elevated temperature for 9 h. Notably, not only

was 48 % of the mother liquor recycled to the initial activation stage but also the zeolite

composition changed markedly depending upon the order of aluminium addition. The

synthesis of nano-kaolinite and xonotlite (6CaO.6SiO 2.H2O) from SLR has also been described

(X. Wang et al., 2021). A hydroxycancrinite intermediate phase was formed by heating SLR

and 6 M NaOH at 220 oC for four hours. As a result of heating the hydroxycancrinite in nitric

acid at 260 oC for 24 h, kaolinite was synthetised. It is noted that an undesirable sodium nitrate

waste stream was created by this approach.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 29
`
The high silicate containing mother liquor produced from the SLR to cancrinite process was

reacted with lime under hydrothermal conditions for two hours to make hydrated calcium

silicate. A second hydrothermal stage was then employed to make xonotlite by heating at 260

oC for 12 h. The initial growth of cancrinite and subsequent reaction to form zeolite materials

has also been reported in studies of kaolin/quartz to zeolite Y (X. Wang & Nguyen, 2016).

The fundamental premise of activation of waste aluminosilicates to form cancrinite has merit

in that it operates at significantly lower temperature than an alkali fusion approach (Ayele et

al., 2016b). However, there remain several gaps in the process knowledge. For example the

study by X. Wang et al. (2021) made products from SLR which may not have a substantial

market in key lithium producing countries such as Australia. Secondly, the extensive use of

high temperature conditions (e.g., heating calcium hydroxide at 900 oC for 2 h) could result in

a process which is prohibitively expensive. The generation of a waste sodium nitrate solution

is also not desirable. Additionally, the presence of salt in the neutralization stage may promote

or inhibit zeolite formation. Nevertheless, previous studies of transforming SLR into value

added products confirm that strategies such as mother liquor recycling and cancrinite formation

are of interest.

Therefore, the aim of the study described in this chapter was to determine the feasibility of

converting SLR to zeolite LTA via a cancrinite intermediate phase. The hypothesis was that if

SLR processing was better understood then zeolite synthesis may be made more cost effective.

The research questions to support this hypothesis were: (1) Can pure zeolites be made from a

dual process? (2) How can the formation of high-quality cancrinite be achieved when

additional aluminium species are added to the feed mixture? (3) What factors influence the

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 30
`
formation of cancrinite? (4) What are the optimal conditions for the acid dissolution and

neutralization steps in the process of synthesizing zeolite LTA from the cancrinite/sodium

aluminate mixture? (5) How do different acid and alkali concentrations affect the quality of the

final product?

The activation of SLR and subsequent zeolite synthesis were conducted at bench scale to prove

the concept in this investigation. Comprehensive analysis of both solid and aqueous streams

was completed to determine product quality and understand the chemistry. A process flow

diagram was created which included mass balance and zeolite yields. Zeolite LTA was

targeted, as it is the largest volume zeolite synthesised by industry due to its use by the detergent

and gas industries.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 31
`
Materials and Methods

Spodumene Leachate Residue and Chemicals


The Spodumene Leachate Residue (SLR) for this study was supplied from a lithium mine in

Western Australia. Analysis by X-ray Fluorescence indicated that the main species present

were SiO2 (66.6 wt%) and Al2O3 (23.7 wt%) (Supplementary Table 1). Quantitative X-ray

diffraction suggested the major components were Hydrogen Aluminium Silicon Oxide (46.9

wt%), quartz (3.93 wt%), lithium sulfate (0.30 wt%) and non-diffracting/unidentified material

(48.86 wt%) (Supplementary Table 2). The particle size of the SLR ranged between 0.76 and

240 µm with a d50 = 47.61 µm and a d90 = 153.02 µm (Supplementary Figure 1). The SLR

grains were relatively featureless (Supplementary Figure 1). Sodium aluminate (NaAlO2) was

purchased from Sigma-Aldrich. Hydrochloric acid (HCl; 32 wt %) and sodium hydroxide

(NaOH; extra pure micro-pearls & 40 wt% solution) were procured as an analytical reagent

(AR) grade from ChemSupply.

Synthesis of Zeolite LTA from Spodumene Leachate Residue


Three general pathways to making zeolite LTA from SLR via a cancrinite stage were

investigated. In Process 1 two crystalline products were made (zeolite LTA and cancrinite)

(Figure 1), whereas Process 2 further reacted cancrinite produced from the high temperature

hydrothermal activation stage to make zeolite LTA. Notably, additional NaAlO2 was added to

the high temperature hydrothermal reactor to adjust the SiO 2/Al2O3 ratio to 2 (Figure 2).

Finally, Process 3 was a variation of Process 2 wherein the additional NaAlO 2 was mixed with

the cancrinite product in contrast to the high temperature activation stage (Figure 3).

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 32
`
Process 1
The aim of this stage was to determine the impact of sodium hydroxide molarity (4 to 10 M),

temperature (100 to 240 ℃), and reaction time (1 to 24 hours) upon the transformation of SLR

to cancrinite. SLR (10 g) was mixed with NaOH (1.917 g NaOH/g SLR), and deionized water

added to adjust NaOH molarity. For reactions at 100 and 150 oC this latter mixture was placed

into a Berghof BR100 High Pressure Reactor. For reactions at higher temperatures (200 and

240 oC) samples were placed in a PTFE lined 100 mL acid digestion vessel (Amar Equipment

PVT Ltd., India) and subsequently located in a rotary oven (KLJX Homogeneous reactor). In

this case the volume of the samples was adjusted to accommodate the reduced equipment

capacity by mixing 5 g of SLR, sodium hydroxide (1.917 g NaOH/g SLR) and deionized water

to control NaOH molarity.

Figure 1: Process flow diagram for making both cancrinite and zeolite LTA from Spodumene
Leachate Residue

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 33
`
All reactant mixtures were agitated at 100 rpm and hydrothermal reaction time was defined as

the time after reaching the set temperature. Each reaction allowed for 60 minutes cooling

before collecting the product slurry. After the reaction was complete the mother liquor and the

solid were separated using an Allegra X-30 Centrifuge. The wet cake was washed with 240

mL of deionised water to reduce pH (pH 11 to 11.5) and finally dried in an oven at 105 ℃

overnight. Once the preferred synthesis conditions for cancrinite were obtained, larger

volumes of this material were made using a 1 L hydrothermal reactor (Amar Equipment, India).

In this case, SLR (60 g), NaOH (115.02 g) and deionized water (410.82 g) were reacted at 240

℃, with agitation (100 rpm) for 1 h.

The resultant mother liquor was then modified by addition of NaAlO 2 to adjust the SiO2/Al2O 3

ratio to 2 along with sufficient water to adjust the sodium hydroxide molarity to 4 M. Zeolite

synthesis was performed by placing sealed containers in a New Brunswick Innova 42R

Incubator for four hours at 80 ℃ while being agitated at 100 rpm. When the synthesis was

complete the slurry was centrifuged to recover the solid product, which was washed multiple

times with deionized water to reduce pH from 14 to 11. During each cycle 30 mL of water was

used. Once the preferred pH was obtained, the slurry (wet cake) was weighed and dried in an

oven at 105 ℃ overnight. The mother liquor from the zeolite synthesis stage was recycled to

the initial high temperature activation stage.

Process 2
The high temperature activation stage was conducted similarly to the corresponding approach

in Process 1. The main difference was the addition of 31.8 g of sodium aluminate along with

SLR (10 g), NaOH (1.917 g /g SLR), and deionized water to adjust NaOH molarity.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 34
`
Cancrinite dissolution trials involved addition of cancrinite (1 g) and hydrochloric acid

solutions (10 mL) of the desired molarity (1, 2, 3, 4, 5, 6, 7, 8 M) into a 50 mL centrifuge tube.

The solution was agitated with a rotary tube mixer for 15 min. Subsequently the samples were

syringe filtered and mother liquor was collected for analysis to determine the presence of major

cations in solution. All experiments were performed in triplicate. Cancrinite (2 g) was added

to a centrifuge tube wherein HCl (20 mL) was added to dissolve the cancrinite. This solution

was then neutralised with varying concentrations of NaOH solution (20 mL). Zeolite synthesis

was performed as described in Process 1.

Figure 2: Process flow diagram for the synthesis of zeolite LTA from Spodumene Leachate
Residue via cancrinite formation; Additional NaAlO 2 added to the feed stream to the high
temperature hydrothermal reactor.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 35
`
Process 3
A variation of Process 2 involved addition of sodium aluminate to the cancrinite product

instead of during the hydrothermal activation stage (Figure 3).

Figure 3. Process flow diagram for the synthesis of zeolite LTA from Spodumene Leachate
Residue via cancrinite formation; Additional NaAlO 2 added to the cancrinite product.

By adding the extra sodium aluminate directly to the cancrinite produced from high

temperature hydrothermal reaction, it was possible to synthesise zeolite from the silica species

present in the mother liquor, as in Process 1.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 36
`
Solution Analysis

Inductively Coupled Plasma Optical Emission Spectroscopy (ICP-OES)


The concentration of sodium, aluminium and silicon from the acid dissolution step and zeolite

synthesis (mother liquor) was determined using a Perkin Elmer Optima 8300 DV ICP-OES

spectrometer. Prior to analysis the samples were filtered through a syringe filter and diluted to

1:100 and 1:1000 with 2% nitric acid (diluent) using a Hamilton auto-diluter.

Ion Chromatography (IC)


Solutions were syringe filtered prior to analysis and diluted using Milli-Q water in a Hamilton

auto-diluter. Chloride analysis was completed using a Dionex ICS-2100 ion chromatography

system. The flow rate used for the suppressed ion chromatographic method was 1.0 mL/min

with an eluent strength of 28.0 mM KOH and an injection volume of 25.0 mL. The elution time

was 10 minutes with a chloride retention time of 3.16 min. The instrument was connected to a

Dionex ICS AS-DV autosampler.

Solid Characterization

Quantitative X-ray Diffraction (XRD)


A PANalytical X’PertPro powder diffractometer was used to identify the crystalline phases in

the final products. The samples were initially mixed with 10 wt% corundum standard powder

(Al2O3). The sample (0.45 g) and standard (0.05 g) were micronized in a McCrone micronizing

mill with 15 mL ethanol; and dried in an oven for four hours at 40 ℃. Dried samples were

collected, and front pressed into sample holders. To identify the product phases JEdit and

TOPAS software were used. The reference patterns for zeolite LTA were obtained from Breck

et al., 1956) and atomic parameters were cross-referenced from PDF entry 04-017-3644.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 37
`
X-ray Fluorescence (XRF)
PANalytical WROXI software application and a PANalytical AXIOS 1.5 kW Wavelength

Dispersive XRF spectrometer were used to determine the elemental quantification of the solid

products. Solid samples (1 g) were mixed with lithium tetraborate: lithium metaborate (50:50)

with 0.05% li-Iodate (9 g) and heated at 1050 ℃ to form a fused glass disc.

Scanning Electron Microscopy (SEM)


Shape and size of zeolite crystals were observed using a JEOL 7001F Scanning Electron

Microscope. The accelerating voltage employed was 5 kV, with a probe current of 10 A at a

magnification between 4000 and 10000x. Each sample was prepared on carbon stamps and

coated with platinum prior to observation.

Particle Size Distribution


A Malvern Mastersizer 3000 instrument equipped with dispersion chamber, ultrasound and

automatic cleaning was used to determine the particle size of the Spodumene Leachate Residue.

Powder was added along with deionized water into the dispersion chamber to reach 15%

obscuration with sonication. Measurements were repeated five times for 30 s and the reference

material used for the program was zeolite LTA.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 38
`
Results and Discussion

Process 1: Synthesis of Cancrinite and Zeolite LTA from SLR

Synthesis of Cancrinite from SLR and Sodium Hydroxide Aqueous Solution


The first stage investigated the impact of reaction temperature (100 to 240°C), NaOH molarity

(4, 5, 7, 10 and 14 M) and reaction time (1, 15 and 24 h) upon cancrinite formation (Figure 4).

100
1h Synthesis 100
wt% Crystalline Cancrinite

1h Synthesis
24h Synthesis
24h Synthesis
80

wt% Non-Diffracting
80

60 60

40 40

20 20

0
0 4 6 8 10 12 14
4 6 8 10 12 14
NaOH Molarity
NaOH Molarity

wt% Crystalline Cancrinite wt% non-Diffracting/unidentified


Reaction Temperature = 150 oC

100
1h Synthesis 100
wt% Crystalline Cancrinite

1h Synthesis
15hrs Synthesis 15h Synthesis
80 24hrs Synthesis
wt% Non-Diffracting

80 24h Synthesis

60 60

40 40

20 20

0
0
4 6 8 10 12 14
4 6 8 10 12 14
NaOH Molarity
NaOH Molarity

wt% Crystalline Cancrinite wt% non-Diffracting/unidentified


Reaction Temperature = 200 oC
Figure 4. Influence of time and NaOH molarity upon cancrinite synthesis from SLR at 150 and
200 ℃ a) Crystalline cancrinite, b) non-Diffracting material. Uncertainties were calculated with
a 95% confidence interval.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 39
`
Reaction at 100 oC with 10 M NaOH solution for 24 hours produced crystalline cancrinite

(Na6Ca1.5(CO3)1.5(Si6Al6O24).2H2O) at 10.4 wt %; in addition to a dominant non-diffracting

fraction of 74.2 wt%, sodalite (9.8 wt%), and quartz (SiO 2) (3.9 wt%) and a small amount of

zeolite LTA (1.7 wt%) (Supplementary Table 3). Murukutti and Jena (2022) indicated that

extended reaction of coal fly ash at 100 oC promoted the transition from zeolite LTA to

hydroxysodalite. It is noted that the XRD pattern for cancrinite was best fitted to the pattern

for Na6Ca1.5(CO3)1.5(Si6Al6O24). However, from the XRF analysis (

Supplementary Table 6) it was apparent that the amount of calcium present in the cancrinite

material was substantially less than the quantity of Si or Al present (oxide mass ratios relative

to CaO of 90 and 32, respectively). Studies of coal fly ash conversion to zeolitic materials

suggested that cancrinite with the formula (Na 8Al6Si6O24(OH)2.2H2O) was formed when

activating fly ash with sodium carbonate at 200 oC for 20 hours (Murukutti & Jena, 2022).

Notably, XRD is a phase analysis technique, not a chemical one. Thus, the results derived

from modelling can match phases or specific crystal structures to patterns with a certain degree

of confidence based on the pattern intensity, the position (Bragg angle or interplanar spacing) ,

and intensities of the diffraction maxima (peaks) (H. Khan et al., 2020). In this case the samples

fitted to a cancrinite pattern with extra Ca because the match was focused on specific structural

details where all members have the same structure. In order to determine the chemical structure

of the material it is important to support XRD results with XRF/elemental results.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 40
`
The data recorded for activation of SLR at 100 oC suggested that conditions were not sufficient

to fully transform the starting mixture to the more thermodynamically stable cancrinite phase

( Hackbarth et al., 1999). This conclusion was supported by reports that cancrinite synthesis

from kaolinite or coal fly ash was observed at significantly higher temperatures such as 200 oC

(Murukutti & Jena, 2022; Wernert et al., 2020) or 250 ℃ (Selim et al., 2018).

Therefore, the reaction temperature was increased to 150 oC while using 4 to 10 M NaOH

solutions. After one hour at 150 oC, the formation of cancrinite in greater amounts (20 wt %)

was observed when compared to the 100 oC analogues (Figure 4). Moreover, there was a

distinct trend whereupon increasing NaOH molarity promoted cancrinite synthesis. See Figure

3. Process flow diagram for the synthesis of zeolite LTA from Spodumene Leachate Residue

via cancrinite formation; Additional NaAlO2 added to the cancrinite . Increasing the reaction

time to 24 hours promoted the formation of cancrinite, with 79 wt% identified when reacted in

10 M NaOH. Even at the lower NaOH molarity of 4 M, the amount of cancrinite recorded was

63 wt%. It was noted that when cancrinite was synthesised, corresponding decreases in the

presence of non-diffracting material (Figure 3) and quartz were evident. Sodalite formation

appeared to be inhibited, especially if reaction occurred for 24 hours (Supplementary Table 4),

which was in accord with previous literature (Murukutti & Jena, 2022). This behaviour was

consistent with sodalite being an intermediate phase transformed to the more

thermodynamically stable cancrinite under high molarity and temperature conditions

(Limlamthong et al., 2021).

Based upon the data acquired at reaction temperatures of 100 and 150 oC it was postulated that

not only NaOH molarities >10 M promote cancrinite formation but also that at higher reaction

temperatures the formation of cancrinite may be accelerated (Figure 3). Accordingly, using

five M NaOH solutions for 1 hour at 200 oC enhanced cancrinite synthesis (38 wt %). As

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 41
`
reaction time increased to 15 and then 24 hours the cancrinite phase was raised to 42 and 68

wt%, respectively. Further increasing the NaOH molarity to 7 M initially promoted the

formation of cancrinite (44 wt%) after one hour of reaction but a slight inhibition of cancrinite

formation was observed when the reaction time was extended (relative to 5M NaOH). Use of

10 M NaOH solutions generally improved the cancrinite formation up to 71 wt% after 24 hours

reaction. However, the appearance of sodalite was noted at this molarity. and clearly, when

using 14 M NaOH, the presence of cancrinite was substantially reduced. Simultaneously the

presence of non-diffracting species and sodalite was enhanced (Supplementary Table 5).

Overall, activation times of 24 hours may not be practical due to excessive energy costs; thus,

the reaction temperature was further elevated to 240 oC while fixing the reaction time as one

hour. In addition, the NaOH molarity was adjusted to either 7 or 10 M with the aim of

promoting cancrinite formation (Table 1). With 10 M NaOH solutions a dominant non-

diffracting component was detected. Concurrently a small amount of cancrinite crystals (15

wt%) was formed. Indeed, the major crystalline phase was identified to be sodalite. In

contrast, reacting SLR at 240 ℃ for one hour with 7 M NaOH substantially increased cancrinite

content to 87.2 wt%.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 42
`
Table 1. XRD analysis of solid products resultant from SLR activation at 240 oC as a function
of NaOH molarity and reaction time
10M NaOH 7M NaOH
Phase 1 hour synthesis 1 hour synthesis (wt%)
(wt%)
Crystalline Quartz (%) 0.0 0.5
Crystalline Cancrinite (%) 15.0 87.2
Crystalline Sodalite (%) 17.60 0.2
Crystalline Zeolite LTA (%) 1.3 2.8
Non-Diffracting/unidentified (%) 66.1 9.3

Hackbarth, et al. (1999) found that at elevated temperatures such as 200 oC, increasing the

NaOH molarity from 12 to 16 M resulted in transformation of cancrinite to sodalite. It is

therefore probable that at a temperature of 240 oC the NaOH molarity would further need to be

reduced. Scanning electron microscope (SEM) images of samples synthesised at 150 and 240

oC showed the presence of needle shaped cancrinite crystals in each instance (Figure 5) (Deng

et al., 2006).

a) 150 o C, 10 M NaOH and 24 hours b) 240 oC, 7 M NaOH and 1 hour


Figure 5: SEM images of solid material created by NaOH treatment of SLR at temperatures
of 150 and 240 oC

As indicated above, the energy consumption is important in terms of engineering optimizat ion

of the SLR to cancrinite method. Thus, energy demand for cancrinite synthesis at 240 oC for

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 43
`
one hour (7 M NaOH, 87.2 wt%) was compared with SLR activation at 150 oC for 24 hours

(10 M NaOH, 79 wt%) using temperature data acquired every minute during heating and

cooling periods. Equation 1 (Luscombe, 2018) was employed to estimate the heat energy input

per minute, using the measured temperature at that step compared to the initial temperature.

Thus, a range of average heat energy values per minute for each time step was generated. The

median of these values allowed calculation of the energy added per minute when the reactor

was heated to reaction temperature.

Equation 2. Heat Energy 𝑸/𝒕 = 𝒎 ∗ 𝑪𝒑 ∗ 𝒅𝑻/𝒕

Where Q is the heat energy (J), m is the mass (g), Cp is the heat capacity (J/kg. K), dT is the

change in temperature (K), and t is the time (s). Once the vessel reached reaction temperature,

the temperature was maintained for the specified reaction time prior to cooling, when reaction

was completed. To estimate the heat losses in the system, and therefore the energy required to

hold the temperature for the set time, a similar approach was taken using the data during the

cool down period. Once again, heat energy was calculated per minute and a median value was

taken, which was the heat loss through the vessel due to radiation. The total heat required was

calculated using Equation 3 (Luscombe, 2018). The heating cycle to reach the required

activation temperature was considered as well as the time that the reaction temperature was

held (24 hours or 1 hour).

Equation 3. Total Energy Required 𝑸𝑻𝒐𝒕𝒂𝒍 = 𝑸𝑰𝒏 ∗ 𝒕𝑯𝒆𝒂𝒕 + 𝑸𝑳𝒐𝒔𝒔𝒆𝒔 ∗ 𝒕𝑯𝒐𝒍𝒅

The results of this calculation revealed that heating the sample at 150 oC for 24 hours consumed

3.7 times more energy than activation at 240 oC for one hour (Supplementary Figure 4).

Additional benefits of using the shorter reaction time included greater throughput which may

translate to reductions in capital expenditure through smaller equipment requirements.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 44
`
Subsequently, using the optimized SLR activation procedure of 240 oC and 7 M NaOH for 1

h, bulk material was produced in 1 L hydrothermal reactors to provide sufficient material for

the remaining tests in this study. The average amount of crystalline cancrinite observed in the

solid product after 5 repetitions was 87 ± 5 wt%.

Mass Balance for Synthesis of Cancrinite from SLR and Sodium Hydroxide Aqueous
Solution
As the activation conditions for SLR to cancrinite have been determined, this situation allowed

a mass balance calculation to identify important criteria such as mother liquor composition and

process yields of cancrinite (Figure 6). The process illustrated was for heating the starting

mixture at 240 oC with 7 M NaOH for one hour. In terms of overall mass balance between the

starting materials and solid/liquid products, 100 % mass balance for Si (18.7 g), Al (7.5 g) and

Na was obtained (66.1 g), whereas a slight discrepancy was noted for the water content (Inlet

410.8 g and outlet 402.1 g). This data suggested that water losses may have occurred during

processes such as centrifuging the sample after hydrothermal reaction to collect solids and

mother liquor.

More specifically the mother liquor and wash water analysis revealed that a total of 9.1 g of Si

had reacted to form the solid product, whereas 9.6 g had unreacted. On this basis the

SiO2/Al2O3 molar ratio of the cancrinite was reduced to 2.32 (c.f. starting value of 4.8 for SLR).

XRF analysis of the cancrinite product (

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 45
`
Supplementary Table 6) similarly estimated a SiO2/Al2O3 molar ratio of 2.5. X. Wang et al.

(2021) also found that cancrinite made from alkali activation with NaOH produced cancrinite

with a SiO2/Al2O3 molar ratio of 2.62. It is noted that the difference in the product composition

(Na9(CO3)1.5 (Si6 Al6 O24).2H2O) and the XRD database pattern

(Na6Ca1.5(CO3)1.5(Si6Al6O24).2H2O) was substitution of sodium ions for calcium ions (which

were not available when SLR was activated with NaOH). Carbonate-cancrinite was more

probable than hydroxy-cancrinite as the measured Loss on Ignition (LOI) of 9.5 wt% was closer

to Na9 Al6 Si6 O24 (CO3 )1.5 . 2H2 O (LOI = 7.5 wt%) rather than (Na8Al6Si6O24(OH)2.2H2O) (LOI

= 3.5 wt%).

Consequently, the stoichiometric conversion of SLR to cancrinite was approximated as shown

in Equation 4

Equation 4: 𝟑 𝑯𝟐 𝑨𝒍 𝟐𝑺𝒊𝟒.𝟓𝑶𝟏𝟑.𝟏. 𝟓𝑯𝟐 𝑶 + 𝟐𝟒 𝑵𝒂𝑶𝑯 + 𝟏. 𝟓 𝑪𝑶𝟐 →


𝑵𝒂𝟗 𝑨𝒍 𝟔 𝑺𝒊 𝟔 𝑶𝟐𝟒(𝑪𝑶𝟑 )𝟏.𝟓.𝟐𝑯𝟐 𝑶 (𝒔 ) + 𝟕. 𝟓 𝑵𝒂𝟐 𝑺𝒊𝑶𝟐(𝑶𝑯)𝟐 (𝒂𝒒) + 𝟏𝟎 𝑯𝟐 𝑶

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 46
`
Figure 6: Process flow diagram including stream tables for (1) conversion of SLR to cancrinite and (2) SLR mother liquor to zeolite LTA

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 47
`
The conversion represented in Equation 3 has a theoretical yield of 48.8 g of cancrinite and 47 g of

dissolved silicate from 60 g of SLR material. The mass balance reported in Figure 6 shows a mass

yield of 50.16 g of cancrinite which represented a 103 % yield (essentially 100 % as inherent errors

in measurement were involved). The theoretical yield of 9.4 g of dissolved silicon matched with the

observed silicon amount of 9.59 g in the mother liquor/wash water. The agreement between the mass

balance and the theoretical yield indicated that Equation 3 was a good approximation of the process.

Synthesis of Zeolite LTA from Recycled Mother Liquor after Cancrinite Synthesis
After cancrinite synthesis, an excess of 20.5 g/L dissolved active silicon species was found in the

mother liquor as well as an excess of NaOH. Upon addition of water to adjust the NaOH molarity (4

M) and sodium aluminate to adjust the SiO 2/Al2O3 ratio to 2, an aluminosilicate gel was rapidly

formed, which was then heated in an incubator at 80 oC. XRD results in Table 2 revealed that zeolite

LTA was formed with negligible by-products such as quartz, cancrinite or sodalite detected.

Table 2. XRD results after the mother liquor recycling process.


XRD phases (wt%) 2-hour synthesis 3-hour synthesis 5-hour synthesis

Quartz 0.2 0.1 0.2

Cancrinite 0.0 0.1 0.0

Sodalite 0.0 0.0 0.0

Zeolite LTA 63.7 67.8 68.9

Non-diffracting 36.1 32.0 31.1

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 48
SEM images confirmed the appearance of zeolite LTA rounded cubes (Figure 7). Bronić, Palčić,

Subotić, Itani, and Valtchev (2012) also reported the presence of zeolite LTA crystals with truncated

edges. Similarly, Basaldella, Kikot, and Tara (1997) intimated that when using SiO2/Al2O3 ratios in

the starting mixture between 1.48 and 1.99 the final zeolite LTA product was cubic with bevelled

edges. To obtain cubes with sharp edges it was necessary to increase the SiO 2/Al2O3 ratio to >2.18.

The effect on LTA nucleation is due to the change in the overall concentration of the species present

in solution, which is higher when more concentrated NaOH is used to neutralize it, as suggested in

the context of promoting crystal formation, accelerating nucleation, and influencing the crystal

growth mechanism.

The success of the mother liquor modification and subsequent synthesis of zeolite LTA was

consistent with published studies of other zeolite synthesis approaches. For example . C. Liu et al.,

(1998) showed that after zeolite crystallization the mother liquor contained sodium silicate species.

In this case, aluminium sulfate was added, which in turn reacted with the active silicon species in the

mother liquor to form an aluminosilicate gel. Zeolite Y was subsequently synthesised, with the caveat

that the presence of zeolite crystals in the mother liquor (>2 wt%) may compromise the quality of the

zeolite product.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 49
`
Reaction time
2 hours 3 hours 5 hours

4,000x 4,000x 4,000x

10,000x 10,000x 10,000x


Figure 7: SEM images of the zeolite product after hydrothermal reaction of mother liquor modified
by water and sodium aluminate addition; Reaction temperature 80 oC; Time = 2 to 5 hours

This study of transforming SLR to cancrinite and zeolite LTA products indicates a key aspect of its

success appears to be the value of the cancrinite. Investigations of cancrinite applications describe

the potential for cancrinite to be used as a sorbent for heavy metal ions such as Pb (II) and Cu (II)

ions when added to contaminated soils (Zheng et al., 2021). Similarly, Wernert et al. (2020)

described the successful use of cancrinite to remove Cd (II) and Pb (II) species from water.

Alternatively, Đặng et al., (2021) indicated that cancrinite catalysed the transesterification reaction

of palm oils and soybean with methanol to make biodiesel. Hence, the dual product approach

described in this study may have merit and is worthy of further exploration.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 50
`
Process 2: Synthesis of Zeolite LTA from SLR via Cancrinite Formation

Addition of NaAlO2 to the SLR Activation Mixture


As indicated in the mass balance for synthesis of cancrinite from SLR and sodium hydroxide aqueous

solution, it was apparent that the mother liquor comprised of unreacted sodium silicate species. To

recycle the mother liquor, it is necessary to remove the dissolved silicate species as these will

accumulate over time. One strategy was to add additional sodium aluminate along with the SLR and

NaOH to create a mixture with SiO2/Al2O3 ratio = 2. (Equation 5).

Equation 5: 𝟐 𝑯𝟐 𝑨𝒍𝟐 𝑺𝒊𝟒.𝟓 𝑶𝟏𝟑. 𝟏. 𝟓𝑯𝟐 𝑶 + 𝟖. 𝟓 𝑵𝒂𝑶𝑯 + 𝟓 𝑵𝒂𝑨𝒍𝑶𝟐 + 𝟐. 𝟐𝟓 𝑪𝑶𝟐 →


𝟏. 𝟓 𝑵𝒂𝟗 𝑨𝒍𝟔 𝑺𝒊𝟔 𝑶𝟐𝟒 (𝑪𝑶𝟑 )𝟏.𝟓 .𝟐𝑯𝟐 𝑶 (𝒔) + 𝟔. 𝟐𝟓 𝑯𝟐 𝑶

In practice, it was necessary to add sodium hydroxide solution in excess of stoichiometric amounts

to a prescribed value which promotes cancrinite formation (Equation 6).

Equation 6: 𝟐 𝐇𝟐 𝐀𝐥 𝟐𝐒𝐢𝟒.𝟓 𝐎𝟏𝟑 .𝟏. 𝟓𝐇𝟐 𝐎 + 𝟒𝟎 𝐍𝐚𝐎𝐇 + 𝟓 𝐍𝐚𝐀𝐥𝐎𝟐 + 𝟐. 𝟐𝟓 𝐂𝐎𝟐 + 𝟑𝟏𝟒 𝐇𝟐 𝐎


→ 𝟏. 𝟓 𝐍𝐚𝟗 𝐀𝐥𝟔 𝐒𝐢𝟔 𝐎𝟐𝟒 (𝐂𝐎𝟑 )𝟏.𝟓.𝟐𝐇𝟐 𝐎 (𝐬) + 𝟑𝟏. 𝟓 𝐍𝐚𝐎𝐇
+ 𝟑𝟐𝟎. 𝟐𝟓 𝐇𝟐 𝐎

Figure 8 shows the collected data when additional aluminate was added to the synthesis mixture in
order to adjust SiO2/Al2O3 ratio to 2.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 51
`
Figure 8: Process flow for cancrinite synthesis from a starting mixture containing SLR, NaOH,
water and sodium aluminate; one hour at 240 oC in hydrothermal reactor.

The solid product resultant from SLR activation at 240 oC for one hour was composed of 26.5 %

cancrinite, 13 wt% sodalite and 59 wt% amorphous (as measured by quantitative XRD). In accord

with the presence of extra soluble aluminium species the amount of silica found in the mother liquor

was substantially less than the case when additional aluminium was not present (ca. 0.64 g compared

to 6.67 g, respectively). However, despite the increased reaction of dissolved silicates with aluminate

species, this process did not enhance the presence of crystalline cancrinite. Inspection of SEM images

for the solid product (Figure 8) revealed that the presence of the cancrinite pillars was significantly

less than observed when additional aluminium was present (Figure 6).

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 52
`
The underlying reason for the inhibition of cancrinite formation was not evident. One hypothesis

was that an excess of silicate species may needed to promote cancrinite growth. This suggestion was

consistent with the well-known synthesis of zeolite X which typically is made using a SiO 2/Al2O 3

ratio in the synthesis gel of 2.9 to 3.8 (whereas the product zeolite had a SiO2/Al2O3 ratio of

approximately 2.5) (Ahmadon et al., 2018; X. Zhang et al., 2013).

Transformation of Cancrinite to Zeolite LTA


Given that carbonate-cancrinite (Na9 Al6 Si6 O24 (CO3 )1.5 . 2H2 O) has a SiO2/Al2O3 ratio of 2 it was

pertinent to examine if acid dissolution followed by neutralization and hydrothermal synthesis

produced zeolite LTA (Figure 3). Table 3 shows that only minor amounts of zeolite LTA were

present in the solid product along with the non-diffracting major product. Use of higher acid

concentrations appeared to be detrimental to zeolite LTA formation.

Table 3. XRD data for zeolite LTA synthesis from cancrinite: four hours reaction time; agitation 100
rpm; reaction temperature 80 ℃; dissolution of cancrinite with either 2 or 3 M HCl followed by
neutralization.
Phase (wt%) Dissolved at 2M HCl Dissolved at 3M HCl
Quartz 0.1 0.1
Cancrinite 1.8 1.2
Sodalite 0.0 0.0
Zeolite LTA 14.9 4.9
Non-diffracting 83.3 93.8

SEM images corroborated XRD data as the presence of cubic zeolite LTA crystals was not prevalent

(Figure 9).

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 53
`
2 M HCl 3 M HCl
Figure 9. SEM images of solid material made from cancrinite: four hours reaction time; agitation
100 rpm; reaction temperature 80 ℃; dissolution of cancrinite with either 2 or 3 M HCl followed by
neutralization.

Addition of NaAlO2 to the Cancrinite Product


Based upon the analysis above one idea to promote the formation of zeolite LTA was the introduction

of excess aluminium species. This practice is used in the zeolite industry wherein the most common

SiO2/Al2O3 ratio appears to be 1.8 ± 0.2 ( Kettinger et al., 1979). In theory, the presence of excess

aluminium species minimizes the formation of by-products.

Acid Dissolution of Cancrinite to Make Active Al and Si species.

As seen above in “Transformation of Cancrinite to Zeolite LTA” the molarity of HCl employed

to create active aluminium and silica species from cancrinite was important. Therefore, the influence

of hydrochloric acid concentration was investigated by contacting cancrinite (1 g) and additional

sodium aluminate (initially set to SiO 2/Al2O3 ratio = 1.14) with 10 mL of hydrochloric acid (HCl) at

different molarities. Figure 10 shows the theoretical maximum amount of Si and Al in solution if the

cancrinite was totally dissolved. Substantial dissolution of cancrinite occurred when using 2 to 4 M

HCl (maximum value of 97 and 89 % aluminium and silica active species). Under the highly acidic

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 54
`
conditions, the active species in solution are reported to be (Al(H2O)6)3+ and H4SiO4 (J.-Q. Wang

et al., 2016). The scope to reduce the HCL concentration was not possible as formation of active

species was limited at a HCl molarity of 1 M. Notably, J.-Q. Wang et al. (2016) reported that

hydroxycancrinite made from kaolin became XRD amorphous after using 0.2 M HCl in a timeframe

of 30 minutes. This observation may indicate that another factor which is potentially important is the

solid/liquid ratio when adding acid to an aluminosilicate material (Jiang et al., 2012).

100 100

Si Active species (%)


Al Active species (%)

90 90

80 80

70 70

60 60

50 50
0 1 2 3 4 5 0 1 2 3 4 5

HCl Molarity (M) HCl Molarity (M)

Figure 10: Influence of HCl molarity on Al and Si active species formation from cancrinite/sodium
aluminate

Although the acid dissolution stage created active silica and aluminium species, it was apparent that

the addition of HCl also made NaCl. As discussed in the following sections, this salt production may

represent a significant financial and environmental problem when making zeolites from SLR.

Neutralization of Acidic Solution of Active Al and Si Species Followed by Zeolite Synthesis

To initiate zeolite formation, an acid neutralization stage was completed to promote homogeneous

gel formation. NaOH (10 mL per gram of cancrinite) at the desired molarity was added into the acid

solution whereupon an aluminosilicate gel was apparent. This gel was then reacted under agitation

conditions (100 rpm) at 80 ℃ for various synthesis times to form zeolite LTA.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 55
`
According to Krongkrachang et al., (2019) the presence of chloride ions in the reaction interfered

with zeolite crystallization. Therefore, only dissolution of cancrinite/NaAlO2 with 2 and 3 M HCl

was studied in relation to the impact of subsequent neutralization and ultimately, zeolite LTA

formation. Table 4 reveals a similar trend in zeolite LTA quality (i.e., reduction in zeolite LTA

amount when higher molarity acid was employed in the cancrinite dissolution stage) to the situation

where no additional NaAlO2 was added to the cancrinite. Notably the highest amount of crystalline

zeolite LTA was 63.1 wt% using the 2 M HCl solution, and only 35.7 wt% when 3 M HCl was

employed in the activation step.

Table 4. Quantitative XRD analysis of zeolite LTA: Impact of cancrinite/sodium aluminate acid
treatment conditions
wt%
Sample names
Quartz Cancrinite Zeolite LTA Non-Diffracting

Cancrinite dissolved in 2M HCl 0.2 4.7 63.1 32.0

Cancrinite dissolved in 3M HCl 0.3 6.1 35.7 58.0

Further information was acquired by application of Scanning Electron Microscopy to image the

zeolite products outlined in Table 4 (Figure 11). Dissolution of cancrinite with 2 M HCl followed by

neutralization using 4M NaOH and hydrothermal synthesis at 80 oC, clearly produced cubic zeolite

LTA crystals (approximately 1 to 1.5 micron in size). In contrast, use of 3 M HCl inhibited zeolite

LTA formation (Figure 11). This observation agrees with Garcia et al., (2018) who noted that the

presence of NaCl suppressed zeolite P formation.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 56
`
a) Cancrinite dissolved in 2M HCl b) Cancrinite dissolved in 3M HCl
Figure 11.SEM images of zeolite LTA: Influence of different HCl concentrations on zeolite LTA
synthesis at 4 h, 80 ℃ and neutralized with 4 M NaOH solution: Cancrinite/NaAlO 2 dissolved in (a)
2 M HCl and (b) 3M HCl.

From the results shown in Table 4 it was pertinent to investigate the influence of sodium hydroxide

molarity upon zeolite LTA growth, as this factor may be important in controlling zeolite formation.

Tests were conducted wherein the cancrinite was activated using a 2 M HCl solution, then neutralized

by varying NaOH molarities, and finally hydrothermally reacted at 80 oC for four hours (Figure 12).

100 100
wt% Crystalline Zeolite LTA

wt% Non-Diffracting

80 80

60 60

40 40

20 20

0
0
2 3 4 5 6 7 8
2 3 4 5 6 7 8
NaOH Molarity (M)
NaOH Molarity (M)

Figure 12: XRD determination of % of crystalline zeolite LTA in solid product as a function of
NaOH molarity used in the neutralization stage: reaction temperature 80 oC; reaction time 4 h;
Si/Al = 0.57

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 57
`
Figure 12 shows that zeolite LTA synthesis rapidly intensified when raising the NaOH molarity

above 2 M; and peaked when using 4M NaOH (63 ± 5%). The solid products formed at higher NaOH

molarities plateaued in terms of zeolite LTA content. As zeolite LTA was formed, the amount of

amorphous material detected reduced in abundance. Some cancrinite was also present in amounts

ranging from 3 to 8 wt %. Due to the relatively small zeolite LTA size quantitative XRD analysis

may have been comprised by line broadening and/or a zeolite with significant amount of defects in

the structure (Raić et al., 2020; Strachowski et al., 2022). Further insight into the impact of sodium

hydroxide concentration is shown in Figure 13. Use of 2 M NaOH evidently did not promote zeolite

crystallization as the material appeared highly amorphous. Whereas the 3 M NaOH solution

promoted crystallization of cubic zeolite LTA, although the presence of less developed cubes was

evident. When neutralizing the acid activated cancrinite with either 4 or 5 M NaOH the formation of

cubic zeolite LTA was widespread.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 58
`
Neutralized in 2M NaOH Neutralized in 3M NaOH

Neutralized in 4M NaOH Neutralized in 5M NaOH

Figure 13. Influence of NaOH molarity in the neutralization stage prior to the formation of zeolite
LTA.

Impact of SiO 2/Al2O3 Molar Ratio upon Zeolite LTA Synthesis


As indicated previously, when making zeolite LTA commercially the SiO2/Al2O3 ratio is often less

that the value of 2 (1.8 ± 0.2) ( Kettinger et al., 1979). The silica/alumina ratio significantly influences

the crystallinity and properties of zeolite LTA. Higher silica/alumina ratios yield more ordered and

crystalline structures, impacting ion exchange capacity, catalytic activity, thermal stability, and

suitability for specific applications (Derouane, 1992).

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 59
`
In a previous section (Addition of NaAlO2 to cancrinite product) the SiO2/Al2O3 ratio was 1.14, thus

it was of interest to study higher values of SiO2/Al2O3 (up to a value of 2). Figure 14 showed that

increasing the SiO2/Al2O3 ratio decreased the amount of crystalline zeolite LTA. Simultaneously the

presence of amorphous material correspondingly increased.

The cancrinite impurity phase was favoured at lower values of SiO 2/Al2O3. Ameh et al., (2017)

examined the impact of sodium aluminate addition to fused coal fly ash in relation to subsequent

hydrothermal synthesis of zeolite LTA. It was found that the crystallinity and purity of zeolite LTA

was promoted when the SiO2/Al2O3 ratio was deceased from 2.14 to 1.02. Zhang et al. (2013) also

observed the formation of zeolite X from hydrogels at SiO2/Al2O3 = 3.5 to 1.5 and zeolite LTA at

SiO2/Al2O3 = 1.0 to 0.5. Mechanistically it was postulated that the low silica concentrations

constrained the building units to be double four membered rings which favoured zeolite LTA growth.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 60
`
100
Dissolved in 2M HCl

wt% Crystalline Zeolite LTA


Dissolved in 3M HCl
80

60

40

20

0
1.2 1.4 1.6 1.8 2.0
SiO2/Al2O3 ratio

100 100
Dissolved in 2M HCl Dissolved in 2M HCl
wt% Crystalline Cancrinite

Dissolved in 3M HCl wt% non-diffracting Dissolved in 3M HCl


80 80

60 60

40 40

20
20

0
0
1.2 1.4 1.6 1.8 2.0
1.2 1.4 1.6 1.8 2.0
SiO2/Al2O3 ratio
SiO2/Al2O3 ratio

Figure 14. Influence of different HCl concentrations on Zeolite LTA synthesis using different Si/Al
ratios, synthesis at 4 h, 80℃ and neutralized with 4M NaOH solution.

The Influence of Hydrothermal Synthesis Time upon Zeolite LTA Formation


Ideally, the synthesis time for zeolite LTA should be relatively short in order to lower production

costs. Figure 15 illustrates the relationship between the amount of crystalline zeolite LTA detected

by XRD analysis as a function of reaction time. After two hours 44.2 % of the solid product was

estimated to be crystalline zeolite LTA. Increasing the reaction to four hours resulted in greater

amounts of zeolite LTA (63.1 wt%) formed; albeit extending the time further did not reveal any

process benefits. One negative aspect was the identification of cancrinite (Table 5) in the range 5 to

8 wt%.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 61
`
100
Zeolite LTA
Non-diffracting
80
wt% Crystalline

60

40

20

0
2 3 4 5 6
Time (hrs)
Figure 15. Influence of synthesis time on zeolite LTA synthesis: reaction temperature 80℃; 0.57
Si/Al ratio; and 100 rpm agitation

Supplementary Figure 5 shows that after two hours reaction there was a mixture of phases present

and that extending the reaction time to four hours correspondingly increased the presence of cubic

zeolite LTA crystals. Thus, there was a good correlation between the XRD data (

Table 5) and SEM images (Supplementary Figure 5).

Table 5. XRD analysis of zeolite LTA activated at 2M HCl and synthesised at a reaction
temperature 80℃; 0.57 Si/Al ratio; and 100 rpm agitation.
Wt%
Zeolite Non-
Synthesis time (h) Quartz Cancrinite Sodalite
LTA Diffracting
2 0.1 7.1 0.0 44.2 48.7
3 0.0 6.8 0.0 55.8 37.4
4 0.2 4.7 0.0 63.1 32.0
5 0.1 7.9 0.0 54.4 37.6
6 0.1 8.1 0.0 57.1 34.7

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 62
`
Mass Balance of the Activation of Cancrinite, Neutralization and Zeolite Synthesis
The question thus arising was the fate of the excess aluminium species added. Due to Lowenstein’s

rule, it is not possible to have Al-O-Al pairs in a zeolite lattice which means that the lowest

SiO2/Al2O3 ratio in the zeolite LTA framework is two. Consequently, the mass balance of the

conversion of cancrinite/sodium aluminate to zeolite LTA was examined (Figure 16). In the acid

dissolution stage, it was evident that not only was the creation of active aluminium and silica species

high (98.4 & 94.4 %, respectively) but also substantial amounts of chloride ions in solution were

present (1.41 g). In relation to the synthesis parameters the molar ratios of the reactants were:

SiO2/Al2O3 = 1.07; Na2O/SiO2 = 4.99; H2O/Na2O = 35.94. These feed ratios were calculated after

correction of sodium ions associated with the chloride ions in solution. Hydrothermal synthesis at 80

oC for four hours produced a wet cake which was 46.5 wt% solids. After drying the solid material,

analysis indicated that the SiO2/Al 2O3 ratio of the zeolite LTA was 1.92. This value was in excellent

agreement with the theoretical formula for zeolite LTA of Na12((AlO2)12(SiO2)12)·27H2O and

supported by the data shown in Figure 13 and Figure 14.

The mother liquor contained minimal silica species (0.005 g) which was in accord with the excess of

aluminium employed in the zeolite synthesis stage to ensure maximum reaction of silica species. In

contrast a significant amount of dissolved aluminium species was present (0.21 g), as was sodium

ions (1.854 g), and chloride ions (1.02 g). Subtraction of sodium ions associated with dissolved

chloride revealed that 1.191 g of sodium ions were probably present as sodium hydroxide. Notably

the first washing stage of the wet cake contained a significant amount of dissolved aluminium,

sodium, and chloride ions. Table 6 illustrates the overall mass balance for the cancrinite to zeolite

LTA route. Within experimental error all elements were satisfactorily accounted for.

Table 6: Mass balance for zeolite LTA formation by conversion of cancrinite

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 63
`
Mass (g)

Feed Mix Dry Solid Mother Liquor Wash Water Balance

Si 0.34 0.33 0 0.01 0

Al 0.61 0.33 0.21 0.08 -0.01

Na 3.69 0.14 1.85 1.7 0

Cl 1.41 0 1.02 0.38 0.01

A final comment is made in terms of the practical aspects of this new route for transforming SLR to

a valuable zeolite (Figure 16). Usually, recycling of mother liquor from the hydrothermal stage is

necessary to recover expensive sodium hydroxide (Pastorello & Troglia, 1987). However, the use of

acid to dissolve the cancrinite inherently means that sodium chloride is made. If, in the process

described in this study, the mother liquor was recycled, then salts may accumulate and ultimately

create a waste stream. The synthesis method is gaining complexity and thus may not be able to

compete with conventional zeolite LTA synthesis.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 64
`
Figure 16. Process flow diagram including stream tables for conversion of cancrinite to zeolite LTA

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 65
`
Conclusions

Zeolite LTA has been successfully synthesized from waste spodumene leachate residue (SLR)

produced during lithium mining from hard rock. Conversion of SLR initially to a cancrinite

intermediate was demonstrated and subsequent acidification, neutralization and hydrothermal

synthesis completed.

The first process involved creation of dual products, namely zeolite LTA and cancrinite. In this

instance, advantage was taken of the fact that the mother liquor contained a high level of active

silicate species. Addition of sodium aluminate to the mother liquor followed by hydrothermal

synthesis created high quality zeolite LTA. A critical issue was the value attributed to cancrinite

which may have use for heavy metal removal, biofuels production, or agriculture.

A second strategy was to add additional aluminium species to the feed mixture to promote cancrinite

formation by reaction with the excess silica species resultant from the fact that SLR had a SiO2/Al2O 3

= 4. However, it was challenging to produce high quality cancrinite by this method, and this was

attributed to a possible need to have an excess of silica species present to drive cancrinite formation.

A third strategy involved use of a cancrinite/sodium aluminate mixture which was then dissolved in

acid, neutralized with NaOH, and hydrothermally reacted. High quality zeolite LTA was formed by

using 2 M HCl and 4 M NaOH for the acid dissolution and neutralization steps. Addition of extra

sodium aluminate to lower the SiO2/Al2O3 to <1 improved zeolite LTA quality. Mass balancing

revealed that the excess aluminium species were in the mother liquor and wash water.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 66
`
The presence of active aluminium species and sodium hydroxide was beneficial if reused in the

synthesis process. However, the presence of sodium chloride was proposed to be problematic in

terms of interference with zeolite synthesis and enrichment in the mother liquor/wash water which

ultimately created a wastewater.

In conclusion, the cancrinite intermediate route for making zeolite LTA is technically feasible if

conditions are chosen carefully. Future research should address the feasibility of Process 1 in greater

detail, especially in terms of cancrinite application and process economics. Determination of a cost-

effective means to remove/recover sodium chloride for Process 3 would also be valuable.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 67
`
Key Insights

This chapter demonstrated the successful synthesis of zeolite LTA from waste spodumene leachate

residue (SLR) through a multi-step process involving the creation of a cancrinite intermediate.

Following the activation process, two products were obtained, the mother liquor and cancrinite. The

addition of sodium aluminate to the mother liquor, followed by hydrothermal synthesis, produced

high-quality zeolite LTA. This is a noteworthy achievement, as a high-purity material can be

produced through this route, with all impurities remaining in the solid product (cancrinite), while the

mother liquor consists mainly of pure silicon species.

However, the production of cancrinite proved to be challenging, requiring additional aluminium

species for its formation. Hydrochloric acid was utilized to dissolve the cancrinite and obtain active

aluminium and silica species, followed by neutralization with NaOH and hydrothermal reaction to

form zeolite LTA. Nonetheless, the presence of sodium chloride was problematic, interfering with

zeolite synthesis and increasing the complexity of the process to recycle mother liquor and washing

water, ultimately leading to wastewater generation. Thus, further research is necessary to identify

alternative methods for cancrinite utilization that do not necessitate the presence of hydrochloric acid.

The next chapter focuses on the improvement of this method. In order to further optimize and improve

the synthesis of zeolite LTA from waste spodumene leachate residue, several areas of research can

be explored. One potential avenue is the investigation of high silica waste materials as a means to

improve the efficiency of the method. The high levels of active silicate species present in the mother

liquor make waste materials with high silica content an attractive option, as they could lead to the

formation of a high purity solution while reducing the production of cancrinite.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 68
`
Another promising area for future research is the exploration of alternative methods for cancrinite

use that do not require the presence of hydrochloric acid. A possible solution would be to investigate

the use of cancrinite as a raw material for other processes, such as phosphate absorption. This would

provide a sustainable solution for the disposal of waste materials, while also reducing the overall

complexity of the process and minimizing the generation of wastewater. These potential

improvements could have a significant impact on the commercial viability of the process and

contribute to a more sustainable approach to waste management.

Chapter 3: Feasibility of Zeolite LTA Synthesis from Spodumene Leachate Residue via Cancrinite Formation 69
`
Chapter 4: Transformation of natural zeolite powder to synthetic zeolite LTA

for water softening and lanthanum coated cancrinite for phosphate removal.

Abstract:

Mining of natural zeolite results in the formation of a powdered waste material which has limited

commercial value due to a combination of processing challenges and low purity. Therefore, a new

approach was taken to transform the powder into synthetic zeolite LTA and lanthanum supported on

cancrinite. Cancrinite was initially made by hydrothermal synthesis of zeolite with sodium hydroxide

solution at 240 oC. Importantly, use of 6 M NaOH solutions was necessary to reduce the amount

cancrinite present and promote the presence of silicates in solution. The resultant mother liquor was

composed of Si, Al, and Na while the bulk of the remaining impurities remained in the solid

cancrinite. Hydrothermal synthesis of the mother liquor with added sodium aluminate to adjust the

Si/Al ratio produced zeolite LTA crystals (95.5 wt%). The reaction time was preferred to be > three

hours under the applied conditions. Value was added to the cancrinite product by coating with

lanthanum chloride to make 25 g of lanthanum per kg of cancrinite. Subsequently phosphate species

were removed from solution (99 % reduction). Future research should focus on the development of

improved lanthanum doped cancrinite formulations and expand the testing to include solutions with

phosphate and multiple competing anions.

Key Words: zeolite, waste material, hydrothermal activation, mother liquor

Chapter 4: Transformation of natural zeolite powder to synthetic zeolite LTA for water softening and lanthanum coated
cancrinite for phosphate removal. 70
`
Introduction

Natural zeolites are widely used for applications such as molecular sieving (Eroglu et al., 2017),

catalysis (Hwang et al., 2002), cation exchange and sorption (Smith, 1985; S. Wang & Peng, 2010),

wastewater treatment (Rodríguez-Iznaga et al., 2018), and wound healing (Y. Li et al., 2012). In

terms of processing the natural zeolite deposits, the stages involved: (1) blasting to free the rocks;

(2) crushing; (3) fractionation; (4) washing; and (5) packaging (Yusupov et al., 2000). Consumer

demand is typically for crushed rocks in the size range 1 to 4 mm. Notably, there is limited industry

requirement for finer zeolite particles produced from the crushing step with most investigations of

natural zeolite conducted in the size range 0.5 to 3 mm (Zijun et al., 2021). Thus, a waste material is

generated with a particle size smaller than 100 µm. This powder not only represents a financial loss

in the process but also is a health hazard on site due the susceptibility of the small particles to be

inhaled (Hnizdo & Vallyathan, 2003; Steenland & Ward, 2014; Zijun et al., 2021).

One approach to the natural zeolite powder waste issue is the conversion of natural zeolite to value

added zeolites. For example, Zijun et al. (2021) used natural zeolite powder to synthetize zeolite

W. A high temperature alkali fusion stage was initially required to activate the natural zeolite powder.

Subsequently, aluminium hydroxide, potassium hydroxide and water were added to the activated

solid material and hydrothermal synthesis completed at 150 oC for 16 h. The material zeolite W

product exhibited a high capacity for ammonium exchange from water (324 meq/100 g). Despite the

success in making zeolite W, this study was probably not economic due to the costly fusion process.

Wruck et al. (2021) studied the influence of natural zeolite deposit upon activation of natural zeolite

and resultant synthesis of Zeolite LTA. Although the natural zeolite samples exhibited different

thermal stability, the quality of the final zeolite LTA product made via an alkali fusion process was

not significantly impacted (ca. 85 wt% zeolite LTA).

Chapter 4: Transformation of natural zeolite powder to synthetic zeolite LTA for water softening and lanthanum coated
cancrinite for phosphate removal. 71
`
Avoidance of the need for high temperature alkali fusion has been described by J.-Q. Wang et al.

(2016) who transformed kaolin and quartz into zeolite Y using hydrothermal alkaline activation

(HAA). The procedure was to heat a mixture of kaolin, quartz, and high molarity sodium hydroxide

solution at 240 oC to form hydroxycancrinite. The hydroxycancrinite was unstable in the acid

solution and as such degraded to form active aluminium and silicate species when contacted with

HCl. Finally, sodium hydroxide was added to form a gel which was then reacted at 90 oC under

appropriate conditions to make zeolite Y. One claimed advantage of this hydroxycancrinite

intermediate process was that the purity of the zeolite Y was improved. A similar approach was also

reported by the same authors in a study regarding transformation of kaolin to zeolite LTA (J.-Q.

Wang et al., 2014). The conventional high temperature activation of kaolin to metakaolin was

replaced by hydrothermal alkaline activation and then the cancrinite phase was acid dissolved,

neutralised, and reacted to make zeolite LTA. However, there was no discussion of the potential

problem relating to the formation of sodium chloride during the NaOH neutralization step.

X. Wang et al. (2021) described the need to value add aluminosilicate-based lithium slag produced

during the extraction of lithium from hard rock ore bodies. Notably, for every tonne of lithium

carbonate made, eight to ten tonnes of waste were created. These authors suggested a strategy

wherein lithium slag was hydrothermally treated at temperatures in the range 220 to 240 oC to produce

a solid hydroxycancrinite and a silicate rich mother liquor. The hydroxycancrinite was subsequently

reacted at elevated temperatures with nitric acid solution to make kaolinite. In addition, the mother

liquor was reacted at 260 oC with calcium oxide to form xonotlite (6CaO⋅6SiO2⋅H2O). The question

thus arose whether the outlined approach for recovering valuable materials from lithium slag

treatment could be transferred and improved for upgrading of natural zeolite waste powder.

Chapter 4: Transformation of natural zeolite powder to synthetic zeolite LTA for water softening and lanthanum coated
cancrinite for phosphate removal. 72
`
In terms of material composition, the lithium slag was characterized by a SiO2/Al2O3 ratio = 4.26 (X.

Wang et al., 2021), while natural zeolite has a SiO 2/Al2O3 ratio up to 9.88 (Wruck et al., 2021).

Moreover, natural zeolites invariably comprise of a range of phases such as mordenite, quartz,

clinoptilolite, stilbite, sanidine, and heulandite, and impurities such as iron, calcium, potassium, and

magnesium (Wruck et al., 2021; Zijun et al., 2021).

To avoid the formation of sodium chloride contamination in the mother liquor it was postulated that

instead of acid dissolving the cancrinite, the cancrinite would be better used as a support for an active

coating. A major environmental problem is the presence of excess amounts of phosphates in water

and wastewater (Hinesly & Jones, 1990; Kassem et al., 2022). Instead of depleting phosphate rock

deposits, the removal and recovery of phosphate species from wastewater also offers the advantage

of sustainability. (T. Zhang et al., 2010). Commercially lanthanum coated bentonite has been

employed to control the phosphate species in lakes (Y. Han et al., 2022; Waajen et al., 2016).

Notably, lanthanum has a strong affinity for phosphate species (S. Dong et al., 2017). There are

successful studies showing the efficiency of the coated materials with lanthanum to remove

phosphates, for instance J. He et al. (2015) coated nanorods with lanthanum as an antimicrobia l

solution for drinking water security. Whereas,. L. Zhang et al. (2011), coated activated carbon fibres

with lanthanum oxide and obtained 97.6 % phosphate removal.

Therefore, the aim of this study was to activate natural zeolite powder at relatively low temperatures

to form cancrinite and highly alkaline, silica rich, mother liquor. Value adding of cancrinite by

coating with lanthanum species was targeted for phosphate removal from wastewater. In turn, zeolite

LTA synthesis from mother liquor made use of excess silicates and alkali such as NaOH plus the

addition of sodium aluminate. The hypothesis was that if natural zeolite powder can be activated to

Chapter 4: Transformation of natural zeolite powder to synthetic zeolite LTA for water softening and lanthanum coated
cancrinite for phosphate removal. 73
`
form cancrinite and mother liquor then two valuable products can be made for agricultural purposes

and the synthetic zeolite market. Research questions addressed included: (1) What are the preferred

conditions for natural zeolite activation? (2) How to minimize the amount of cancrinite present and

maximise zeolite LTA? (3) What is the composition of the mother liquor and is the presence of

impurities reduced? (4) Does the reaction time significantly influence zeolite LTA formation? and

(5) What is the performance of lanthanum doped cancrinite for phosphate removal from aqueous

solutions? The basic methodology involved bench scale hydrothermal synthesis of cancrinite from

natural zeolite powder.

Chapter 4: Transformation of natural zeolite powder to synthetic zeolite LTA for water softening and lanthanum coated
cancrinite for phosphate removal. 74
`
Materials and Methods

Natural Zeolite and Chemicals


Natural zeolite was supplied by Zeolite Australia with a particle size of 7.16 μm. The elemental

composition as inferred by X-ray Fluorescence indicated that the Si/Al ratio was greater than 4.

Table 7: XRF analysis of natural zeolite powder


Mass % (Dry Basis)
Oxide Natural Zeolite Oxide Natural Zeolite
Na2O 1.49 Mn2O3 0.12
Al2O3 12.54 Fe2O3 1.98
SiO2 71.08 NiO 0.01
MgO 0.58 SrO 0.03
P 2O 5 0.06 ZrO2 0.04
SO3 0.04 BaO 0.04
K2O 1.75 HfO2 0.01
CaO 3.32 PbO 0.05
TiO2 0.21 Sum 99.89

Quantitative X-ray diffraction (XRD) analysis indicated the presence of 14.0 wt% quartz, 1.9 wt%

cancrinite, 11.3 wt% sanidine, 41.1 wt% heulandite, 11.0 wt% mordenite, and 20.7 wt% non-

diffracting/amorphous material. Sodium aluminate (NaAlO 2) and lanthanum chloride hexahydrate

(LaCl3.6H2O) were purchased from Sigma-Aldrich and sodium hydroxide (NaOH; extra pure micro-

pearls) was procured as an analytical reagent (AR) grade from ChemSupply.

Chapter 4: Transformation of natural zeolite powder to synthetic zeolite LTA for water softening and lanthanum coated
cancrinite for phosphate removal. 75
`
Transformation of Natural Zeolites to Synthetic Zeolites
Natural Zeolite Activation
In the activation stage, 600 g of natural zeolite powder was initially mixed with NaOH (1.917 g

NaOH/g natural zeolite) and deionized water to control sodium hydroxide molarity. The mixture was

then placed in a PTFE lined 1000 mL acid digestion vessel (Amar Equipment, India) at 240 ℃ for

one hour. After the activation period two products were separated from using an Allegra X-30

centrifuge: solid product (hydroxycancrinite) and supernatant (mother liquor rich in active silicate

species and sodium hydroxide).

Zeolite LTA Synthesis from Mother Liquor


A SiO2/Al2O3 ratio = 2 was achieved by adding extra NaAlO 2 solution along with sufficient water to

adjust the sodium hydroxide molarity to 4 M. This mixture produced aluminosilicate gels that were

placed in sealed containers in a New Brunswick Innova 42R Incubator at 80 ℃ for four hours and

100 rpm. When the zeolite synthesis was completed, the final product was centrifuged to recover the

zeolite slurry. Then, the “wet cake” was washed with deionized water to reduce pH from 14 to 11.

(30 mL of deionized water was used every cycle). After the preferred pH was reached, the zeolite

slurry was weighed, and dried in an oven at 105 ℃ overnight.

Cancrinite Coating with Lanthanum


Cancrinite (1 g) was accurately weighed into a 50 mL centrifuge tube and mixed with 4 mL of a

lanthanum chloride (LaCl3.6H2O) solution which had a molarity of 1, 2, and 3 M. The solution was

agitated for 24 hours on a rotary stirrer wheel. The mixture was then centrifuged, and the supernatant

collected, solid washed with deionised water, and dried in an oven at 105 °C overnight. The overall

transformation of natural zeolite to zeolite LTA/lanthanum loaded cancrinite is shown in Figure 17.

Chapter 4: Transformation of natural zeolite powder to synthetic zeolite LTA for water softening and lanthanum coated
cancrinite for phosphate removal. 76
`
Figure 17: Concept for transformation of natural zeolite powder to synthetic zeolite LTA and
lanthanum coated cancrinite.

Phosphate Sorption Isotherms on Modified Cancrinite


Samples of modified cancrinite with masses between 0.1 and 2.5 g were added to twelve 250 mL

Nalgene bottles along with a 200 mL aqueous solution of anhydrous sodium phosphate dibasic

(Na2HPO4) at 100 mg/L phosphate concentration. The samples were then placed in a Innova 42R,

New Brunswick Scientific shaking incubator at 25℃ and agitated at 150 RPM. The samples were

stirred for 24 hours to guarantee that equilibrium was obtained. The samples were filtered through a

syringe filter and diluted to 1:10 with Milli-Q water to analyse the solution. All the samples were

completed in duplicates to ensure the accuracy of the data.

The equilibrium isotherms for phosphate sorption on lanthanum modified cancrinite were modelled

using the Langmuir, Freundlich, and Aranovich-Donohue models.

Chapter 4: Transformation of natural zeolite powder to synthetic zeolite LTA for water softening and lanthanum coated
cancrinite for phosphate removal. 77
`
The Langmuir equation, where: Csolution is the dissolved K concentration in soil solution (mgKL -

1), Csolid is the soil sorbed K concentration (mgKkg-1), km is the K adsorption maximum (mgKkg-

1) and kα is the affinity constant (LmgK -1) (C. Wang et al., 2017)

∁ 𝒔𝒐𝒍𝒖𝒕𝒊𝒐𝒏 𝟏 ∁ 𝒔𝒐𝒍𝒖𝒕𝒊𝒐𝒏
= +
∁ 𝒔𝒐𝒍𝒊𝒅 𝐤𝐦 ∗ 𝐤𝛂 𝒌𝒎
Equation 7. Langmuir equation

The Freundlich equation, where kf is the Freundlich constant (mgKkg-1) and C solid and C solution

were defined earlier (C. Wang et al., 2017)

𝑪𝒔𝒐𝒍𝒊𝒅 = 𝒌𝒇 ∗ 𝑪 𝒔𝒐𝒍𝒖𝒕𝒊𝒐𝒏𝟏/𝒏
Equation 8. Freundlich equation

The Aranovich-Donohue equation, where q is the adsorbed concentration, c is the aqueous

concentration, km and kα are the Langimuir constants and Ka and nA are fitting parameters (Chu &

Hashim, 2023)

𝒌𝒎 ∗ 𝒌𝜶
𝒒=
(𝟏 + 𝒌𝜶)(𝟏 − 𝒌𝑨𝒄) 𝒏𝑨
Equation 9. Aranovich-Donohue equation

Solution Analysis

Inductively Coupled Plasma - Optical Emission Spectroscopy (ICP-OES)


A Perkin Elmer Optima 8300 DV ICP-OES spectrometer was used to determine the composition of

dissolved ions in the mother liquor (especially Si, Al, and Na). Before the analysis, the samples were

Chapter 4: Transformation of natural zeolite powder to synthetic zeolite LTA for water s oftening and lanthanum coated
cancrinite for phosphate removal. 78
`
filtered through a syringe filter and diluted to 1:100 and 1:1000 with 2% nitric acid (diluent) using a

Hamilton auto-diluter.

Ion Chromatography (IC)


A Dionex ICS-2100 integrated ion chromatography system was used to complete phosphate

quantification. The system was equipped with a conductivity detector and a potassium hydroxide

eluent generator. The conditions for the suppressed ion chromatographic method were eluent

strength of 28.00 mM KOH, flow rate of 1.00 mL/min, injection volume of 25.00 ml, column

temperature of 30 °C, cell temperature of 35 °C, elution time of 10 minutes and operating pressure

of 2000 psi. Multi-element standards were prepared in-house to calibrate this method. Samples were

diluted with Milli-Q water and syringe filtered prior to analysis.

Solid Characterization

Quantitative X-ray Diffraction (XRD)


Powder X-ray diffraction patterns were collected using a PANalytical X’PertPro powder

diffractometer. The internal standard used to identify the crystalline phases in the solid products

consisted of corundum. Typically, the zeolite samples were mixed with 10 wt% corundum standard

powder (Al2O3). The samples (1.8 g) and standard (0.2 g) were micronized for five minutes in a

McCrone micronizing mill with 15 mL ethanol; and dried in an oven at 40 ℃ for four hours. Dried

samples were back pressed into sample holders. To identify the material phases JEdit software and

TOPAS were used. The reference patterns for zeolite LTA were obtained from Breck et al. (Breck

et al., 1956) and atomic parameters were cross-referenced from PDF entry 04-016-9920.

Chapter 4: Transformation of natural zeolite powder to synthetic zeolite LTA for water softening and lanthanum coated
cancrinite for phosphate removal. 79
`
X-ray Fluorescence (XRF)
The PANalytical Wide Ranging Oxide Calibration (WROXI) method was used to determine the

elemental quantification of the solid products. Solid samples (1 g) were mixed with lithium

tetraborate: lithium metaborate (50:50) with 0.05 % li-Iodate (9 g) and heated at 1050 ℃ to form a

fused glass disc.

Scanning Electron Microscopy (SEM)


The morphology of the samples was observed using a JEOL 7001F Scanning Electron Microscope.

The accelerating voltage employed was 5 kV, with a probe current of 8 A at a magnification between

4,000 and 10,000 x. Each sample was prepared on carbon stamps and coated with five nm of platinum

prior to observation to reduce charging across the surface.

Chapter 4: Transformation of natural zeolite powder to synthetic zeolite LTA for water softening and lanthanum coated
cancrinite for phosphate removal. 80
`
Results and Discussion

Hydrothermal Activation of Natural Zeolite Powder


The first stage determined the impact of the sodium hydroxide molarity upon cancrinite

formation. Tests were conducted at a reaction temperature of 240 ℃ and one hour synthesis

time. The starting mixture was described by a SiO 2/Al2O3 ratio = 9.59, and Na 2O/SiO2 ratio

=1.33. The NaOH molarity ranged from 10 M (H 2O/Na2O=21.17), 8 M (H2O/Na2O=25.40), 6

M (H2O/Na2O=32.45), 5 M (H2O/Na2O=38.08) and 4 M (H2O/Na2O=46.54). The amount of

active Si species in solution was dependent upon the NaOH molarity. Alkaline solutions

increase the solubility of silicate minerals and these can form soluble complexes such as silicate

ions (SiO32-), leading to greater silicon dissolution. This phenomenon is a result of enhanced

mineral solubility and complex formation in alkaline conditions (D’Elia et al., 2018).

The highest amount of Si species (73 % of total Si present) in the mother liquor was found

when using 6 M NaOH Figure 2. Whereas higher molarities such as 10 M NaOH contained

substantially less active Si species (35 %). Y. Liu et al. (2021) also concluded that the preferred

molarity of the NaOH solution was 6 M when using microwave heating to convert potassic

rocks to cancrinite. However, in this example a reaction time of two hours at 200 oC was

recommended.

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated
cancrinite for phosphate removal
81
Figure 18. Active Si in the mother liquor (%) and solids recovery (g) at different NaOH
molarities

Quantitative XRD was used to elucidate the dependence of NaOH molarity upon creation of

active Si species in solution (Table 8).

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated
cancrinite for phosphate removal
82
Table 8. Solid materials obtained after natural zeolite powder activation at 240 oC and one hour synthesis.

Quantitative XRD Analysis (wt %)


NaOH
Molarity Solid (g) Quartz Cancrinite Sanidine Heulandite Mordenite Amorphous
(M)
10 41.3 10.26 10.26 6.58 0.00 0.10 72.81
8 30.8 12.63 7.95 9.66 2.59 0.00 66.04
6 16.7 0.46 71.10 1.03 0.53 0.06 26.82
5 27.1 0.45 74.17 0.30 1.21 1.54 22.34
4 14.6 0.43 51.49 0.00 1.30 2.08 44.70

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated cancrinite for phosphate removal 83
After the activation process it was apparent that the mass of solid collected was highest when

cancrinite formation was not favoured (8 and 10 M NaOH). Indeed, the smallest quantity of

solid was obtained when using 6 M NaOH (16.7 g). When using 8 and 10 M NaOH solutions

the solid was comprised of relatively small amounts of cancrinite (7.95 and 10.26 wt%,

respectively) and comparably large amounts of quartz (12.63 & 10.26, respectively), sanidine

(9.66 and 6.58 wt%, respectively), and non-diffracting material (66.04 and 72.81 wt%,

respectively).

Further analysis of the solid products shown in Table 8 was completed using SEM (Figure 19).

Activation of the natural zeolite powder with 6 M NaOH solution promoted growth of

cancrinite crystals as evidenced by the characteristic rod-like shape ( Buhl & Petrov, 2021). In

contrast, activation with 10 M solutions produced material which was more heterogeneous in

appearance, decorated with small amounts of cancrinite, and exhibited ill-defined elongated

plate like materials.

Activation with 6 M NaOH Activation with 10 M NaOH

Figure 19: SEM images of solids formed after natural zeolite activation using 6 M and 10 M
NaOH: reaction temperature = 240 oC; reaction time = 1 h; 10,000x magnification.

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated
cancrinite for phosphate removal
84
According to Hermeler et al. (1991) increasing the NaOH concentration produced an

intermediate phase besides cancrinite and sodalite. This intermediate phase was described as

cubic natrodavyne ( Buhl, 1991; Hermeler et al., 1991). The difference between cancrinite and

this intermediate compound is the crystal structure, while cancrinite exhibits a 3-dimensiona l

framework this new compound shows one-dimensional positional disorder. The results also

agreed with previous studies that transformed aluminosilicate waste materials (coal fly ash,

kaolin, and raw muscovite) to higher value products, where cancrinite was reported to form at

temperatures higher than 200 ℃ (Mahima, et al., 2020; Murukutti & Jena, 2022; Selim et al.,

2018; Wernert et al., 2020). Furthermore, Hackbarth et al. (1999), reported that to avoid the

formation of by-products (especially sodalite) during cancrinite synthesis the temperature

should not remain below 200 ℃.

Stability of Mother Liquor Solution


The previous section described how to activate natural zeolite to make in addition to solid

material a solution comprising of sodium hydroxide and dissolved silica species (mother

liquor). For practical purposes it is necessary to understand if the mother liquor is stable or

not. Depending upon the answer, either fresh mother liquor had to be made for each experiment

or larger volumes could be stored and used as required. Therefore, mother liquor was stored

in polypropylene bottles and preserved in a fridge for a period up to 25 days (Table 9).

Importantly, the solutions, whether at 4 or 5 M NaOH, displayed a remarkable degree of

stability with no evidence to indicate formation of precipitates. This observation was consistent

with high pH values between 13 and 14 which preserved the dissolved silicates in the mother

liquor (Krauskopf, 1956).

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated
cancrinite for phosphate removal
85
Table 9. Mother liquor composition as a function of storage time and NaOH molarity: natural
zeolite activation at 240 oC for 1 h.

4M NaOH 5 M NaOH
SAMPLE COLLECTION
Al (g) Si (g) Na (g) Al (g) Si (g) Na (g)
Activation (Day 0) 0.06 10.80 50.42 0.06 10.96 48.25
Day 3 0.06 11.12 50.36 0.06 11.27 47.89
Day 4 0.06 11.14 48.94 0.06 11.53 45.61
Day 11 0.06 11.00 47.76 0.06 11.51 44.95
Day 18 0.06 11.04 47.52 0.06 11.53 46.33
Day 25 0.06 10.90 46.78 0.06 11.27 44.76

Mass Balance for Synthesis of Zeolite LTA from Natural Zeolite Powder

Examination of the mass balance for the transformation of natural zeolite powder to zeolite

LTA was completed to understand the chemistry and process engineering involved (Figure 20).

Overall, a mass balance was recorded with: Si in and Si out 18.9 g; Al in and Al out 3.9 g; Na

in and Na out 66.8 g, and water in and water out 528 g. After the activation stage, it was evident

that the highest amount of dissolved Si species was in the mother liquor (13.90 g, which was

74 % of the total Si present). In contrast, only a minor amount of Al ions was present in the

mother liquor (0.09 g). The molar ratios for the activation parameters were: SiO2/Al2O3 = 9.31;

Na2O/SiO2 = 2.16; and H2O/Na2O = 20.19. Hydrothermal activation at 240 oC for one hour at

100 rpm produced 72.91 g of zeolite slurry. After drying the slurry, the final product was a

solid with 71.10 wt% cancrinite crystals. The first washing step of the zeolite slurry contained

1.84 g Si and 1.64 g Al (which can be reused in the next activation batch).

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated
cancrinite for phosphate removal
86
Zeolite LTA synthesis conditions were: SiO2/Al2O3 = 2.81; Na2O/SiO2 = 4.94; and H2O/Na2O

= 14.59. Reaction temperature was 80 oC and reaction time was four hours. Then the conditions

for zeolite LTA were adjusted by adding extra sodium aluminate (28.60 g) and water (282.50

g) to reduce NaOH molarity. After the synthesis the mass balance showed that the amount of

aluminium in the process was not enough to complete the zeolite synthesis since the excess of

Si in the supernatant was 6.93 g and 2.91 g Al.

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated
cancrinite for phosphate removal
87
Figure 20. Mass balance for activation of natural zeolite powder using 6 M NaOH at 240 oC for one hour followed by hydrothermal synthesis of

zeolite LTA at 80 oC for 4 h

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated cancrinite for phosphate removal 88
Table 10 shows the XRF analysis of the cancrinite after the hydrothermal activation at 240 oC and

one hour. The results showed that most natural zeolite impurities remained in the solid product

(cancrinite) after the activation stage.

Table 10: XRF analysis of solid product from activation of natural zeolite with 6 M NaOH at 240 oC
for one hour
Oxide Mass % (Dry Basis) Oxide Mass % (Dry Basis)
Na2O 19.88 Mn2O3 0.01
Al2O3 24.56 Fe2O3 0.00
SiO2 40.47 NiO 0.00
MgO 0.58 SrO 0.02
P 2O5 0.06 ZrO2 0.04
SO3 0.04 BaO 0.04
K2O 0.18 HfO2 0.01
CaO 3.30 PbO 0.03
TiO2 0.16 Sum 100

Zeolite LTA Synthesis from Mother Liquor

The synthesis recipe was SiO2/Al2O3 = 2.81; Na2O/SiO2 = 4.94; and H2O/Na2O = 14.59. The reaction

temperature was 80 oC and the sample was agitated at 150 rpm. Figure 21 indicates that the zeolite

LTA cubic crystals were not prevalent in the SEM images recorded for the sample which was reacted

for two hours. Indeed, the presence of a more disordered phase was noted. In contrast, extending

the reaction time to three hours promoted the formation of zeolite LTA as evidenced by the more

well-defined zeolite LTA material. The quantitative XRD data in Table 11 supports the conclusions

from inspection of the SEM images in Figure 21.

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated cancrinite
for phosphate removal 89
After two hours, less than half of the material was transformed into zeolite LTA (47.4 wt%).

However, increasing the reaction time to three hours resulted in almost double the amount of zeolite

shown at two hours (83.8 wt%). It was also noted that further increasing the time did not present any

extra benefits as the amount of material and the crystalline % stayed stable for up to five hours.

Synthesis time: 2h Synthesis time: 3h


Figure 21: Influence of synthesis time upon zeolite LTA synthesis: 80 oC, 150 rpm and
SiO2/Al2O3=2.81

Table 11. XRD determination of percentage of crystalline zeolite LTA in solid product as a function
of synthesis time: reaction temperature 80 ◦C; SiO 2/Al2O3 = 2
SYNTHESIS XRD ANALYSIS
Zeolite Zeolite
Time Dry Quartz Cancrinite Sodalite
LTA g /g LTA Amorphous%
(hrs) solid (g) wt% wt% wt%
NZ wt%
2 30.90 0.52 0.15 0.00 0.04 47.42 52.39
3 43.40 0.72 0.30 0.00 0.01 83.84 15.85
4 42.95 0.72 0.32 0.00 0.00 82.88 16.81
5 43.60 0.73 0.03 0.00 0.00 80.87 18.80

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated cancrinite
for phosphate removal 90
Table 12 shows the XRF analysis of the natural zeolite powder and how the impurities have moved

after the activation stage to the cancrinite material while the zeolite LTA synthetized from the mother

liqueur (after activation) was relatively pure: K 2O=0.28%; TiO2=0.05%, Mn2O3=0.05% ;

Fe2O3=0.22% and PbO=0.02%.

Table 12. XRF analysis of zeolite LTA made from natural zeolite powder.
Mass % (Dry Basis)
Oxide Zeolite LTA Oxide Zeolite LTA
Na2O 17.60 Mn2O3 0.05
Al2O3 28.55 Fe2O3 0.22
SiO2 34.19 NiO 0.00
MgO 0.00 SrO 0.00
P 2O 5 0.00 ZrO2 0.00
SO3 0.00 BaO 0.00
K2O 0.28 HfO2 0.00
CaO 0.00 PbO 0.02
TiO2 0.05 Sum 101.7

Influence of SiO 2/Al2O3 Molar Ratio upon Zeolite LTA Synthesis


Table 11 indicates that the wt% zeolite made was ≤83.8 wt%. Therefore, the challenge was to increase

the amount of zeolite LTA formed. Reducing the SiO2/Al2O3 ratio to slightly less than two should

drive the zeolite LTA formation due to the excess of aluminium present compared to stoichiometric

values (Rozhkovskaya et al., 2021). Figure 22: Zeolite LTA SiO2/Al2O3=1.90; Na2O/SiO2=5.11;

H2O/Na2O= 4.82, 4-hour synthesis and 80 oC represents the SEM image of a zeolite LTA made with

SiO2/Al2O3 = 1.90; Na2O/SiO2 = 5.11; H2O/Na2O = 14.82; four hours synthesis 80oC and 150 rpm.

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated cancrinite
for phosphate removal 91
Figure 22: Zeolite LTA SiO2/Al2O3=1.90; Na2O/SiO2=5.11; H2O/Na2O= 4.82, 4-hour synthesis and
80 oC

The SEM images suggested that an abundance of zeolite LTA was produced. This finding was

consistent with the XRD analysis which showed that the sample was now composed of 95.5 wt%

zeolite LTA, with only 0.28 wt% cancrinite, 0.04 wt% sodalite, 0.03 wt% quartz. and 4.11 wt%

amorphous content. For every 60 g of natural zeolite, 62.85 g of zeolite LTA was produced.

Phosphate Sorption on Cancrinite

Synthesis of Lanthanum Doped Cancrinite


The wet impregnation route was chosen to contact aqueous solutions of lanthanum chloride with the

cancrinite sample. The molarity of La 3+ ions was adjusted in the range 1 to 3 M at a constant solution

volume of 4 mL. The wt% of lanthanum on the cancrinite was determined to be 5.25 ± 0.15, 6.5 ±

0.3 and 11 ± 0.8 for initial lanthanum molarities of 1, 2 & 4, respectively.

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated cancrinite
for phosphate removal 92
Sorption Equilibrium Isotherms of Phosphate on Lanthanum Modified Cancrinite

The sorption capacity of cancrinite and cancrinite coated with lanthanum were investigated using

solutions comprised of 126 mg/L of phosphate species at 30 oC. When no coating was present on the

cancrinite material the uptake of phosphate ions from solution was zero. . Xiaotian et al., (2022)

reported that phosphate sorption on cancrinite present in red mud waste from bauxite refining is

relatively low. Indeed, increasing the cancrinite surface wettability was beneficial to enhancing

phosphate species uptake. In contrast, introduction of a lanthanum coating promoted phosphate

uptake (Figure 23). The most used isotherm models (Langmuir and Freundlich) did not satisfactorily

fit the recorded equilibrium data as the sum of errors squared (ERRSQ) value was 372 and 406,

respectively. A superior data fit was obtained when the Aranovich Donohue model was used

(ERRSQ = 74). This finding suggested that more than one sorption site for phosphate species on

lanthanum modified cancrinite is available (Millar, Couperthwaite et al., 2017). A recent study by

Y. Yang et al. (2021) supports this view as data was acquired which indicated that both monolayer

and multi-layer sorption processes occurred when using lanthanum doped ZSM-5 zeolite to

selectively recover phosphate from wastewater. For the Aranovich Donohue model the maximum

loading of phosphate species was predicted to be 19.8 mg/g of coated cancrinite, the “n” value was

0.12 and the Langmuir equilibrium coefficient (K L) was 3.23 L/mg.

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated cancrinite
for phosphate removal 93
Figure 23: Equilibrium isotherms for phosphate sorption on lanthanum modified cancrinite:
Langmuir (top), Freundlich (middle) and Aranovich-Donohue models (bottom)

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated cancrinite
for phosphate removal 94
The solutions after the equilibration process were also analysed for the presence of lanthanum ions

in the supernatant (Figure 24). For all the samples no lanthanum species were recorded which

suggested that the coatings were relatively stable. However, there was a linear trend in increasing

sodium ions in solution as a function of greater solid mass used in the isotherm tests. In addition,

smaller amounts of silicon species were steadily released until 1.5 g of cancrinite sorbent was added.

Higher solid mass content resulted in the gradual decrease in silicates emitted into solution.

Figure 24: Release of ions into solution as a function of from cancrinite mass during phosphate
equilibrium tests

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated cancrinite
for phosphate removal 95
Conclusions

This study successfully converted natural zeolite waste powder into two products which potentially

represent higher value; zeolite LTA and lanthanum loaded cancrinite. Activation of the natural

zeolite was achieved at relatively mild conditions. Additionally, the molarity of the NaOH used was

critical to produce the greatest amount of zeolite LTA. Employment of high pH solutions promoted

the deposition of precipitates in the cancrinite phase.

As the mother liquor from cancrinite formation was relatively pure, the quality of the zeolite LTA

produced from hydrothermal synthesis was increased. This discovery solved previous issues where

the removal of impurities was challenging, such as when acid washing is applied. The new approach

outlined also avoided problems in previous approaches wherein acid was used to dissolve cancrinite

as this resulted in a sodium chloride waste stream.

Cancrinite as a waste product causes problems of sustainability and economic viability. Thus, doping

of the cancrinite powder was chosen as it was useful for removing phosphate species from water and

wastewater resources. The initial performance of the cancrinite based sorbent was acceptable and

worthy of future studies to enhance the physical properties of the sorbent material. Expansion of the

zeolite manufacturing process to other high silica content waste resources is also suggested.

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated cancrinite
for phosphate removal 96
Key Insights

This chapter has successfully converted natural zeolite waste powder into two higher value products:

zeolite LTA and lanthanum-loaded cancrinite. The quality of the zeolite LTA produced through

hydrothermal synthesis was improved by obtaining relatively pure mother liquor from the activation

of the waste material. This discovery has addressed previous challenges by avoiding the use of

hydrochloric acid, which produced a non-sustainable and economically unviable sodium chloride

waste stream. Doping of the cancrinite powder was found to be useful in removing phosphate species

from water and wastewater resources. The initial performance of the cancrinite-based sorbent was

promising, and future studies aimed at enhancing the physical properties of the sorbent material are

suggested. The expansion of the zeolite manufacturing process to other high silica content waste

resources is also recommended. However, the efficiency of this method is limited to high silica

materials, and the process can be more complex than other methods, making it challenging to

implement on an industrial scale.

Therefore, the next chapter will focus on developing a simpler method to activate aluminosilicate

waste material that has the potential to be scaled up for industry. The next method needs to adjust the

reaction conditions to prevent the formation of cancrinite, potentially simplifying the process and

reducing the production of waste materials. Additionally, the new method has identified specific

optimal conditions for activation, including NaOH molarity, activation temperature, and quenching

time. This level of control has the potential to lead to more consistent and reproducible results, which

are essential for industrial-scale production.

Chapter 4: T ransformation of natural zeolite powder to synthetic zeolite LT A for water softening and lanthanum coated cancrinite
for phosphate removal 97
Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene

Leachate Residue using a New Quench Method

Abstract

Increasing demand for lithium has also increased the production of waste materials. The

aluminosilicate-based waste product from hard rock refining is generically termed “lithium

slag” or spodumene leachate residue (SLR) if spodumene was the mineral source. The

hypothesis was that if an economical approach was devised for the synthesis of zeolites from

waste or natural aluminosilicates then uptake of “green” zeolites by industry may be

accelerated. This study has successfully addressed the optimization of SLR activation and

zeolite LTA synthesis from the activated material using a novel one-pot process, named the

“Quench Method”. The activation of SLR was completed as a function of temperature (25 to

150◦ C), sodium hydroxide concentration (2 to 14 M), and time (10 to 1440 min) where 100◦ C,

10M NaOH and 60 minutes contact time were the preferred conditions. After activation, the

sodium hydroxide molarity and Si/Al ratio were adjusted for optimal zeolite synthesis

conditions. Si/Al ratio of 0.95 at 80◦ C and four hours synthesis resulted in 56 wt% Crystalline

zeolite LTA, although Si/Al ratio of 0.85 at 80◦ C and three hours resulted in a similar amount

of crystalline zeolite LTA (52 wt%) proving that the synthesis was improved by adding an

excess of aluminium in the solution. This study successfully validated the concept of using a

quench method to initially activate waste aluminosilicate materials (such as SLR) and then

adjust reaction conditions to form valuable zeolite products such as zeolite LTA. Future studies

should focus on the development of continuous hydrothermal reactors which can more rapidly

decrease the reaction temperature and molarity to improve control of the zeolite LTA synthesis.

Key Words: spodumene leachate residue; zeolite; activation; lithium; slag

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a Ne w Quench
Method 98
Introduction

Lithium and its compounds have been historically used in diverse application such as batteries,

lubricating greases, pharmaceuticals, glass, ceramics, and metallurgy (Choubey et al., 2016;

Collins et al., 2020; G. Han et al., 2018; Z. Liu, Zhu et al., 2019; P. Meshram et al., 2014; Sun

et al., 2014). For example, high-capacity lithium batteries are already prevalent in mobile

devices and are strong candidates for renewable energy storage ( Flexer et al., 2018).

Substantial growth of lithium battery use is expected as demand for electric vehicles increases

(Setoudeh et al., 2020). Lithium is mainly extracted from hard rock ores (lithium minerals such

as spodumene, petalite and lepidolite) (G. Han et al., 2018) and aqueous resources such as

lithium rich brines or salt lakes ( Flexer et al., 2018). At present, the majority of lithium

production around the world belongs to: six mineral operations in Australia, two brine

operations each in Argentina and Chile, and one brine and one mineral operation in China

(USGS, 2020). Indeed, Australia is currently the world’s largest producer of lithium ( Flexer

et al., 2018; Vikström et al., 2013). Spodumene is the favoured resource for processing into

lithium, as lepidolite and petalite typically contain a high fluoride content High fluoride content

in lepidolite and petalite increases operational costs, poses environmental challenges, and limits

market demand, affecting the profitability of lithium processing. (Vikström et al., 2013).

With increasing demand for lithium there is also increased production of waste materials. The

aluminosilicate based waste product from hard rock refining is generically termed “lithium

slag” or spodumene leachate residue (SLR) if spodumene was the mineral source (Zampori et

al., 2012). SLR is the solid waste formed after: (1) heat treatment at high temperatures to

transform α spodumene to the more active ß spodumene; and (2) leaching with sulfuric acid to

recover lithium ions.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 99
The general formula for SLR is HAlSi2O6 and as this material is commonly stored outdoors

there is a possibility of environmental damage due to its high alkalinity and potential for

leaching of heavy metals into the surrounding soil and water sources. It may also take up

valuable land space and contribute to visual pollution (D. Chen et al., 2012). Consequently,

SLR has been used as cement clinker in concrete, as a raw material for ceramic glazed tiles ,

and in activated clays (G. Han et al., 2018). However, the utilisation rate is only about 10 %

(Beushausen et al., 2012) which means there is a clear need to discover alternate uses for SLR.

Use of SLR to make synthetic zeolites is one possible avenue of research as this approach

makes use of the high aluminosilicate content of SLR. Zeolites are typically microporous or

mesoporous with the general composition: Mx/n (AlO2) x (SiO2) y ∙ zH2O where M is the cation

that compensates the negatively charged framework, n is the cation valency, y/x is the Si/Al

ratio, and z is the water content. Zeolite A or LTA is the zeolite used primarily by industry on

a mass basis (making up 73 % of global production (Ayele et al., 2016a)) due primarily to its

use in laundry detergents as a water softener (S. U. Meshram et al., 2014).

The concept of transforming waste and naturally occurring aluminosilicates into higher value

products such as zeolites has been studied in recent years in relation to the development of

green synthesis methods (H. Liu et al., 2014; Pan et al., 2019). Notably, due to the poor

reactivity of waste or natural aluminosilicates an additional activation stage is commonly

applied. For example, thermal treatment of SLR has been studied in relation to construction

industry utilisation (Z. Liu et al., 2019). It was discovered that the amorphous content increased

from ca. 17 to 51 % upon heating at 700 oC for two hours. Whereas,. Zhuang, et al. (2014)

reported that hydroxide dissolution followed by hydrothermal synthesis produced a zeolite

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 100
LTA/Faujasite mixture from lithium slag. Lin et al. (2015) added that it was necessary to

recycle the waste solution (“mother liquor”) when synthesising zeolites from lithium slag due

to economic considerations. Fusion pre-treatment of aluminosilicate waste materials to

activate the silica and alumina species is commonly used. The basic idea is to heat a mixture

of waste and sodium hydroxide at temperatures above the melting point of NaOH for several

hours. Hence, D. Chen et al. (2012) fused lithium slag with sodium hydroxide at 600 oC for

four hours and then aged and hydrothermally reacted the resultant solid at 95 oC for eight hours

to make zeolites.

The need for activation of waste materials appears to be a key difference compared to

commercial zeolite synthesis wherein aluminosilicate gels are created from soluble silicates

and aluminates. Therefore, efforts have focussed on developing low cost means of activating

waste and natural aluminosilicates. Fusion with NaOH is viewed as too expensive to be

relevant to industry (Ojumu et al., 2016). Consequently, a novel process was developed, which

involved the activation of rectorite and kaolin clay through a method known as “High

Concentration Alkali Solution” (HCAS) (T. Li et al., 2012; H. Liu et al., 2015; Yue et al.,

2014; Yue et al., 2015). The basis of this approach was hydrothermal reaction of the clay with

high molarity (ca. 15 M) NaOH solutions at temperatures of 200 oC or higher. The resultant

solid was then separated from the solution and reacted under appropriate conditions to make

zeolites such as LTA, ZSM-5 and USY. The critical insight was that a depolymerizat ion-

polymerisation strategy was necessary to make quality zeolites.

H. Liu et al. (2019) compared thermal, alkali fusion and HCAS activated rectorite upon the

synthesis of Ti-ZSM-5 zeolite. It was found that titanium species were partially incorporated

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 101
in the lattice framework of the zeolite when employing the alkali fusion or HCAS methods

whereas with thermal activation only extra-framework titania was recorded. Consequently, the

catalytic activity of the titanium incorporated ZSM-5 was greater for 1-octene

hydroisomerization and aromatization.

Miao et al. (2016) described the transformation of potassic rocks to low sodium zeolite X

(LSX) using the sub-molten salt approach. A mixture of powdered potassic rock, potassium

hydroxide, and water was heated with agitation in the temperature range 180 and 200 °C for

up to seven hours. The mass concentration of KOH was critical, with a value of 75 %

recommended, as the resultant product exhibited only peaks in the X-ray diffraction (XRD)

patterns ascribed to potassium hydroxide. J. Yang et al. (2017) more recently described how

to activate kaolin under relatively mild conditions in order to make them amenable for

conversion into zeolite LTA and zeolite Y. These authors indicated that the HCAS method

developed in their research group was not amenable to industry use. Therefore, a new strategy

termed Quasi-Solid Phase Activation Process (QSP) was developed. Initially, kaolin was

kneaded with sodium hydroxide and then a small fraction of water was added. Subsequently,

this material was extruded into shapes 1.5 mm in diameter and 2 to 3 cm in length. The

extrudate was finally heated to the required temperature in a belt type calcination oven. The

principal advantages relative to the HCAS approach were: (1) lower activation temperature;

(2) less water required; (3) continuous processing (as opposed to batch for HCAS); and (4)

reduced energy consumption. Yue et al., 2020) demonstrated that >95 % depolymerisation of

the rectorite was obtained by heating at 170 oC for one hour.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 102
Despite the advances in activation approaches it was evident that an ideal ‘One pot’ synthesis

method for waste-to-zeolite synthesis was missing (X. Ji et al., 2018). Consequently, this study

was based upon the innovative idea of initially activating the SLR using a relatively high

concentration of NaOH at a particular temperature for a certain time. Then additional water

was added to rapidly lower the NaOH molarity and control reaction temperature (Figure 25).

Therefore, the aim of this study was to: (1) discover SLR activation conditions which produce

high amounts of active silica and alumina; (2) adjust synthesis conditions by water addition in

order to make crystalline zeolite LTA. The hypothesis was: “If an economical approach to

synthesis of zeolites from waste or natural aluminosilicates then uptake of “green” zeolites by

industry may be accelerated”. Research questions addressed to support the hypothesis

included: (1) Which temperature and molarity of NaOH are preferred for the activation of SLR?

(2) What is the quality and yield of Zeolite LTA from activated SLR? (3) It is possible to

optimize the bench-scale synthesis of zeolite LTA from SLR powder, which was performed,

and extensive characterization of the activated product and zeolite was completed.

Figure 25: Illustration of the “Quench” concept for making zeolite from waste aluminosilicate
materials.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 103
Materials and Methods

Spodumene Leachate Residue & Chemicals


The SLR used for this study came from a lithium mining operation in Western Australia.

Sodium hydroxide (NaOH) (40 wt%) solution, sodium hydroxide extra pure micro-pearls and

hydrochloric acid (32 wt%) were purchased from ChemSupply Australia as analytical reagent

(AR) grade. Sodium aluminate (NaAlO 2) was purchased from Sigma-Aldrich.

Activation of Spodumene Leachate Residue (SLR)


Activating SLR for zeolite synthesis boosts reactivity and removes impurities, preparing SLR

for effective zeolite formation and influencing the quality of the final product. The activation

behaviour of SLR was ascertained as a function of temperature (25 to 150 ℃), sodium

hydroxide concentration (2 to 14 M), and time (10 to 1440 min). Room temperature SLR

dissolution trials involved addition of SLR (1 g) and 5 mL of sodium hydroxide (40% w/v)

solution at the desired molarity into 50 mL centrifuge tubes. This mixture was then activated

on a rotary tube mixer for up to 24 h. Activation at 75 oC was completed using a heating block

(HACH DRB 200), by mixing 1 g of SLR and 5 mL of sodium hydroxide (40% w/v solution)

in 10 mL tubes at the desired molarity. Samples were collected at various intervals from an

activation time of 10 minutes to 24 hours. The 100 ℃ activation tests were similarly completed

using 4 to 10 M NaOH solution which was prepared by addition of 1 g of SLR with 1.92 NaOH

g extra pure micro-pearls and deionised water. At the highest temperature of 150 oC a Teflon

lined Berghof BR100 High Pressure Reactor was utilized for the activation step. In this case,

samples were prepared using 20 g of SLR, 100 mL of sodium hydroxide (40% w/v solution)

and sufficient water to make the desired molarity. Notably, heating the high pressure reactor

to 150 oC took 50 minutes and cooling of the reactor to ambient temperature required another

60 minutes. Hence all reaction times in this vessel were assumed to be timed at 150 oC (and

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 104
not the total heating time including the ramp up phase). Each test was completed in duplicate

to ensure the veracity of the data. After the end of the activation process the samples were

filtered in order to collect the solid phase and the mother liquor. The supernatant for all samples

were collected and appropriately diluted for analysis using inductively coupled plasma optical

emission spectroscopy (ICP-OES). The solid products from this alkaline activation step were

characterised using XRD analysis.

Determination of active alumina and silica species in the solid product was achieved by acid

dissolution using the methodology described by Wang et al. (J.-Q. Wang et al., 2014). Solid

samples were initially oven dried at 105 oC to reduce the water content then 0.05 g of fused

product was mixed with 5 mL of 2 M HCl in a 50 mL centrifuge tube and agitated on a rotary

wheel for 30 min. The supernatant was subsequently diluted and analysed by ICP -OES. The

total amount of active Si and Al species was a combination of those detected in the solid and

liquid phases (Equation 10 & Equation 11).

𝑨𝒍 𝒊𝒏 𝒔𝒐𝒍𝒊𝒅 (𝒈) +𝑨𝒍 𝒊𝒏 𝒎𝒐𝒕𝒉𝒆𝒓 𝒍𝒊𝒒𝒖𝒐𝒓 (𝒈)


Equation 10: 𝑻𝒐𝒕𝒂𝒍 𝑨𝒄𝒕𝒊𝒗𝒆 𝑨𝒍 (%) = ∗ 𝟏𝟎𝟎
𝑨𝒍 𝒊𝒏 𝑺𝑳𝑹 (𝒈)

𝑺𝒊 𝒊𝒏 𝒔𝒐𝒍𝒊𝒅 ( 𝒈) +𝑺𝒊 𝒊𝒏 𝒎𝒐𝒕𝒉𝒆𝒓 𝒍𝒊𝒒𝒖𝒐𝒓 (𝒈)


Equation 11: 𝑻𝒐𝒕𝒂𝒍 𝑨𝒄𝒕𝒊𝒗𝒆 𝑺𝒊 (%) = ∗ 𝟏𝟎𝟎
𝑺𝒊 𝒊𝒏 𝑺𝑳𝑹 (𝒈)

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 105
Conversion of Activated Spodumene Leachate Residue to Zeolite LTA
Activation of SLR accompanied by subsequent zeolite synthesis was performed using a

Berghof BR100 High Pressure Reactor at either 100 or 150oC. SLR (10 g), NaOH (19.17 g)

and sufficient deionized water to control molarity (7 or 10 M) was added to the reactor.

The SLR was thus activated as a function of time (1 to 18 hours) at the desired temperature.

After SLR activation was completed the solution/slurry was reduced in both temperature (to

80 oC) and NaOH molarity (3.95 M). One constraint was that the time to reduce reactor

temperature from 100 to 80 ℃ in the same reactor was approximately 40 minutes for each

sample. Temperature was monitored with a thermocouple connected to the reactor panel

controller. Meanwhile a solution of deionized water with NaAlO2 (5.33 g) was prepared to

adjust the molarity of the synthesis and SiO 2/Al2O3 ratios for zeolite LTA. The solution was

heated at 80 ℃ and agitated at 100 rpm in a New Brunswick Innova 42R Incubator for four

hours.

Once the temperature of the solution containing the activated SLR decreased to 80 ℃, the

reactor was opened, and the pre-heated solution of water and sodium aluminate rapidly added

to the slurry. Then the reactor was re-sealed and maintained at 80℃ for the desired synthesis

time while stirring at 100 rpm. After the reaction was completed, the product was separated

by using a centrifuge. The mother liquor was collected and analysed using ICP-OES. The solid

product was washed with deionised water to reduce the pH and remove the free chemicals in

the product. Six washing cycles were completed using 120 mL of water for each cycle. After

the washing step the wet cake was weighed and dried in the oven at 105 °C. Then the dried

powder was characterised using Quantitative X-ray Diffraction (XRD), X-ray Fluorescence

(XRF) and Scanning Electron Microscopy (SEM). Equation 12 shows the stoichiometric

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 106
reaction to produce zeolite LTA from SLR and sodium aluminate. Hence, the theoretical yield

of zeolite LTA was calculated to be 1.78 kg for every 1 kg of SLR used.

Equation 12: 8 𝑯𝟐 𝑨𝒍 𝟐𝑺𝒊 𝟒.𝟓𝑶𝟏𝟑.𝟎. 𝟓 𝑯𝟐 𝑶 + 𝟐𝟎 𝑵𝒂𝑨𝒍𝑶𝟐 + 𝟑𝟒 𝑯𝟐 𝑶 + 𝟏𝟔 𝑵𝒂𝑶𝑯 →

𝟑 𝑵𝒂𝟏𝟐 𝑨𝒍𝟏𝟐 𝑺𝒊𝟏𝟐 𝑶𝟒𝟖. 𝟏𝟖𝑯𝟐 𝑶

Solid Characterization
X-ray Fluorescence (XRF)
Elemental quantitation of the solid products was carried out using a PANalytical AXIOS 1.5

kW Wavelength Dispersive XRF spectrometer and PANalytical WROXI software application.

Powdered samples were prepared as a fused glass disc with a standard composition of 50:50

lithium tetraborate: lithium metaborate with 0.05 % li-Iodate. Loss of Ignition (LOI) was

determined by heating solid samples at 1050 ˚C for two hours.

Quantitative X-ray Diffraction (XRD)


Crystalline phases in the solid products were obtained using a PANalytical X’PertPro powder

diffractometer. Powdered zeolite samples were mixed with 10 wt% of corundum standard

(1.800 g (±0.004) and 0.200 g (±0.004), respectively). Samples and standard were micronized

and homogenized with zirconia in a McCrone micronizing mill and dried in an oven for four

hours; prior to creating a pressed powder. To analyse the product phases TOPAS (V6, Bruker)

and JEdit software were used.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 107
Scanning Electron Microscopy (SEM)
To observe the shape and size of the zeolite products a JEOL 7001F Scanning Electron

Microscope was used. The samples were prepared on carbon stamps, and then coated in

platinum. The accelerating voltage was 5 kV.

Solution Analysis
The mother liquor and acid dissolution samples were analysed using a Perkin Elmer Optima

8300 DV ICP-OES spectrometer to primarily determine the concentrations of aluminium,

silicon, and sodium. Dilution of the samples was conducted in a Hamilton auto-diluter using

2% nitric acid as diluent.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 108
Results and Discussion

Characterization of Spodumene Leachate Residue


The elemental composition of SLR is displayed in Table 13, as measured by X-ray fluorescence

(XRF).

Table 13. XRF analysis of SLR


Oxide Mass % (Dry Basis) Oxide Mass % (Dry Basis)

SiO2 66.276 P 2O5 0.008


Al2O3 25.005 SO3 1.681
Fe2O3 0.624 Cr2O3 0.05
Na2O 0.118 NiO 0.007
K2O 0.276 ZnO 0.005
MgO 0.127 PbO 0.005
CaO 0.063 BaO 0.006
TiO2 0.019 Mn3O4 0.133
LOI 4.985

From the XRF data it was concluded that the bulk formula for the SLR was 2 H2 Al2 Si4.5 O13.0.5

H2O. Quantitative X-ray Diffraction (XRD) analysis of SLR revealed the sample to be

primarily Hydrogen Aluminium Silicon Oxide and non-diffracting material (Table 13The

slightly higher Si value derived from the XRF data may be related to the presence of quartz.

The small amount of lithium sulfate presumably arose as a residual of the lithium extraction

process using sulfuric acid.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 109
Table 14. Quantitative XRD analysis of SLR
Phase SLR delta Wt%

Hydrogen Aluminium Silicon Oxide 31.92

Quartz 3.85

Lithium Sulfate 0.19

Non-diffracting/unidentified 64.04

Sum 100

Scanning Electron Microscopy (SEM) images supported the XRD data in that the material

appeared random in size and shape (Figure 26). In terms of particle size, SLR comprised of

relatively small grains with a particle size range between 0.71 and 48 (µm); with a d50 = 11.9

(µm).

10

8
Volume Density (%)

0
0.01 1 100 10000
Size Classes (μm)

SEM Image Particle Size Distribution

Figure 26: SEM imaging and particle size of SLR

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 110
Activation of SLR Material

Creation of Active Silica and Alumina Species from SLR


A baseline study was conducted wherein the activation of SLR was completed at ambient

temperature using 2 to 14 M NaOH solutions. In general, increasing the NaOH molarity and

contact time promoted activation of both Si and Al species (Figure 27). This observation

agreed with previously published data relating to dissolution of fly ash (Kuenzel & Ranjbar,

2019). Regarding the active species in the solid phase, the interpretation of the mother liquor

data indicated that 10 and 14 M solutions favoured the activation of SLR. Overall, the

activation process was relatively slow with only 14.2 and 14.74 % of Si and Al activated after

24 hours.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 111
Active Si Species in Mother Liquor at 25℃ Active Al Species in Mother Liquor at 25℃

16 16
2M NaOH 2M NaOH
Si Active Species (%)

14

Al Active Species (%)


3M NaOH
14 3M NaOH
12 4M NaOH 4M NaOH
12 10M NaOH
10M NaOH
10 14 M NaOH 10 14M NaOH
8 8
6 6
4 4
2 2
0
0
0 5 10 15 20 25
0 5 10 15 20 25
Activation time (hrs)
Activation time (hrs)

Active Si Species in Solid at 25℃ Active Al Species in Solid at 25℃

16 16

Al Active species (%)


14
Si Active species (%)

14
12 12
10 10
8 8
6 6

4 4
10M NaOH 2 10M NaOH
2 14M NaOH 14M NaOH
0
0
0 5 10 15 20 25
0 5 10 15 20 25
Activation Time (hrs)
ActivationTime (hrs)

Figure 27: Active Si and Al species in (a) mother liquor and (b) solid as a function of NaOH
molarity and contact time at 25 oC

Therefore, the activation temperature was raised to 75 ℃; whereupon a substantial increase in

both active Si and Al species was apparent (Figure 28). In general, as the molarity of NaOH

increased from 2 to 14 M the formation of active Si and Al species was not only greater but

also the activation rate was higher. Reaction time had a significant impact when using lower

molarity NaOH solutions but minimal impact when using 10 or 14 M NaOH solutions.

Additionally, it appeared that use of either 10 or 14 M NaOH solutions activated the SLR to

approximately the same extent (ca. 76 and 80 % for Si and Al, respectively).

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 112
Total Si Active Species Total Al Active Species

100 100

Al Active Species (%)


Si Active Species (%)

80 80
2M NaOH 2M NaOH
60 3M NaOH 3M NaOH
4M NaOH
60
4M NaOH
10M NaOH 10M NaOH
40 14M NaOH 14M NaOH
40

20 20

0 0
0 5 10 15 20 25 0 5 10 15 20 25
Activation time (hrs)
Activation time (hrs)
Si Active Species in mother liquor Al Active Species in mother liquor

100 100
2M NaOH
Si Active Species (%)

2M NaOH
Al Active Species (%)
3M NaOH 3M NaOH
80 4M NaOH 80
4M NaOH
10M NaOH 10M NaOH
14M NaOH
60 60 14M NaOH

40 40

20 20

0 0
0 5 10 15 20 25 0 5 10 15 20 25

Activation time (hrs) Activation time (hrs)


Si Active Species in solid Al Active Species in solid

100 100
2M NaOH
Al Active Species (%)
Si Active Species (%)

3M NaOH
80 4M NaOH 80
10M NaOH
14M NaOH 2M NaOH
60 60 3M NaOH
4M NaOH
10M NaOH
40 40 14M NaOH

20 20

0 0
0 5 10 15 20 25 0 5 10 15 20 25
Activation time (hrs) Activation time (hrs)

Figure 28: Active Si & Al species from SLR as a function of temperature (75℃), NaOH
molarity, and contact time.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 113
The active species of aluminium were identified to reside mainly in the solid phase. In contrast,

the distribution of active Si species was approximately equal between the solid and liquid

phases. To gain a deeper insight as to the activation behaviour at 75 oC, the corresponding

quantitative XRD data for the solid phase was examined (Table 15).

Table 15: Quantitative XRD data for SLR activated with NaOH at 75 oC
NaOH Reaction Wt %
Concentration Time
Zeolite Non-
(M) Quartz Cancrinite Sodalite Zeolite X
(min) LTA diffracting
3 1080 2.10 0.75 0.19 0.64 0.78 95.54
3 1440 8.16 11.19 6.27 1.31 0.80 72.28

4 1080 2.49 0.38 0.92 1.03 4.82 90.36


4 1440 3.15 0.00 0.00 4.33 9.96 82.56

10 10 2.06 2.74 5.91 0.31 0.00 88.98


10 30 2.11 4.16 11.74 0.54 0.01 81.43
10 1080 2.08 8.21 15.83 0.73 0.17 72.98
10 1440 1.93 19.26 12.37 1.59 0.08 64.77

14.3 10 2.22 2.91 3.91 0.54 0.00 90.41


14.3 30 2.42 7.27 11.63 0.78 0.00 77.90
14.3 1080 2.63 18.57 15.62 0.81 0.06 62.30
14.3 1440 1.66 24.47 13.65 1.10 0.16 58.96

Regardless of the NaOH molarity used the general trend was a diminution in the presence of

non-diffracting material as the activation time was increased. For the low molarity NaOH

solutions (3 and 4 M) the appearance of crystalline phases was not significant until after 1440

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 114
minutes of activation. Alternatively, when 10 or 14 M NaOH solutions were employed, the

appearance of sodalite (Na8Al6Si6O24(OH)2) and cancrinite (Na 8Al6Si6O24(OH)2.2H2O) was

recorded. The cancrinite amount increased with activation time; whereas growth of sodalite

occurred during the first 1080 minutes but diminished after that time. This outlined behaviour

was in accord with the knowledge that sodalite can be transformed into the more stable

cancrinite. P. Zhang, et al. (2019) indicated that the formation of cancrinite occurred via initial

creation of an amorphous phase which transformed into zeolite LTA, then sodalite, and

ultimately cancrinite. Factors which accelerated the production of cancrinite included

increasing reaction time (X. Liu et al., , 2013), higher reaction temperature (Esaifan et al.,

2019), and elevated alkalinity levels (Wernert et al., 2020). As an example, Esaifan et al.

(2019) indicated that use of ≥ 10 M NaOH solutions when activating kaolin resulted in

cancrinite formation. Alternatively, these authors outlined that lower molarity NaOH solutions

(≤ 4 M) promoted formation of zeolite LTA, zeolite P and analcime.

Notably, the fact that active aluminium species resided mainly in the solid phase suggested

that NaAl(OH)4 (which is formed in alkaline, aqueous solutions) was barely present; and that

the active aluminium was probably part of an aluminosilicate material. Yue et al. (2020)

extensively characterized the solid products resultant from activation of rectorite clay using a

quasi-solid phase strategy. Analytical techniques included X-ray diffraction, solid state NMR,

infrared spectroscopy, Raman spectroscopy, Scanning Electron Microscopy (SEM) and X-ray

photoelectron spectroscopy (XPS). A multitude of species were detected including

Na6Al6Si6O24, Na4Al4Si6O20, Na4Al4Si4O16, Na4Al4Si5O18, and NaOH. Ultimately the

depolymerization process formed various monomers (Na 4SiO4 and NaAlO2), dimeric silica

(Na6Si2O7) and silica chains (Na 2SiO3). From Table 15 it was additionally shown that a

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 115
fraction of the Al species were present in both hydroxysodalite (Na 8(Al6Si6O24(OH)2.nH2O)

and cancrinite (Na 6Al6Si6O24(OH)2.2H2O) material.

On the basis that the Si/Al molar ratio in the SLR was 2.25, it was deduced that approximately

half of the Si present was residing in the mother liquor. Based upon the study of Yue et al.,

2020) monomeric Na4SiO4 (Equation 13) or dimeric Na6Si2O7 (Equation 14) were considered

as the final reactive forms of silica formed when depolymerizing aluminosilicates.

Equation 13:

2 𝐻2 𝐴𝑙2𝑆𝑖4.5 𝑂13 . 1.5𝐻2 𝑂 + 40 𝑁𝑎𝑂𝐻 → 9 𝑁𝑎 4 𝑆𝑖𝑂4 + 4 𝑁𝑎𝐴𝑙𝑂2 + 25 𝐻2 𝑂

Equation 14:

4 𝐻2 𝐴𝑙2 𝑆𝑖4.5 𝑂13 . 1.5𝐻2 𝑂 + 62 𝑁𝑎𝑂𝐻 → 9 𝑁𝑎 6 𝑆𝑖2 𝑂7 + 8 𝑁𝑎𝐴𝑙𝑂2 + 41 𝐻2 𝑂

Due to the outlined increase in activation of SLR when using a temperature of 75 oC, it was of

interest to examine whether a higher temperature of 100 oC was beneficial (Figure 29).

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 116
Total Si Active Species Total Al Active Species

100 100
Si Active Species (%)

Al Active Species (%)


80 80

60
60
40

40 20 4M NaOH
4M NaOH 5M NaOH
5M NaOH 7M NaOH
7M NaOH 10M NaOH
10M NaOH
20 0
0 5 10 15 20 25 0 5 10 15 20 25
Activation time (hrs)
Activation time (hrs)
Si Active Species in mother liquor Al Active Species in mother liquor

100 14
4M NaOH
Si Active Species (%)

12
Al Active Species (%)

5M NaOH
80 7M NaOH
10M NaOH
10
60 8

40 6
4
20 4M NaOH
5M NaOH 2
7M NaOH
10M NaOH
0 0
0 5 10 15 20 25 0 5 10 15 20 25
Activation time (hrs) Activation time (hrs)

Si Active Species in solid Al Active Species in solid

100 100
Si Active Species (%)

Al Active Species (%)

80 80

60 60

40 40

20 4M NaOH 20 4M NaOH
5M NaOH
5M NaOH 7M NaOH
7M NaOH 10M NaOH
10M NaOH
0 0
0 5 10 15 20 25 0 5 10 15 20 25
Activation time (hrs) Activation time (hrs)

Figure 29: Active species from SLR as a function of temperature (100 ℃), with different
NaOH molarity and contact time

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 117
The higher activation temperature notably improved the extent of active Si formation compared

to lower temperature data; albeit the use of NaOH molarity < 7 M still required 24 hours to

achieve almost complete Si activation. XRD analysis suggested that there was a clear trend of

increasing cancrinite formation not only as a function of the NaOH molarity but also when the

contact time was longer (Figure 30).

wt% Crystalline Cancrinite wt% Crystalline Sodalite

100 100
wt% Crystalline Cancrinite

4M NaOH 4M NaOH

wt% Crystalline Sodalite


5M NaOH
5M NaOH
7M NaOH
80 7M NaOH
10M NaOH 80 10M NaOH

60 60

40 40

20 20

0 0
0 5 10 15 20 25 0 5 10 15 20 25
Activation time (hrs) Activation time (hrs)

wt% Crystalline Zeolite LTA Non-diffracting/unidentified

100
Non-Diffracting/unidentified %

100
wt% Crystalline Zeolite A

4M NaOH 4M NaOH
5M NaOH 5M NaOH
80 7M NaOH
10M NaOH 80 7M NaOH
10M NaOH

60
60
40
40
20
20
0
0
0 5 10 15 20 25 0 5 10 15 20 25
Activation time (hrs) Activation time (hrs)
Figure 30: XRD analysis of active solid phase formed from reaction of SLR with NaOH
solution at 100 ℃; function of NaOH molarity and contact time.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 118
For example, for the 10 M NaOH solution the amount of cancrinite formed increased from 10.3

to 48.0 wt% as the contact time progressed from 10 to 360 minutes, and it improved to 67%

after 24 hours. Simultaneously, the presence of sodalite diminished from 18.8 to 12.9 wt% and

the detection of non-diffracting species decreased from 67.5 to 35.3 wt%. A minor zeolite

LTA phase was also identified which grew from 0.8 to 2.2 wt % with increasing contact time.

Cancrinite is the most stable phase formed when converting aluminosilicate gels to mixtures

of sodalite, zeolite LTA, and zeolite X (Ríos et al., 2009). Notably, Stewart (1941) indicated

that cancrinite is actually unstable in acid solutions and thus can, in theory, be classified as

“reactive”.

Further heating of SLR at 150 ℃ with a 10 M NaOH solution resulted in a reduction in

formation of active Si species after a one-hour contact time (ca. 5%) (Figure 31). Extension of

the reaction time only resulted in a slight decrease in active Si content. Similarly, use of 3 M

NaOH under similar conditions not only produced fewer active Si species, but also increasing

the contact time actually inhibited formation of active species.

XRD analysis (Figure 32) indicated that the amount of cancrinite increased with alkali

concentration and not with contact time. For example, the 10 M NaOH sample increased the

percentage of cancrinite in the first hour from 16 to 53%, before the amount stabilised until the

last contact time (18 hours); while the 3M NaOH sample remained stable at approximately 7%

until the four-hour mark. Subsequently, the amount of cancrinite significantly increased to

16.5%. The data in Figure 31 suggests that the stabilisation of cancrinite in the samples was

due a deficit of aluminium which agreed with the total amount of active species (Al in mother

liquor + Al in solid product) in the sample.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 119
Total Si Active Species Total Al Active Species

100 100
3M NaOH
3M NaOH
90 10M NaOH
Si Active Species (%)

Al Active Species (%)


10M NaOH
80
80
70 60

60 40
50
20
40
30 0
0 5 10 15 20 25 0 5 10 15 20 25
Activation time (hrs) Activation time (hrs)

Si Active Species in mother liquor Al Active Species in mother liquor

100 100
3M NaOH
Si Active Species (%)

3M NaOH
Al Active Species (%)

10M NaOH
80 80
10M NaOH

60 60

40 40

20 20

0 0
0 5 10 15 20 25 0 5 10 15 20 25

Activation time (hrs) Activation time (hrs)

Si Active Species in solid Al Active Species in solid

100 100
3M NaOH
10M NaOH
Si Active Species (%)

Al Active Species (%)

80 80

60 60

40 40

3M NaOH
20 20 10M NaOH

0 0
0 5 10 15 20 25 0 5 10 15 20 25
Activation time (hrs) Activation time (hrs)

Figure 31. Active species from SLR as a function of temperature (150℃), NaOH molarity
and contact time

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 120
wt% Crystalline Cancrinite wt% Crystalline Sodalite

100 100
wt% Crystalline Cancrinite

3M NaOH

wt% Crystalline Sodalite


10M NaOH
3M NaOH
80 80 10M NaOH

60 60

40 40

20 20

0 0
0 5 10 15 20 25 0 5 10 15 20 25
Activation time (hrs) Activation time (hrs)
wt% Crystalline Zeolite LTA Non-diffracting/unidentified

100 100
Non-Diffracting/unidentified %
wt% Crystalline Zeolite A

3M NaOH
80 3M NaOH
10M NaOH
90 10M NaOH

60 80
70
40
60
20
50
0
40
30
0 5 10 15 20 25
0 5 10 15 20 25
Activation time (hrs)
Activation time (hrs)
Figure 32. XRD analysis of active solid phase formed from reaction of SLR with NaOH
solution at 150 ℃; function of NaOH molarity and contact time.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New Quench
Method 121
Zeolite Synthesis after SLR Activation
Zeolite Synthesis without Pre-activation of SLR
In order to determine the baseline transformation of SLR to zeolite products an

experiment was conducted wherein SLR was reacted with an NaOH solution into

which had been added sodium aluminate in order to adjust the SiO 2/Al2O3 ratio to 0.95

(Figure 33).

a)
b)
Phase Wt%

Quartz 1.41

Cancrinite 2.55

Sodalite 1.22

Zeolite LTA 31.92

Non-diffracting 62.90

Figure 33: a) XRD and b) SEM image of the zeolite product produced by hydrothermal
reaction of non-activated SLR with with NaOH in a sodium aluminate solution at 80
oC for 4 h: SiO /Al O ratio = 0.95
2 2 3

The XRD analysis indicated that directly converting SLR to zeolite LTA is

challenging. Indeed, the dominant product was non-diffracting in character, and there

were also significant amounts of zeolite LTA and a minor fraction of sodalite detected.

Figure 33 also shows an SEM image of the solid product which comprised of two

distinct crystal types. In agreement with the XRD data, sodalite was evidenced by the

characteristic “thread-ball” or “urchin” shape (Daramola et a., 2017) and cubic

crystals ascribed to the presence of zeolite LTA were apparent.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New
Quench Method 122
The sodalite and zeolite LTA materials appeared to be growing out of a relatively

featureless aluminosilicate phase. This conclusion has also been reported by Andaç et

al. (2005) who demonstrated that before zeolite crystal growth, an amorphous phase

forms with the sodium aluminosilicate solution (Andaç et al., 2005).

Pre-activation of SLR with 10 M NaOH at 100 or 150 oC Prior to Synthesis of


Zeolite LTA
The influence of two activation temperatures (100 and 150 oC) upon the quality of

zeolite were evaluated according to XRD analysis (Figure 34). It was apparent that

zeolite LTA was predominantly synthesised from SLR activated at 100 oC. Moreover,

a reaction period of four hours resulted in greater formation of zeolite LTA (54 wt%)

compared to when reaction was conducted for 17 hours (25.6 wt%). Extended reaction

times reduced the presence of zeolite LTA and promoted the growth of cancrinite and

non-diffracting material. Regardless of the reaction time, when using the 100 oC

activated SLR, minimal sodalite was detected. When 150 oC activated SLR was

hydrothermally treated, the solid product was comprised mainly of cancrinite and non-

diffracting material. Increasing the reaction time from 4 to 17 hours notably enhanced

the production of cancrinite (43.2 to 55.4 wt%) and decreased the presence of non-

diffracting material (39.2 to 32.1 wt%).

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New
Quench Method 123
SLR Pre-activation Temperature
100 ℃ 150 ℃

70 70
Cancrinite
Cancrinite
Zeolite LTA
60 Non-Diffracting 60 Zeolite LTA
Non-Diffracting
wt% Crystalline

wt% Crystalline
50 50

40 40

30 30
20 20
10 10
0 0
4h 17h 4h 17h
Synthesis time (hrs) Synthesis time (hrs)

Figure 34. Influence of SLR activation temperature upon Zeolite LTA synthesis: 20 g
SLR, 19.17 g NaOH, 5.33 g NaAlO 2, 80 ℃ synthesis temperature, 3.95 M NaOH

Hence, a closer inspection of the zeolite products made using 100 oC activated SLR

was completed; in order to ascertain the optimum reaction time at 80 oC. Figure 34

shows that a four-hour reaction time was indeed the best condition for making zeolite

LTA. The growth of zeolite LTA inversely correlated with the amount of non-

diffracting material identified, as in Figure 35

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New
Quench Method 124
wt% Crystalline Zeolite LTA wt% Non-Diffracting/unidentified

wy% Crystalline Zeolite LTA


80 80
70

wt% Non-Diffracting
60 60
50
40 40
30
20
20
10
0
0 2 4 6 8 10 12 14 16 18 0
0 2 4 6 8 10 12 14 16 18
Synthesis time (hrs)
Synthesis Time (hrs)

wt% Crystalline Cancrinite wt% Crystalline Sodalite

80 80
wt% Crystalline Cancrinite

wt% Crystalline Sodalite


60 60

40 40

20 20

0 0
0 2 4 6 8 10 12 14 16 18 0 2 4 6 8 10 12 14 16 18
Synthesis time (hrs)
Synthesis time (hrs)

Figure 35. Importance of hydrothermal synthesis reaction time at 80 oC for SLR


activated at 100 oC

Sodalite was present in limited amounts throughout the synthesis period and

remarkably constant in value. In contrast, cancrinite was at the highest level (17.1

wt%) during the initial stage of the synthesis reaction and decreased to a relatively

stable value of ca 6.6%.

Scanning electron microscopy (SEM) images were also gathered in order to provide

deeper insight into the transformation of activated SLR into zeolite products (Figure

36). After one hour of hydrothermal synthesis the solid agglomerated product was not

only roughly spherical but also in the size range of 3 to 4 microns. Numerous globular

materials decorated the surface of the spheres.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New
Quench Method 125
1h 2h

3h 4h

6h 18h

Figure 36: SEM images of solid zeolite product formed as a function of time after
hydrothermal reaction using 3.95 M NaOH and Si/Al ratio = 0.95

Increasing the reaction time to two hours induced a loss of the majority of spherical

structures and growth of a platelet material.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New
Quench Method 126
The first evidence for formation of cubic crystals of zeolite LTA was also observed;

albeit their presence was not common. After three hours of reaction the formation of

cubic zeolite LTA crystals significantly increased. Concomitantly the amount of

platelet material decreased. After four hours reaction the solid product was seen to be

almost entirely comprised of zeolite LTA crystals (Figure 36).

As the reaction time was further extended to six hours the presence of zeolite LTA

crystals was diminished, and growth of spherical material promoted. After 18 hours

reaction the degradation of zeolite LTA was clearly evident; as was the formation of

larger, relatively featureless agglomerates which were accompanied by the

aforementioned spherical material.

Finally, XRF analysis was used to determine the composition of the zeolite product

made using a temperature of 80 oC for four hours (

Table 16).

Table 16. XRF analysis of solid product from Zeolite LTA synthesis using SLR
activated at 100 oC
Oxide Mass % (Dry Basis) Oxide Mass % (Dry
Basis)
SiO2 36.72 P2O5 0.00
Al2O3 29.43 SO3 1.026
Fe 2O3 0.279 Cr2O3 0.018
Na2O 19.53 NiO 0.002
K2O 0.074 ZnO 0.00
MgO 0.055 PbO 0.008
CaO 0.02 BaO 0.002
TiO2 0.037 Mn3O4 0.064
LOI 13.5

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New
Quench Method 127
The molar SiO2/Al2O3 ratio was 2.11 and the general composition was

Na13.1Al12.0Si12.7O50.1.15.6 H2O.

Pre-activation of SLR with 10 M NaOH at 100 oC Prior to Synthesis of Zeolite


LTA with SiO2/Al2O3 ratio of 1.9 in the Feed Mixture

Figure 37. SEM images of solid zeolite product formed after hydrothermal reaction
using 3.95 M NaOH and Si/Al ratio = 0.83 and 3-hour synthesis time

Since

Table 16 indicated that an excess of Si was present in the zeolite product, the amount

of NaAlO2 added to the synthesis mixture was adjusted to provide a slight excess of

aluminium species. Commercially, zeolite LTA is often made using excess aluminium

in order to guide the synthesis to zeolite LTA (Ameh et al., 2017). Figure 37 indicates

that the additional aluminium accelerated the formation of zeolite LTA as maximum

zeolite LTA content was noted after three hours of reaction (52.0 wt%) instead of four

(56.0 wt%). However, the amount of zeolite LTA in the solid product was not

increased.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New
Quench Method 128
60

wt% Crystalline 50

40

30 wt% Crystalline Cancrinite


wt% Crystalline Sodalite
20 wt% Crystalline Zeolite LTA
wt% Non-Diffracting/unidentified
10

0
2.0 2.5 3.0 3.5 4.0
Synthesis time (hrs)
Figure 38. Zeolite synthesis at 80 oC as a function of reaction time: SiO 2/Al2O3 ratio
= 0.83

These results showed that zeolite LTA synthesis is not only affected by the reaction

time but also by Si/Al content. For example, at three hours of reaction and an Si/Al

ratio of 0.95, 30.97wt% zeolite LTA was obtained whereas using the same reaction

time but increasing the aluminium content to an Si/Al ratio of 0.85, 52.0 wt% was

obtained. This conclusion agreed with previous research reported by Ameh et al.

(2017), who demonstrated that a decrease of Si/Al ratio in zeolite LTA synthesis has

a direct proportionality on the crystal size of the final product. Furthermore, it was

proved that crystallinity and framework structure can be improved or compromised by

slightly adjusting the aluminium content in the solution (Ameh et al., 2017).

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New
Quench Method 129
Conclusions

This study successfully validated the concept of using a quench method to initially

activate waste aluminosilicate materials (such as SLR) and then adjust reaction

conditions to form valuable zeolite products such as zeolite LTA.

Activation of SLR was dependent primality on two factors, molarity of the sodium

hydroxide and activation temperature. Testing revealed that an NaOH molarity of 10

M and a reaction temperature of 100 oC for no more than 1 hour was optimal.

Lowering the activation temperature reduced the activation rate which was not

desirable due to energy costs associated with heating the activation mixture. Whereas

use of higher activation temperatures favoured the formation of cancrinite as did higher

alkali molarities. Significantly, high temperature fusion processes usually reported in

zeolite literature are not required.

After quenching the reactant solution to reduce temperature, decrease NaOH molarity

and adjust SiO2/Al2O3 ratio; zeolite LTA was synthesised as verified by XRD patterns

and SEM images. The identification of substantial amounts of non-diffracting material

using quantitative XRD was not consistent with SEM images which revealed that

samples were highly crystalline.

Future studies should focus on development of continuous hydrothermal reactors

which can more rapidly decrease the reaction temperature and molarity to improve

control of the zeolite LTA synthesis.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New
Quench Method 130
Key Insights

This chapter presented an innovative solution that not only reduces waste but also

fosters a sustainable approach towards the production of valuable zeolites. By

optimizing the activation of SLR and synthesizing zeolite LTA using the "Quench

Method", this study successfully demonstrates the feasibility of converting waste

materials into valuable products. The activation of SLR involved varying the

temperature, sodium hydroxide concentration, and time of the process. Subsequently,

zeolite LTA synthesis was achieved by fine-tuning the sodium hydroxide molarity and

Si/Al ratio. The addition of a slight excess of aluminium in the solution further

improved the synthesis process. The results of this study hold promise for promoting

sustainability in the lithium industry by repurposing waste materials into high value

zeolites.

In order to optimize the method proposed, the utilization of continuous reactors has

been recommended. Nonetheless, one of the main challenges encountered in the

employment of continuous reactors for zeolite synthesis is the issue of viscosity. As

the synthesis progresses, the concentration of the reaction mixture increases, which in

turn causes the viscosity of the mixture to rise. This can lead to issues such as reactor

clogging, uneven heat distribution, and difficulties in controlling the reaction

conditions (Z. Liu, Zhu,et al., 2019). Nevertheless, this issue has been resolved in the

design by employing a combination of a continuous reactor and a batch reactor (see

Supplementary Figure 6). Specifically, the continuous reactor will be solely utilized

for the activation phase to obtain active Si and Al species while being capable of faster

quenching before it reaches the next reactor. By incorporating Industry 4.0 principles,

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New
Quench Method 131
notably machine learning, in conjunction with continuous reactors for zeolite

synthesis, several benefits can be derived.

The incorporation of machine learning algorithms can significantly enhance the

optimization of zeolite synthesis in continuous reactors, leading to reduced production

time and costs, improved product quality, and minimized waste. These algorithms can

analyse real-time data collected from sensors and other sources to dynamically adjust

the operating parameters of the reactor, including temperature, pressure, and flow rate.

This approach can effectively maximize the efficiency of the synthesis process.

Moreover, the predictive capabilities of machine learning algorithms can anticipate

potential problems and provide early warnings, enabling timely corrective actions to

be taken, thereby reducing downtime, and minimizing the need for manual

intervention, which is often time-consuming and prone to human error.

The chapter of this research focuses on the development of a robust dataset for machine

learning in zeolite synthesis. that study illustrates the difficulties encountered in

creating appropriate datasets for machine learning in the synthesis of zeolite LTA.

Comparing zeolite synthesis parameters between different laboratories is inherently

challenging, which can impair the accuracy of predictions generated by machine

learning. Thus, researchers must establish their own database for zeolite synthesis,

despite the time-consuming nature of this process which necessitates conducting

numerous experiments. Zeolite LTA is an exemplary case study for this chapter due to

its well-established synthesis process.

Chapter 5: “One Pot” Synthesis of Zeolites: Activation of Spodumene Leachate Residue using a New
Quench Method 132
Chapter 6: Factors which Influence Development of Robust Data Sets for

Machine Learning and Statistical Analysis of Zeolite Synthesis

Abstract

Machine learning is an important part of zeolite science which relies upon a reliable and

comprehensive dataset. However, there exist experimental challenges for zeolite researchers to

create suitable datasets and avoid problems associated with the use of contradictory data. This study

illustrates the situation using zeolite LTA synthesis as an example. First, quantitative X-ray

diffraction (XRD) was preferred to qualitative data as it accounted for non-diffracting material.

Secondly, an accurate mass balance was required since a substantial amount of valuable chemicals

were in the mother liquor and the wash water. Zeolite synthesis data from various hydrothermal

reactor configurations was shown to significantly influence zeolite LTA synthesis. The heat transfer

behaviour of the reactors was of relevance as this controlled heating rates, actual reaction

temperatures and cool down periods. Pre-heating of the zeolite synthesis mixtures was a potential

means of mitigating variation in zeolite synthesis, as it minimised variations in the heating up time.

The influence of stirring or static conditions while the reactants were converted to zeolite product

was another factor which altered zeolite product quality. Static conditions were favourable in terms

of zeolite properties but industrially this may not be an option as the larger vessels involved require

stirring to maintain homogeneity. In summary, it is inherently problematic to compare zeolite

synthesis parameters between laboratories. This latter finding also makes it difficult to use web

crawling approaches to mine data as the collected information is biased. Nevertheless, addressing

the issues allows meaningful results in a researcher’s laboratory.

Key Words: zeolite LTA; synthesis; machine learning; quantitative XRD; statistics

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and Statistical Analysis of
Zeolite Synthesis 133
Introduction

Recent years have witnessed the growth of machine learning and statistical analysis of zeolite

synthesis and structural types (Conterosito et al., 2020; Gaillac et al., 2020; Jensen et al., 2019;

Moliner, et al., 2019; Muraoka, et al, 2019; S. Yang et al., 2010). For example Moliner, et al. (2019)

described the use of principal component analysis (PCA) to dimensionally reduce powder X-ray

Diffraction (XRD) patterns to three structural principal components, when studying the synthesis of

ITQ-21 and ITQ-30 zeolites. As an outcome, predictions were made regarding the composition of

the zeolites and the degree of crystalline material present. Alternatively, Jensen et al. (2019)

demonstrated the ability of a Generative Neural Network to discover new organic structure directing

agents which could be employed to make novel zeolite structures. Likewise, S. Ma et al. (2020)

used machine learning approaches to develop improved zeolite synthesis strategies. Due to the ability

of machine learning to work with big data there is potential to accelerate the use of high throughput

methods in zeolite science (Gaillac et al., 2020; Moliner, et al., 2019; Raccuglia et al., 2016; Serra et

al., 2007).

Despite the identified potential for machine learning approaches there is an issue with respect to

access to reliable training data, and ensuring that the model is not over-fitted to the training data

(Selvaratnam & Koodali, 2021). As such fundamental criteria to be satisfied, include provision of

clean data sets (Shahzad et al., 2017). For example, when examining zeolite synthesis, the data

should be repeatable, consider all the experimental variables, and be precise. In practice, data sets

tend to be too small, although there are indications as to how to manage small datasets in material

science (Y. Zhang & Ling, 2018a).

There also appear to be issues with the experimental data collected. For example, interpretation of

X-ray Diffraction (XRD) data is a point of debate in terms of generating a quality data set. XRD is

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and Statistical Analysis of
Zeolite Synthesis 134
the most common method for determining zeolite product quality, but most publications use only

qualitative XRD methods and not quantitative XRD analysis. Qualitative XRD compares the relative

area of reflections ascribed to crystalline peaks for each phase present and does not account for non -

diffracting/amorphous content (J.-Q. Wang et al., 2014). Alternatively, quantitative X-ray diffraction

(qXRD) involves the addition of an internal standard which subsequently allows the speciation of a

sample including any amorphous/non-diffracting material present (Längauer et al., 2021).

Accordingly, Gougazeh and Buhl (2014) described the use of Rietveld analysis using TOPAS

software to determine the amount of zeolite LTA, hydroxysodalite, quartz and non-diffracting

material in a material made from hydrothermal reaction of clay.

In terms of zeolite synthesis, the literature rarely indicates the zeolite yield when studying synthesis

conditions. For example, numerous publications show they have made a zeolite based upon XRD

patterns (Jafari et al., 2014; J.-Q. Wang et al., 2014); but no mention of the overall process yield is

made (Yu et al., 2018). This situation makes it difficult to construct a mass balance, and in some

instances determine how to recycle alkali solutions, dissolved aluminium, or dissolved silicates.

There are also problems relating to how comparable data is between different laboratories. For

instance, many studies make zeolites in PTFE lined hydrothermal vessels which are inherently

problematic in that the sample does not rapidly heat when placed in an oven (Hang Chau et al., 2000;

Johnson & Arshad, 2014; Katsuki et al., 2005). Indeed, due to variables such as vessel

volume/geometry, thickness of steel wall, and use of PTFE liner or not, the actual reaction

temperature achieved and time to do so, makes it difficult when trying to understand synthesis data.

Hence, in industry the starting solutions are often pre-heated to the reaction temperature to create a

more robust and economical process (Kettinger, Laudone, & Pierce, 1979).

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and Statistical Analysis of
Zeolite Synthesis 135
A new approach to improving heating characteristics is the use of continuous reactors where a tubular

reactor is rapidly heated in for example a pre-heated oil bath (Z. Liu, Wang et al., 2019).

To illustrate the challenges relating to quality data sets this study focussed on zeolite LTA

(Na12((AlO2)12(SiO2)12)·27H2O) as this zeolite remains important due to large scale use in

applications such as laundry detergents (Collins et al., 2020). From an industry perspective the focus

is to make zeolite LTA with the highest yield, optimal physical properties, and with minimal cost.

Notably, the economics for zeolite LTA production have not been addressed in published literature

to any significant extent. Hong et al. (2017) reported a cost study based upon fusion of coal fly ash

followed by hydrothermal synthesis of zeolite LTA. Such a process was suggested to have a positive

net present value (NPV) and payback in ca. seven years. Overarching these technical objectives was

the need to also be environmentally responsible and minimize waste streams (Casci, 2005; Fawer et

al., 1998).

Zeolite LTA synthesis is a well-researched area with many papers published in the academic

literature. However, the primary focus has been in relation to zeolite synthesis chemistry whereas

process engineering aspects have been not addressed in detail. The economics of zeolite LTA

manufacture relates to factors such as price/consumption of raw materials (especially sodium

hydroxide), utilities demand, process design, reaction time and temperature, capital expenditure,

labour and maintenance requirements, and disposal costs for waste streams. A mass balance of the

reactants and products is required as the basis of process design.

Another issue that is often neglected is the composition of the mother liquor recovered when the

zeolite slurry is separated into solid and liquid fractions. Recycling of excess sodium hydroxide to

the initial mixing stage is critical in terms of making the process viable (Bauer, 2008).

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and Statistical Analysis of
Zeolite Synthesis 136
Zeolite LTA has an inherent problem since it produces a highly alkaline solution when added to

water. Extensive washing is often required ( Hong, Maneerung, Koh, Sibudjing, & Wang, 2017) and

this situation can result in an imbalance in water demand/use for the production unit. Consequently,

some companies evaporate a portion of the water in the mother liquor and then reuse this water in the

washing stage (Deabriges, 1982). Following the washing stage zeolite LTA is typically spray dried

to form a uniform powder. Depending upon the application for the zeolite LTA material it may also

be shaped into extrudates or spherical beads (Bingre, Louis, & Nguyen, 2018). Common binders

include clay, pseudoboehmite, and water glass. The mixed product is then calcined at high

temperature to improve both mechanical and textural properties (Bingre et al., 2018).

Ultimately this situation means that optimization of the entire zeolite synthesis process is often not

completed, and the quality of data required for machine learning is flawed. Therefore, the aim of this

study was to develop strategies to improve zeolite LTA synthesis data which could subsequently be

used in machine learning. The hypothesis was that “if process engineering and machine learning

approaches are combined then this may result in more economical, higher quality and

environmentally sustainable synthesis of zeolite LTA”. Research questions which addressed the

hypothesis included: (1) Does quantitative XRD provide superior data when characterizing zeolites?

(2) What is the advantage of ensuring that a mass balance is reported when making synthetic zeolites?

(3) What is the variability of zeolite physical characteristics if made using the same conditions in

different hydrothermal reactor vessels? (4) Can pre-heating of reactant solutions facilitate the growth

of more uniform zeolite powder? (5) Do stirring or static conditions favour superior zeolite synthesis?

Various hypothermal vessels were employed to synthesise zeolite LTA using a range of synthesis

variables including: SiO2/Al2O3 ratio; H2O/Na2O ratio; Na2O/SiO2 ratio; reaction temperature; and

reaction time. Ultimately this study was designed to prepare zeolite researchers to maximise machine

learning outcomes for zeolite synthesis strategies; by understanding that need for data quality.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and Statistical Analysis of
Zeolite Synthesis 137
Materials and Methods

Chemicals
Sodium aluminate (NaAlO2), sodium metasilicate pentahydrate (Na 2SiO3•5H2O), and NaOH (extra

pure micro- pearls) were purchased from Chem-Supply, Australia as analytical reagent grade (AR).

Zeolite LTA Synthesis


For the samples at room temperature, sodium hydroxide was dissolved at ambient temperature in de-

ionised water. Then the sodium metasilicate was added to the sodium hydroxide solution. In a

separate container the sodium aluminate was dissolved into an equivalent amount of de-ionised water.

Ideally, the volumes of the aluminate and silicate solutions should be approximately the same to

encourage homogeneous mixing. The two solutions were rapidly combined into a 125 mL Nalgene

bottle and vigorously stirred for 30 seconds; to again promote uniformity of the resultant gel. The

mixture was then aged at ambient temperature for 0.5 h. After aging the Nalgene bottle was placed

in an incubator (Innova 43R, New Brunswick) at 80 oC for the synthesis period.

For the pre-heated feed mixture, sodium hydroxide pearls and sodium metasilicate were dissolved in

deionized water located in a stainless-steel beaker. This mixture was then heated to 80 ℃. In a

second beaker a solution of deionized water and sodium aluminate was heated to 80℃. Once both

solutions reached reaction temperature, they were combined into a 125 mL Nalgene bottle in which

an aluminosilicate gel rapidly formed. Then, the reactant mixture was placed in an incubator at 80℃

and agitated at 100 rpm for a specific synthesis time. Note that after forming the gels both samples

(stirred and static) were shaken in the bottles to destroy the gel structure before going into the

synthesis reactor.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and Statistical Analysis of
Zeolite Synthesis 138
Once zeolite synthesis was completed, the samples were removed from the incubator and centrifuged

to separate the mother liquor from the solid product. After the removal of the mother liquor the solid

product was washed until the pH was ≤11. The product was then dried and crushed at which point it

was stored in a sealed vessel. In terms of precursor conditions, the H2O/SiO2 molar ratio was varied

between 15 and 35 which was in accord with the comprehensive study by Kostinko (1985) which

suggested this was the preferred range for zeolite LTA formation.

Material Characterization

X-ray fluorescence (XRF)


A wavelength dispersive X-ray fluorescence (PANalytical Axios WD-XRF) instrument was

employed to record the zeolite composition. First 1 g of zeolite powder was placed in a 95/5 % Pt/Au

crucible. Then 10 g of vitreous 50:50 lithium tetraborate: lithium metaborate flux was introduced.

The samples were subsequently placed in the fusion equipment (TheOx, Claisse Scientific) for 20

minutes at 1050 °C. The resultant glass discs were analysed and a value for the Loss on Fusion (LOF)

was estimated.

X-ray diffraction (XRD)


X-ray diffraction (XRD) traces were acquired using a Panalytical X'Pert wide angle X-ray

diffractometer with Co Kα radiation (1.7903 Å). XRD patterns were collected every 0.02° in the

range of 5 to 90° 2θ using a rate of 30 seconds per step. Quantitative XRD analysis was conducted

by introduction of a 10 wt% corundum internal standard. XRD data was interpreted by TOPAS v5

software which used the Rietveld method. The reference patterns for zeolite LTA were referenced

from PDF-4 entry 04-016-9920.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and Statistical Analysis of
Zeolite Synthesis 139
Scanning Electron Microscopy (SEM)
A JEOL 7001F Scanning Electron Microscope (SEM) imaged the zeolite materials. Zeolite samples

were prepared by placing on carbon tape and coating with platinum. An accelerating voltage of 5 kV

and probe current of 8 was employed.

Particle Size Distribution


Particle size distribution of the zeolite product was acquired using a Mastersizer 3000 instrument

(Malvern Panalytical, UK). The powders were mixed with deionized water and ultrasonicated for a

time of 30 minutes.

Solution Analysis

Inductively Coupled Plasma - Optical Emission Spectroscopy (ICP-OES)


Solutions were filtered with a 0.45 µm syringe filter and analysed by Inductively Coupled Plasma

Optical Emission Spectroscopy (ICP-OES). Appropriate standards were used to calibrate the

machine.

pH and Conductivity
The solution pH was determined using a labCHEM-CP benchtop from TPS (Australia). Regular

calibration was enacted using standard buffer solutions (pH 4, 7 and 10) supplied by Rowe Scientific,

Australia.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and Statistical Analysis of
Zeolite Synthesis 140
Results and Discussion

The Use of Quantitative and/or Qualitative XRD


The application of X-ray diffraction to characterize zeolite products is ubiquitous.

There are a range of approaches which generate data which range from qualitative to

quantitative. For example, qualitative XRD simply identifies the crystalline phases

present and indicates if the materials are in trace, minor, or major amounts. The next

level of detail is semi-quantitative wherein not only is the crystalline phase identified

but also a relative percentage of each mineral is provided (derived from application of

the normalized reference intensity ratio). Semi-quantitative analysis can also suggest

what the amorphous/non-diffracting amount is in the sample. Spiking the sample can

result in more accurate data interpretation. In theory, the preferred XRD analysis

method is quantitative in nature. The basis of this technique is to incorporate an

internal standard such as corundum and then complete Rietveld analysis. However,

quantitative XRD requires more complex sample preparation, and more in-depth

analysis (Shams et al., 2021). Inherent to machine learning is the need for accurate

data input, hence it appears necessary to focus on quantitative XRD analysis of zeolite

samples.

Figure 39 shows the quantitative XRD data recorded for the solid products made at 80

oC using a gel composition of SiO 2/Al2O3 = 2; Na2O/SiO2 = 2.5 and H2O/Na2O = 15

to 35 as a function of reaction time. A trend which was common to all H 2O/Na2O

values employed was a clear maximum in zeolite LTA content as a function of reaction

time. The time at which this maximum zeolite LTA content was identified was greater

as the H2O/Na2O ratio was increased, with the optimal synthesis time identified at 4

hours. In addition, the formation of impurity phases such as cancrinite and sodalite

was observed, especially when the H 2O/Na2O ratio was 15 (Table 17). As indicated by
Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 141
Roland (1989) sodalite is a known impurity phase when reaction conditions are sub-

optimal. Similarly, Buhl (2016) reported creation of both sodalite and cancrinite as

by-products when performing autothermal synthesis of zeolite LTA. Cancrinite is

regarded as the thermodynamically most stable species in the Na 2O-SiO2-Al2O3-H2O

system (Peng et al., 2018).

A study by Kostinko (1982) addressed the relationship of synthesis variables when

making zeolite LTA, X and Y. It was shown that careful control of reaction conditions

was required to avoid making mixed phases of either zeolite LTA and hydroxysodalite

or zeolite LTA and zeolite X. An H2O/Na2O ratio of 25 was claimed to produce

relatively pure zeolite LTA within a one-hour reaction time at 100 oC (SiO2/Al2O3 =

2; Na2O/SiO2 = 2.4). Accordingly, in Table 17 it is also evident sodalite and cancrinite

formation formed when using high molarity alkali solutions (H 2O/Na2O < 20).

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 142
100 100
Zeolite LTA Zeolite LTA
Non-diffracting Non-diffracting
80 80

wt% Crystalline
wt% Crystalline
60
60

40
40

20
20

0
0 1 2 3 4 5
1 2 3 4 5
Time (hrs)
Time (hrs) H2O/ Na2O = 20
H2O/Na2O = 15
100 100
Zeolite LTA
Non-diffracting
80 80

wt% Crystalline
wt% Crystalline

60 60
Zeolite LTA
Non-diffracting
40 40

20
20

0
0 1 2 3 4 5
1 2 3 4 5
Time (hrs)
Time (hrs)
H2O/ Na2O = 30
H2O/Na2O = 25
100

80
wt% Crystalline

60
Zeolite LTA
Non-diffracting
40

20

0
1 2 3 4 5

Time (hrs)
H2O/ Na2O = 35
Figure 39: Quantitative XRD patterns for Zeolite LTA: SiO2/Al2O3 = 2; Na2O/SiO2 =
2.5; H2O/SiO2 = 15 to 35; Temperature = 80 oC: Note that trendlines are only a
visual aid.

Notably the maximum amount of crystalline zeolite LTA estimated from quantitative

XRD analysis was 90.24 wt% (when H 2O/Na2O ratio = 35). In comparison, qualitative

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 143
XRD analysis appeared to overestimate the amount of crystalline zeolite LTA present

(Table 17).

Table 17: Summary of quantitative XRD analysis of zeolite samples made at 80 oC;
SiO2/Al2O3 = 2; Na2O/SiO2 = 2.5 and H2O/Na2O = 15 to 35; stirring 100 rpm
wt% Zeolite
wt%
Reaction wt% wt% wt% Non- LTA
H2O/Na2O Zeolite
Time (h) Cancrinite Sodalite diffracting (Qualitative
LTA
XRD)
15 1 0.00 0.02 64 35.85 99.92
15 2 1.8 1.06 59 38.46 95.26
15 3 4.37 4.99 34 56.44 78.51
15 4 5.63 7.37 21 65.55 62.12
15 5 5.03 5.42 20 69.15 66.13

20 1 0.00 0.00 38.15 61.68 99.56


20 2 0.00 0.19 68.85 30.88 99.61
20 3 0.30 0.48 75.26 23.88 98.87
20 4 1.15 1.36 60.09 37.38 95.98
20 5 1.29 2.53 49.53 46.64 92.82

25 1 0.18 0.00 8.82 90.94 97.35


25 2 0.02 0.00 77.75 22.17 99.90
25 3 0.00 0.01 73.28 26.63 99.88
25 4 0.09 0.00 79.72 20.13 99.81
25 5 0.05 0.12 56.62 43.15 99.58

30 1 0.10 0.05 26.64 73.18 99.33


30 3 0.07 0.02 79.40 20.41 99.77
30 4 0.06 0.01 82.87 16.97 99.81
30 5 0.00 0.00 79.27 20.58 99.81

35 1 0.14 0.01 30.01 69.83 99.47


35 2 0.00 0.00 54.39 45.14 99.14
35 3 0.00 0.04 84.74 14.92 99.59
35 4 0.08 0.02 90.24 9.46 99.66
35 5 0.00 0.00 78.51 21.09 99.48

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 144
Notably, with the quantitative XRD method the data was consistent with the general

formation of zeolites presented by (Cundy & Cox (2005). For example, in Figure 39

when the H2O/Na2O ratio was 35, the initial induction/nucleation period was evident

wherein the initial zeolite growth occurs. Increasing reaction time further promoted

zeolite LTA crystal growth until a maximum amount of crystalline zeolite LTA is

present. Finally, the metastability of zeolite LTA was apparent for reaction time >

four hours when amorphous material was forming at the expense of zeolite LTA. In

contrast, the qualitative data did not agree with accepted zeolite synthesis theory due

to the lack of insight into the amount of amorphous/non-diffracting material present.

It is noted that previous literature has reported relatively high levels of

unidentified/amorphous material when making zeolite from waste aluminosilicates.

For example, . Längauer et al. (2021) estimated the amorphous/unidentified content of

zeolites when synthesised from coal fly ash to be in the range of 55 wt%.

SEM images confirmed the appearance of zeolite LTA in the samples (Figure 40). For

the sample prepared at H2O/Na2O = 15, initially after one hour of reaction the presence

of sub-micron sized crystals was observed as part of a larger cluster. Extension of the

reaction time broke up the clusters and the presence of well-defined crystals lessened.

By five hours of reaction the material present was not only relatively amorphous in

appearance but highly dispersed. In contrast, SEM images recorded when the

H2O/Na2O ratio was increased to 35 indicated a different pattern of behaviour. After

one hour the sample appeared to lack distinct crystals with the material best

characterized as amorphous in character. However, lengthening of the reaction time

to three hours promoted the growth of cubic zeolite LTA crystals out of an amorphous

mass. The distinctive cubic crystals of zeolite LTA became more abundant after five

hours of reaction and became relatively well dispersed. The particles were visually
Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 145
estimated to range from 0.25 to 2 microns. The trends observed in the SEM images

correlated well with the quantitative XRD analysis but did not agree with the results

of qualitative XRD analysis (Table 17 & Figure 39).

H2O/Na2O = 15 H2O/Na2O = 35

1 hour

3 hours

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 146
5 hours
Figure 40: SEM images of zeolite made at 80 oC as a function of time: SiO 2/Al2O3 =
2; Na2O/SiO2 = 2.5 and H2O/Na2O = 15 or 35 as indicated

Mass Balance for Zeolite LTA Synthesis


A mass balance for the synthesis of zeolite LTA at 80 oC for four hours (SiO2/Al2O3 =

2; Na2O/SiO2 = 2 and H2O/Na2O = 30) is shown in Figure 41. To facilitate

examination of the mass balance for zeolite LTA synthesis the ideal stoichiometric

equation is shown in Equation 15.

Equation 15: 𝟏𝟐 𝑵𝒂𝟐 𝑺𝒊𝑶𝟑. 𝟓 𝑯𝟐 𝑶 + 𝟏𝟐 𝑵𝒂𝑨𝒍 𝑶𝟐 →


𝑵𝒂𝟏𝟐 𝑨𝒍𝟏𝟐𝑺𝒊𝟏𝟐 𝑶𝟒𝟖.𝟐𝟕𝑯𝟐 𝑶 + 𝟐𝟒 𝑵𝒂𝑶𝑯 + 𝟐𝟏 𝑯𝟐 𝑶

Alternatively, the actual experimental conditions which relate to the molar oxide ratios

presented above, are illustrated in Equation 16.

Equation 16: 𝟓 𝑵𝒂𝟐 𝑺𝒊𝑶𝟑. 𝟓 𝑯𝟐 𝑶 + 𝟓 𝑵𝒂𝑨𝒍𝑶𝟐 + 𝟓 𝑵𝒂𝑶𝑯 + 𝟐𝟕𝟐. 𝟓 𝑯𝟐 𝑶 →


𝟎. 𝟒𝟏𝟕 𝑵𝒂𝟏𝟐 𝑨𝒍𝟏𝟐 𝑺𝒊𝟏𝟐 𝑶𝟒𝟖 . 𝟐𝟕𝑯𝟐 𝑶 + 𝟏𝟓 𝑵𝒂𝑶𝑯 + 𝟐𝟖𝟏. 𝟓 𝑯𝟐 𝑶

In terms of a general mass balance for the synthesis process the total input was 110.1

g which was balanced by the product streams: (1) mother liquor; (2) wet zeolite cake;

and (3) material lost to the washing stage. The predicted mass of zeolite LTA formed

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 147
was 11.1 g if all starting aluminium and silicon species in the initial gel had been

completely transformed. However, the actual zeolite LTA product mass was 8.40 g.

When consideration of the composition of the washing stage and mother liquor were

considered, the value of 8.40 g was optimal when the aluminium and silicon losses

were accounted for.

The mother liquor did not contain major amounts of aluminium and silicon species,

but it did comprise of substantial quantities of sodium (presumably from the presence

of excess sodium hydroxide as suggested in Equation 16). Such large volumes of

sodium hydroxide represent a notable cost to zeolite producers and thus industry

usually recycles mother liquor to the synthesis stage.

For example, Deabriges (1982) described the necessity to recycle both mother liquor

and wash water to recover valuable chemicals. Notably, these streams were heated to

vaporize excess water which was not required for the zeolite synthesis stage. This

vaporized water could then be condensed and reused for washing purposes to reduce

overall water consumption. It was thus apparent that in terms of using machine

learning to maximise the efficiency of zeolite synthesis that collection of mother liquor

analysis data was critical.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 148
Sodium Metasilicate Pentahydrate
g mol
Mass 12.94 0.06
0 2.93 0.18 Mother Liquor (ML)
Si 1.71 0.06 Mass 78.77 g
Na 2.80 0.12 68.5 mL
H2O 5.49 0.30 Si 1495 mg/L
Al 788.7 mg/L
Solid NaOH Na 68020 mg/L
g mol
Mass 4.88 0.12 g mol
O 1.95 0.12 Si 0.10 0.004
Na 2.80 0.12 Aluminosilicate Gel Al 0.05 0.002
H 0.12 0.12 g mol Na 4.66 0.203
Mass 110.10
Water Added to NaOH O 6.83 0.43 Water Mass Loss Washing 7.99 g
g mol Si 1.71 0.06 Incubator - 80oC- 4h
Mass 43.64 2.42 Al 1.65 0.06 Hydrothermal synthesis-100 Slurry Washing Solution
Na 7.01 0.31 RPM g mol
Sodium Aluminate H2O 92.77 5.15 Si 0.29 0.01
Mass 5 g Al 0.35 0.01
g mol Wet Cake Na 1.85 0.08
Na 1.40 0.06 23.34 g
Al 1.65 0.06
O 1.95 0.12
Dried Solid Residue
Water Added to Sodium Aluminate Mass 8.34 g
g mol
Mass 43.64 2.42 g mol
Si 1.32 0.047
Mass Balance Major Cations Al 1.24 0.046
Inlet (g) Outlet (g) Outlet Total (g) Na 0.50 0.022
ML Washing Solid
Si 1.71 0.1 0.29 1.32 1.71
Al 1.65 0.05 0.35 1.24 1.64
Na 7.01 4.66 1.85 0.5 7.01

Figure 41: Mass balance for the synthesis of zeolite LTA

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and Statistical Analysis of Zeolite S ynthesis 149
Nevertheless, most zeolite synthesis publications have not analysed the mother liquor,

or indeed described the impact of mother liquor recycling upon not only product

quality but also process economics. Recently, Probst et al., (2022b) showed that

mother liquor recycling when making zeolite N improved the growth of the zeolite

phase due probably to the presence of seeds in the reactant mixture. The

technoeconomics of zeolite N synthesis was also recently shown to be unfavourable if

mother liquor was not recycled (Probst et al., 2022a). In addition, these authors

reported that it was also desirable to recover the first stage wash water as it contained

dissolved silicon and aluminium species in amounts worthy of reusing. This

observation was also in accord with the above mentioned study by Deabriges (1982)

and the data shown in Figure 41 in this investigation.

Influence of Reactor Design upon Zeolite Synthesis


Figure 42 illustrates the temperature profiles when using either a reaction vessel placed

in a heated incubator or a hydrothermal reactor vessel. The reactant mixture did not

achieve the set-point of 80 oC until 2.5 hours after placing the reactor vessel into the

incubator. In contrast, when using the heated hydrothermal reactor, the reaction

temperature overshot the target value of 80 oC, then eventually stabilized at 80 oC after

1.6 hours. The presence of overheating in teflon lined batch reactors is viewed as a

critical factor which should be monitored or at least modelled (Deneyer et al., 2020).

However, most zeolite synthesis publications simply state that they placed the

hydrothermal vessel in a heated oven. For example, Xiao-Yong et al., (2016) used a

teflon lined stainless steel pressure vessel which was placed in an oven at 100 oC for

one hour to convert reactive aluminosilicate gel to zeolite LTA.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 150
As no monitoring of the temperature within the pressure vessel was conducted it was

unknown what the actual temperature profile during reaction was. It is noted that at a

hydrothermal reaction temperature of 100 oC, formation of hydroxysodalite was

promoted (Burriesci et al., 1984). Therefore it is reasonable to assume the actual

reaction temperature in the study by (Xiao-Yong et al. (2016) was most likely lower

than 100 oC.

SEM images indicated that the solid collected after hydrothermal reaction in the

incubator comprised of relatively small, cubic crystals (≤ 1 microns) (Figure 42).

Whereas the zeolite LTA crystals were mostly larger (ca. 2 microns) when made in the

Berghof hydrothermal reactor. It was noted that zeolite formed in the Berghof reactor

was less uniform in size compared to the zeolite sample prepared in the heated

incubator.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 151
Aluminosilica gel Temperature ( oC)

Aluminosilica gel Temperature ( oC)


100 100
80 80
60 60
40 40
20 20
0 0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
Reaction Time (hrs) Reaction Time (hrs)

a) Incubator (b) Berghof Hydrothermal Reactor

(c) Zeolite synthesis in the incubator (d) Zeolite synthesis in the Berghof
after 3 hrs synthesis- 75 wt% reactor after 3 hrs synthesis- 98
zeolite LTA wt% zeolite LTA
Figure 42. Heating profile of vessel in heated incubator at 80 oC. Aluminosilicate gel
conditions: SiO2/Al2O3 = 2; Na2O/SiO2 = 2.5, H2O/ Na2O = 42

For both zeolite LTA products, the amount of crystalline zeolite LTA was estimated

to be 75 and 98 wt% for incubator and Berghof reactors, respectively. Ju et al., ,

(2006) examined the variation in characteristics when zeolite LTA was made in either

a conventional teflon lined, a stainless-steel reactor vessel, or a continuous

microreactor. In terms of zeolite particle size, the smallest particles were obtained

when using the microchannel reactor. Deneyer et al. (2020) discussed the mechanism

for zeolite synthesis wherein at lower reaction temperatures the number of nuclei

present was relatively high.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 152
In turn this situation led to formation of small zeolite particles. Conversely, at higher

reaction temperatures where fewer nuclei were generated the average particle size was

increased. The data shown in Figure 42 was consistent with the interpretation of

Deneyer et al. (2020) in that the temperature in the incubated samples was in general

lower that the corresponding temperature in the Berghof hydrothermal reactor.

Impact of Pre-Heating Reactants Prior to Zeolite Synthesis


As outlined in the previous section, various types of hydrothermal vessels have been

used to synthesise zeolites. This situation is inherently problematic in terms of

comparing results from different research groups. One approach to reducing the

problem of differing reactor geometry was to pre-heat the feed solutions prior to

introduction to the hydrothermal reactor (Kettinger et al., 1979). Consequently, both

sodium aluminate and sodium silicate solutions were heated to 80 oC prior to mixing

in the reactor vessel (in this case, the heated incubator). Figure 43 shows that using

H2O/Na2O = 15 resulted in some zeolite LTA formation but also cancrinite and

sodalite contaminants (Table 20). The corresponding SEM image also suggested that

a lack of zeolite LTA was present as minimal “cubes” were observed (Figure 43).

In addition, the SEM image displayed a heterogeneous material which presumably was

a mixture of amorphous matter, cancrinite and sodalite. This result contrasted with the

64.1 to 20.4 wt % zeolite LTA identified in Table 17 after 1 to five hours of reaction

when the feed solutions were not pre-heated, respectively. Pre-heating the aluminate

and silicate solutions accelerated the formation of sodalite, cancrinite and non-

diffracting material at the expense of the metastable zeolite LTA. Hence, use of low

values of H2O/Na2O in the reactant gel combined with pre-heating dictated that the

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 153
zeolite synthesis reaction needed to be terminated rapidly. This conclusion was

supported by industry practice where zeolite can be synthesised in only 5 minutes (

Kettinger et al., 1979).

For the zeolite synthesised with H 2O/Na2O = 25, pre-heating of the precursor solutions

promoted the formation of zeolite LTA (c.f. Table 17 to Table 20). The amount of

crystalline LTA was in the range of 73 to 80 wt% when the precursor solutions were

not pre-heated, whereas when the solutions were pre-heated to 80 oC the amount of

crystalline zeolite LTA was in the range 79 to 90 wt %. There was also no evidence

of zeolite LTA degradation after six hours of reaction for the pre-heated samples

(which contrasted with the data in Table 17). Increasing the H2O/Na2O ratio in the

precursor material to 35 along with synthesising zeolite using pre-heated solutions was

conducted. It was noted that the after two hours reaction for the pre-heated sample the

amount of crystalline zeolite LTA was only 11.44 wt % (Table 20) which contrasted

with the corresponding material made with no preheating (Table 17) where zeolite

LTA was 54.39 wt%. Even with longer reaction times the value for crystalline zeolite

LTA was always less than the value for non-preheated samples. In general, these

results agreed with Pina et al, (2004), who pre-heated solutions between 60 and 65 oC

to produce greater amounts of crystalline zeolites.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 154
100
Zeolite LTA
80
Non-diffracting

wt% Crystalline 60

40

20

0
1 2 3 4 5 6
Time (hrs)

H2 O/Na2 O = 15 H2 O/Na2 O = 15; 46% Zeolite LTA

100

80
wt% Crystalline

60
Zeolite LTA
40 Non-diffracting

20

0
1 2 3 4 5 6
Time (hrs)

H2 O/Na2 O = 25 H2 O/Na2 O = 25; 86% Zeolite LTA

100

80
wt% Crystalline

60
Zeolite LTA
40 Non-diffracting

20

0
1 2 3 4 5 6
Time (hrs)

H2 O/ Na2 O = 35 H2 O/Na2 O = 35; 44% Zeolite LTA

Figure 43. Quantitative XRD patterns for Zeolite LTA: SiO2/Al2O3 = 2; Na2O/SiO2 =
2.5; H2O/SiO2 = 15 to 35; Temperature = 80 ℃, pre-heated samples to 80℃; Reaction
Time = 3 h; Reactant mixture stirred at 100 rpm

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 155
Table 18. Summary of quantitative XRD patterns for Zeolite LTA: SiO 2/Al2O3 = 2;
Na2O/SiO2 = 2.5; H2O/ Na2O = 15 to 35; Temperature = 80 ℃, Agitation 100 rpm,
pre-heated gel at 80℃ before synthesis

H2O/ wt%
Tim wt% wt% wt%
Na2 Zeolite
e (h) Cancrinite Sodalite Amorphous
O LTA
15 1 6.87 5.85 21.73 65.46
15 2 8.10 10.90 44.07 36.80
15 3 6.41 12.36 46.28 34.70
15 4 7.81 12.97 51.16 27.97
15 5 7.43 13.24 47.54 31.64
15 6 7.79 11.05 38.96 42.14

20 1 0.27 0.27 86.59 12.83


20 2 0.82 0.73 81.11 17.28
20 3 8.89 7.73 61.60 21.48
20 4 1.64 2.10 66.23 29.90
20 5 1.79 3.40 63.29 31.50
20 6 1.83 4.07 63.41 30.60

25 1 0.52 0.00 18.57 80.91


25 2 0.00 0.00 78.80 20.99
25 3 0.00 0.06 86.10 13.66
25 4 0.07 0.16 79.28 20.21
25 5 0.03 0.17 87.48 12.18
25 6 0.06 0.27 89.57 9.88

30 1 0.23 0.06 20.41 79.29


30 2 0.15 0.00 68.65 31.15
30 3 0.00 0.01 90.16 9.59
30 4 0.00 0.00 88.84 10.97
30 5 0.00 0.00 84.44 15.24
30 6 0.02 0.02 87.96 11.74

35 1 0.09 0.01 22.59 77.30


35 2 0.43 0.00 11.44 88.14
35 3 0.13 0.00 43.67 56.15
35 4 0.00 0.00 72.83 26.68
35 5 0.00 0.00 84.59 14.74
35 6 0.00 0.00 84.66 14.96

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 156
Although it was evident that a complex relationship existed among the various

synthesis variables, machine learning approaches offer a potential means to

comprehend the solution chemistry involved.

Importance of Agitation During Hydrothermal Zeolite Synthesis

Four modes of agitation have been described during the synthesis of zeolites (Deneyer

et al., 2020). The first option is to not stir the reactant mixture. Alternatively,

conventional stirring of the reactor can be achieved by placement of a stirrer shaft

vertically in the hydrothermal synthesis vessel. It is noted that various stirrer

configurations are employed to homogenize solutions (Casci, 2005). In a laboratory

setting reactor vessels can also be shaken or tumbled (Yu et al., 2018). In this study

the incubator employed a shaker plate upon which the sample bottles were placed.

Additionally, the reaction mixture was rotated around the central axis of a reactor

vessel. As indicated by Casci (2005), industry mostly conducts zeolite synthesis by

stirring the reactant mixture.

It was essential to analyse the variations in the collected data while synthesizing zeolite

LTA using pre-heated feed solutions and without stirring. (Figure 44). In this example

the feed conditions were pre-heating to 80 oC; SiO2/Al2O3 = 2; Na2O/SiO2 = 2.5;

H2O/SiO2 = 20 to 30; synthesis temperature = 80 oC; reaction time = three hours; no

stirring of reaction mixture. Comparison of the results obtained with the stirred

samples (Table 20) to the static ones (Table 21), revealed that the samples made using

static conditions produced a similar or higher amount of zeolite LTA. Indeed,

especially after a two-hour synthesis period all the samples consisted of > 90 wt%

crystalline zeolite LTA.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 157
The influence of stirring reactant solutions or leaving them static has often been

reported to be a critical factor in making high quality zeolites. For example, Hagio et

al. (2021) found that use of static synthesis conditions facilitated the growth of EAB-

type zeolite in a matter of hours when precisely controlling reaction temperature. In

contrast, previous efforts to synthesise EAB-type zeolite all required many days to

react when employing agitated reactant mixtures. Alternatively, Kubota et al. (2000)

discovered that only with agitated conditions was it possible to make CFI-type zeolite.

It was also noted that when the H2O/Na2O ratio increased from 25 to 30 the shape of

the crystalline zeolite LTA changed from cubes with sharp edges to rounded cubes,

morphology of zeolite crystals changes depending on the synthesis conditions (Bronić

et al., 2012; Zhan, Li, Zhang, Han, & Chen, 2013). Ideal morphology of zeolite LTA

crystals has been described as a sharp edge cube (Zhan et al., 2013). However, Ayele,

et al. (2016a) reported the importance of using cubes with rounded corners as

detergents because of the incrustation of textiles when zeolite LTA with sharp edges

was used. Y. Liu, et al. (2021) produced zeolite LTA with rounded corners using a

microwave hydrothermal method, and they suggested that the high Ca 2+ ion exchange

(335.6 mg CaCO3 g-1) was due the cubic morphology with rounded edges.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 158
100

80
wt% Crystalline
60
Zeolite LTA
Non-diffracting
40

20

0
1 2 3 4 5
Time (hrs)

H2O/Na2O = 25 98.6% Zeolite LTA

100

80
wt% Crystalline

60
Zeolite LTA
40 Non-diffracting

20

0
1 2 3 4 5
Time (hrs)

H2O/Na2O = 30 98.5 Zeolite LTA

Figure 44. Quantitative XRD data for Zeolite LTA: SiO 2/Al2O3 = 2; Na2O/SiO2 = 2.5;
H2O/SiO2 =25 to 30; Reaction Temperature = 80 ℃; pre-heated samples at 80 ℃;
Reaction Time = 3 h; No stirring

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 159
Table 19. Summary of quantitative XRD patterns for Zeolite LTA with pre-heated and
non-stirring samples: SiO2/Al2O3 = 2; Na2O/SiO2 = 2.5; H2O/ Na2O = 20 to 30;
Temperature = 80 ℃

wt% wt% wt% wt%


H2O/Na2O Time (h)
Cancrinite Sodalite Zeolite LTA Amorphous
25 1 0.05 0.02 55.85 43.98
25 2 0.35 0.04 97.94 1.63
25 3 0.14 0.04 98.59 1.21
25 4 0.34 0.03 96.88 2.69
25 5 0.11 0.06 96.10 3.66

30 1 0.29 0.04 37.60 62.05


30 2 0.09 0.02 90.35 9.48
30 3 0.07 0.05 98.53 1.22
30 4 0.02 0.06 98.67 1.17
30 5 0.00 0.05 96.26 3.60

Influence of Stirring Methodology as a Function of Varying Na2O/SiO2 and


H2O/Na2O Ratio upon Zeolite LTA Synthesis
Based upon the findings in the previous section where the H2O/Na2O ratio was mainly

changed, it was pertinent to determine if the stirring methodology during zeolite LTA

synthesis was dependent on the Na 2O/SiO2 and H2O/Na2O ratios in the gel

composition. In the first instance the H 2O/Na2O ratio was varied from 20 to 30 and the

Na2O/SiO2 ratio was adjusted to= 2 instead of 2.5). The synthesis temperature was

maintained at 80 oC (pre-heated experiments) and the SiO 2/Al2O3 ratio was 2.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 160
100 100

Crystalline Zeolite LTA%

Crystalline Zeolite LTA%


80 80

60 60

40 40

H2O/Na2O=20 H2O/Na2O=20
20 20
H2O/Na2O=25 H2O/Na2O=25
H2O/Na2O=30 H2O/Na2O=30

0 1 2 3 4 0 1 2 3 4
Time (hrs) Time (hrs)

Na2O/SiO2=2 - Stirring Na2O/SiO2 = 2- No-Stirring

Figure 45. Quantitative XRD patterns for Zeolite LTA: SiO 2/Al2O3 = 2; Na2O/SiO2 =
2; H2O/ SiO2 = 20 to 30; Temperature = 80 ℃ - pre-heated samples at 80℃

In terms of the solid products formed the only two phases identified were zeolite LTA

and amorphous/non-diffracting material (Table 20). The impact of the water content

in the zeolite made using Na 2O/SiO2 = 2 was significant, regardless of stirring or not

stirring the feed mixture (Figure 45). When H2O/Na2O = 20 the amount of zeolite

LTA decreased as the reaction time was increased from the optimal value of one hour.

This situation was presumably representative of the metastable state of zeolite LTA

under the experimental conditions. In this regard, the stirred solution promoted the

degradation of zeolite LTA compared to the non-stirred solution. Use of stirred and

non-stirred samples when H2O/SiO2 = 25 produced distinct changes in the rate of

zeolite formation. After one hour synthesis at 80 oC the non-stirred reactor produced

a product which was 85 wt% zeolite LTA [Figure 7]. In contrast, the stirred reactant

solution resulted in a significant inhibition of zeolite growth (only 20 wt % zeolite

LTA) due to the solution agitation.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 161
Table 20. Summary of quantitative XRD patterns for Zeolite LTA with pre-heated
samples and non-stirring: SiO2/Al2O3 = 2; Na2O/SiO2 = 2; H2O/ SiO2 = 20 to 30;
Temperature = 80 ℃
wt%
Synthesis Stirring wt% wt% wt%
H2 O/Na2 O Zeolite
Time (h) (rpm) Cancrinite Sodalite Amorphous
LTA
20 1 100 0.08 0.07 87.42 12.35
20 2 100 0.33 0.55 81.35 17.62
20 3 100 0.42 0.97 75.01 23.55
20 4 100 0.61 1.38 66.62 31.27

25 1 100 0.28 0.00 19.98 79.64


25 2 100 0.00 0.00 86.81 13.09
25 3 100 0.00 0.00 82.23 14.55
25 4 100 0.02 0.01 85.66 14.15

30 1 100 0.33 0.00 5.68 93.96


30 2 100 0.55 0.00 30.25 69.15
30 3 100 0.00 0.01 94.37 5.36
30 4 100 0.00 0.04 96.47 3.22

20 1 0 0.02 0.05 90.31 9.54


20 2 0 0.11 0.05 94.96 4.81
20 3 0 0.12 0.73 84.59 14.40
20 4 0 0.25 1.79 69.63 28.14

25 1 0 0.00 0.03 85.08 14.64


25 2 0 0.00 0.05 93.90 5.95
25 3 0 0.00 0.03 91.05 8.76
25 4 0 0.07 0.00 95.83 3.95

30 1 0 0.54 0.00 15.31 83.97


30 2 0 0.00 0.06 39.86 60.08
30 3 0 0.18 0.09 100.00 0.00
30 4 0 0.07 0.04 97.10 2.65

Samples with H2O/Na2O = 30 needed at least three hours to reach 94 wt% and 100

wt% zeolite LTA crystals for stirring and static samples, respectively. Whether stirred

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 162
or non-stirred solutions were employed, the zeolite LTA content was highest for the

H2O: SiO2 ratio was 30.

Figure 46 shows SEM images between stirred and non-stirred samples when the ratio

of H2O/SiO2 was 30. In the initial stages of reaction, the product was clearly a mixture

of non-diffracting and crystalline material (in accord with Table 20). After three hours

of reaction both the stirred and non-stirred samples exhibited the dominant presence

of crystalline zeolite LTA (Figure 46). When stirring was applied during the reaction

to make zeolites the size and shape of the stirred sample was evident.

Indeed, the zeolite crystals were not only larger but also formed of sharp edges. This

observation was in agreement with Andaç et al. (2005) and Palcic et al., (2012), who

mentioned the benefit of agitation for the formation of zeolite phases and crystal

growth. In contrast, when the synthesis mixture was not stirred the zeolite LTA was

smaller in size and comprised of rounded edges on the crystals. When comparing the

data in Table 19 with Table 20, the reduction of Na2O/SiO2 to 2 appeared to mainly

inhibit the growth of zeolite LTA at short reaction times of one to two hours.

Particle Size Distribution


When used as a detergent builder several physical properties of zeolite LTA are

important. The particle size distribution is critical in relation to zeolite LTA

effectiveness to soften water. If the zeolite crystals are > 10 microns then the rate of

calcium/magnesium ion exchange is limited by diffusion considerations (Pastorello &

Troglia, 1987). In addition, these larger zeolite grains may deposit on clothing and/or

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 163
cause problems with blocking of pipes in laundry machines. Alternatively, zeolite

LTA with a particle size < 1 micron may promote matting of clothing material due to

excessive permeation into the fabric.

Figure 47 shows the light scattering data for the zeolite LTA samples illustrated in

Table 20 and Figure 46. In general, two features were present, namely a broad

distribution assigned to amorphous and/or agglomerated material, or a sharp profile at

lower size classes which presumably corresponded to the formation of crystalline

zeolite LTA. Development of zeolite LTA crystals was promoted using static

solutions in agreement with the data in Figure 45 and Table 20. The broadening

observed even with the samples with the highest crystallinity suggested that the zeolite

LTA crystals were aggregated. For example, Table 21 shows that for the zeolite LTA

made using H2O/Na2O =30 and no stirring for four hours, that 90 % of the particles

were less than 11.2 microns and 50 % were less than 5.4 microns. In relation to its use

as a detergent, in the zeolite LTA only 10 % of zeolite crystals were less than one

micron in size.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 164
STIRRING 100 rpm NO-STIRRING

1h Synthesis - 5.7% Zeolite LTA 1h Synthesis -15.3% Zeolite LTA

2h Synthesis - 30.3% Zeolite LTA 2h Synthesis - 39.9% Zeolite LTA

3h Synthesis - 94.4% Zeolite LTA 3h Synthesis - 100% Zeolite LTA

4h Synthesis - 96.5% Zeolite LTA 4h Synthesis - 97.1% Zeolite LTA


Figure 46. Formation of zeolite LTA crystal over time at 80 oC with preheated samples;
H2O/Na = 30, Na 2O/SiO2 = 2, SiO2/Al2O3 = 2

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 165
8 8
1h Synthesis 1h Synthesis
2hrs Synthesis 2hrs Synthesis
3hrs Synthesis 3hrs Synthesis
4hrs Synthesis 4hrs Synthesis
Volume Density (%)

Volume Density (%)


6 6

4 4

2 2

0
0
0 20 40 60 80 100 120
0 20 40 60 80 100 120
Size Classes (um)
Size Classes (um)

H2O/Na2O=20 - 100 rpm H2O/Na2O=20 – 0 rpm

8 8
1h Synthesis 1h Synthesis
2hrs Synthesis 2hrs Synthesis
3hrs Synthesis 3hrs Synthesis
4hrs Synthesis 4hrs Synthesis
Volume Density (%)
Volume Density (%)

6 6

4 4

2 2

0
0
0 20 40 60 80 100 120
0 20 40 60 80 100 120
Size Classes (um)
Size Classes (um)
H2O/Na2O=25 – 100 rpm H2O/Na2O=25 – 0 rpm

8 8
1h Synthesis
1h Synthesis 2hrs Synthesis
2hrs Synthesis 3hrs Synthesis
3hrs Synthesis 4hrs Synthesis
Volume Density (%)

4hrs Synthesis 6
Volume Density (%)

4
4

2
2

0
0 0 20 40 60 80 100 120
0 20 40 60 80 100 120
Size Classes (um)
Size Classes (um)
H2O/Na2O=30 – 100 rpm H2O/Na2O=30 – 0 rpm
Figure 47. Particle Sizing for zeolite LTA: SiO 2/Al2O3 = 2; Na2O/SiO2 = 2; H2O/Na2O
= 20 to 30; temperature = 80 ℃ - pre-heated samples at 80 ℃

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 166
Table 21. Summary of particle sizing analysis for Zeolite LTA with pre-heated
samples and no-stirring: SiO2/Al2O3 = 2; Na2O/SiO2 = 2; H2O/ SiO2 = 20 to 30;
Temperature = 80 ℃
Stirring Reaction Time Particle Sizing Volume Density
H2 O/Na2 O
(RPM) (hours) Dx (10) Dx (50) Dx (90)
20 100 1 0.69 5.89 26.7
20 100 2 0.91 8.80 33.7
20 100 3 0.84 9.84 38.2
20 100 4 1.47 14.1 46.7
25 100 1 7.26 30.9 74.6
25 100 2 1.05 10.4 42.6
25 100 3 2.38 13.7 39.5
25 100 4 0.86 6.76 21.1
30 100 1 10.1 26.6 56.3
30 100 2 8.29 28.6 62.0
30 100 3 1.28 7.31 25.9
30 100 4 1.01 5.18 15.8
20 0 1 0.63 4.50 33.8
20 0 2 0.66 6.61 35.2
20 0 3 0.94 14.7 57.7
20 0 4 1.05 8.99 37.2
25 0 1 1.17 11.7 46.8
25 0 2 0.98 7.43 30.1
25 0 3 0.85 6.56 32.1
25 0 4 0.87 4.02 11.2
30 0 1 10.3 28.6 61.2
30 0 2 9.49 29.7 68.7
30 0 3 0.72 3.24 16.7
30 0 4 1.02 4.52 11.2

Therefore, the inclusion of zeolite particle sizing data appears to be amenable to

machine learning approaches if a range of particle sixing volume density values are

employed.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 167
Machine Learning Insights

In terms of machine learning data sets, one approach involves automated extraction of

relevant data from published journals (Kim et al., 2017). The fundamental strategy is

to convert human readable text to machine-readable text using a combination of trained

machine learning and natural language extraction algorithms. When applied to

relevant articles a database is compiled which can then be filtered to give materials

synthesis guidance. For example, Jensen et al. (2019) extracted data relevant to

germanium containing zeolites. The machine learning approach successfully

illustrated the relationship between zeolite synthesis parameters and the framework

density of germanium containing zeolites. However, it was also evident that there was

considerable scatter in the data points mined from a range of publications. This

observation is supported by the results of this study which has revealed the difficulties

in trying to use data from other research organizations.

Thus, when trying to optimize zeolite synthesis using machine learning methods an

inherent challenge is the need to work with small data sets (Y. Zhang & Ling, 2018).

Automated extraction of data from published sources can significantly expand the data

set size but the downside is that errors in the data can occur due to all the reasons

illustrated in this investigation. Moreover, collection of all the necessary data to

improve the machine learning model is often not performed by the researchers.

Adjiman et al. (2021) highlighted the need to integrate process engineering concepts

in the macroscale to fundamental science at the nanoscale. This study for example

examined the macroscale impact of stirring or not stirring solutions. Despite the

favourable synthesis of zeolite LTA with no stirring during hydrothermal treatment

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 168
process, constraints evolve on the macroscale in terms of keeping the vessel contents

homogeneous in character.

Application of the data in this study to a machine learning investigation of zeolite LTA

synthesis has subsequently been shown to be informative (Conroy et al., 2022). The

data set was interrogated by a range of machine learning algorithms from the basic to

more advanced: linear regression, ridge regression, regression tree, random forest,

XGBoost and artificial neural network (ANN) models. Simpler models were

characterized by an R2 value of approximately 0.5 whereas tree-based models

increased R2 to between 0,6 and 0.7 R2. However, use of an ANN model was highly

successful as shown by the R2 value of 0.84. Although the data suggested that a deep

neural network (DNN) approach may further improve the fitting accuracy, the

learnings from the small data set agreed with known chemistry of zeolite LTA

synthesis.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 169
Conclusions

This study has shown that it is challenging to create data sets for synthesis of zeolite

LTA which are suitable for machine learning. A critical method for zeolite

characterization is XRD yet the number of publications using quantitative XRD

methods is very limited. It is acknowledged that a greater amount of effort is required

to obtain quantitative XRD patterns due to the need for an internal standard and more

complex data processing. However, data science methods need comprehensive

structural information if credible predictions are to be made.

Mass balance of processes is fundamental to the chemistry of zeolite synthesis and this

approach ultimately provides more extensive data to interpret. The fate of unreacted

species such as sodium hydroxide is not only important for process economics but also

for promotion of subsequent zeolite synthesis.

A key hindrance to data uniformity is the influence of the dimensions and design of

hydrothermal vessels upon zeolite synthesis. Researchers need to be aware that

heating rates, sample temperature profiles (lower/high temperatures than targeted) and

cooling rates all vary between reactor types. This fact makes it difficult to compare

data with that reported in other publications by different research teams.

Pre-heating of reactant solutions is recommended to minimize the variation in reaction

temperature upon zeolite synthesis. Although reported in industry this suggestion has

not been widely adopted in the research literature.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 170
No agitation of the reactant mixture under hydrothermal synthesis conditions was

shown to increase the quality of zeolite LTA. However, when scaling-up zeolite

production it is usually necessary to stir the large vessels involved to enhance

homogeneity of the solutions or slurry.

Subsequent application of machine learning principles validated the usefulness of even

a small dataset in predicting zeolite synthesis outcomes. Therefore, it is recommended

that researchers should capture more data during zeolite synthesis and follow the

lessons from this study.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 171
Key Insights

This chapter has emphasized the challenges involved in creating suitable datasets for

machine learning in the synthesis of zeolite LTA and proposes solutions to address

these challenges. The limited use of quantitative X-ray diffraction (XRD) methods in

published literature, the lack of reporting of the mass balance of processes, and the

impact of vessel dimensions and design on the synthesis process are identified as

critical factors that need to be considered. The chapter recommends pre-heating of

reactant solutions to minimize variation in reaction temperature and avoiding agitation

of the reactant mixture to improve the quality of zeolite LTA. Comparing zeolite

synthesis parameters between laboratories is inherently difficult, and using web

crawling approaches to mine data is impractical due to bias in the collected

information. However, addressing these issues can lead to meaningful results in a

researcher's laboratory.

The next chapter will test the dataset created in this chapter using a progressive

machine learning methodology. The aim is to identify the relationship between zeolite

synthesis descriptors and evaluate the potential for machine learning to predict the

quantitative output of synthesis routes. The hypothesis is that by applying statistics

and machine learning principles, it may be possible to enable pre-evaluation and

increase zeolite yield and performance. The chapter will apply various machine

learning algorithms to zeolite LTA synthesis data, including linear regression, ridge

regression, regression tree, random forest, XGBoost, and artificial neural network

models.

Chapter 6: Factors which Influence Development of Robust Data Sets for Machine Learning and
Statistical Analysis of Zeolite Synthesis 172
Chapter 7: Evaluation and application of machine learning

principles to Zeolite LTA synthesis

Bethany Conroy1, Richi Nayak2, Andrea Lucia Hidalgo Rocha 1, and Graeme J.

Millar1

1. School of Mechanical, Medical & Process Engineering, Science and

Engineering Faculty, Queensland University of Technology (QUT),

Brisbane, Queensland, Australia.

2. School of Computer Science, Faculty of Science, Queensland University of

Technology (QUT), Brisbane, Queensland, 4000, Australia

* Corresponding author. E-mail: graeme.millar@qut.edu.au

Published (2022): Microporous and Mesoporous Materials, Volume 335.


DOI: https://doi.org/10.1016/j.micromeso.2022.111802

This chapter is a copy of the journal paper published.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
173
Statement of Contribution of Co-Authors for Thesis by Published Papers

The authors listed above have certified that:

They meet the criteria for authorship in that they have participated in the conception,

execution, or interpretation, of at least that part of the publication in their field of

expertise.

They take public responsibility for their part of the publication, except for the

responsible author who accepts overall responsibility for the publication.

There are no other authors of the publication according to these criteria.

Potential conflicts of interest have been disclosed to (a) granting bodies, (b) the editor

or publisher of journals or other publications, and (c) the head of the responsible

academic unit, and

They agree to the use of the publication in the student’s thesis and its publication on

the QUT’s ePrints site consistent with any limitations set by publisher requirements.

Title of the publication:

Evaluation and application of machine learning principles to Zeolite LTA

synthesis

Status of the publication: Published

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
174
Abstract

A progressive machine learning methodology was utilised to not only identify the

relationship between zeolite synthesis descriptors but also evaluate the potential for

machine learning to predict the quantitative output of synthesis routes. The hypothesis

was if statistics and machine learning principles are applied, then it may enable pre-

evaluation and result in potential increases to zeolite yield and performance. Various

machine learning algorithms were applied to zeolite LTA synthesis data, including

linear regression, ridge regression, regression tree, random forest, XGBoost and

artificial neural network models. Major findings included the use of input synthesis

variables and the product yield for model training. Additionally, the use of both

quantitative and qualitative X-ray diffraction (XRD) data was required to accurately

determine product composition ("hybrid XRD" approach). Models, including linear

regression, ridge regression and regression trees, returned R2 values less than 0.5

indicating the complexity inherent with the problem. Embedded tree-based models,

including random forest and XGBoost, resulted in testing accuracies equal to R2

=0.620 and R2 = 0.700, respectively. An ANN model achieved the highest accuracy

among all machine learning algorithms of R2 = 0.84. Notably, this model was the most

accurate because it exploited non-linear and complex relationships within a

multidimensional and intercorrelated dataset such as that obtained from zeolite

synthesis. Despite reaching an accuracy greater than 80%, the ANN model accuracy

continued to increase by increasing the network size, indicating that advanced deep

learning models should be considered as part of future work.

Keywords: Zeolite Synthesis; Artificial Neural Network; Data; Machine Learning;


Principal Component Analysis.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
175
Introduction

Zeolites are used in many industrial applications, including petroleum refining

(Alkhlel & de Lasa, 2018), water treatment ( Millar, Couperthwaite, et al., 2016; Wen

et al., 2018), gas adsorption (Poursaeidesfahani et al., 2019), agriculture (Nakhli, et

al., 2017), animal feed additives (Valpotic, 2017) and green chemistry (Y. Li, Li, &

Yu, 2017; Lima et al., 2019). At present, over 200 natural and synthetic zeolite

structures have been recorded, with zeolite LTA, X, USY, mordenite, Beta and ZSM-

5 being the most widely used (Collins, et al. 2020). The synthetic zeolite market is

projected to grow by 2.6 % between 2018 and 2023 to reach a value of USD 5.9 million

per annum (Y. Dong, Lin, & He, 2017).

Numerous factors impact the zeolite synthesis mechanism, product quality, yield, and

performance (Cundy & Cox, 2005). Thus, zeolite synthesis is an increasingly

complicated, multidimensional problem (Moliner et al., 2019). Machine learning is

of increasing interest to the zeolite community as it employs mathematical models that

learn from past data to extract patterns and intervariable relationships within a dataset

(Figure 48). The ability to map intricate behaviours (semi-)autonomously using

machine learning enables the materials science field to progress in the absence of

comprehensive knowledge of advanced mechanisms (Moliner et al., 2019).

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
176
Figure 48: Histogram showing citations of publications with the keywords "machine
learning AND zeolite", "neural network AND zeolite", and "principal component
analysis AND zeolite".

Current research relating to machine learning and zeolite science includes

characterisation and structure discovery for petrochemical cracking and gas adsorption

(Xiao et al., 2020). Another example is the study by Jensen et al., (2019) which

utilised data mining techniques to extract and compile a database of zeolite synthesis

routes. Zeolite literature has often used machine learning to predict the ideal physical

characteristics of zeolites but has not yet predicted the synthesis routes required to

obtain these characteristics (Jensen et al., 2019; Moliner et al., 2019; Muraoka, Sada,

Miyazaki, Chaikittisilp, & Okubo, 2019; Xiao et al., 2020). Studies have used

machine learning algorithms to understand relationships between the synthesis

descriptors to predict the qualitative rather than the quantitative output (Jensen et al.,

2019; Moliner et al., 2019; Muraoka et al., 2019; Xiao et al., 2020).

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
177
Understanding the data, intervariable relationships, and correlations is paramount in

multidimensional problems. Principal Component Analysis (PCA) is an unsupervised

machine learning technique that utilises statistical methodologies to reduce

dimensionality (Jolliffe, 2002). PCA identifies variables of high and low influence in

a dataset through projection analysis, allowing the removal of redundant features to

improve learning efficiency and reduce computational cost (Jolliffe, 2002). For

example, Conterosito et al. (2020) utilised PCA and correlation analysis on parameters

obtained from Rietveld refinement to heighten relationships present in the CO 2/Xe

sorption data on Zeolite Y .

For systems modelled using regression machine learning, the use of regression trees

and random forest models to predict the physical properties of synthesised zeolites has

been documented. Jensen et al. (2019) used literature extracted data to develop a

random forest regression model that predicted the zeolite framework density from

features including the Silicon, Germanium, Aluminium, Hydroxide, and water content,

as well as reaction time and temperature. The study found the model accuracy to be

R 2 = 0.7 (Jensen et al., 2019). Alternatively, (Muraoka et al. (2019) trained

classification models to predict the resultant zeolite using the same dataset, including

the framework density. The results showed that the Extreme Gradient Boosting

(XGBoost) model outperformed all other models with a 75 to 80 % test accuracy

(Muraoka et al., 2019). In contrast, Moliner et al. (2019) utilised Artificial Neural

Networks (ANN) to predict internal relationships between the zeolite synthesis

descriptors. However, each publication lacks transparency when reporting how these

models have been constructed, trained, and tested to obtain these results.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
178
Artificial Neural Networks (ANN) (Figure 49) are non-linear systems that model

complex multidimensional relationships through a series of nodes and connections

reminiscent of a biological brain (Moliner et al., 2019). While ANNs have been used

previously in zeolite science, as shown by Moliner et al. (2019), few studies have

ventured further in model complexity, for example, applying advanced deep learning

techniques. Deep Neural Networks (DNN) provide a greater ability to solve complex

problems and identify hidden patterns and relationships within multivariate datasets

using multiple network layers between the input and output (Figure 49). While the

tools required to implement DNN are available, the lack of sufficient data in zeolite

synthesis applications raises the question of whether DNN should be applied.

Figure 49. Visualisation of the different neural network structures (ANN left, DNN
right).

Zeolite LTA (Na12 [(AlO2 )12 (SiO2 )12 ] ⋅ 27H2 O is widely used in industry (Collins et

al., 2020) and obtained by hydrothermal synthesis of aluminosilicate gels (Cundy &

Cox, 2005).

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
179
Fundamental synthesis parameters which impact the formation of zeolite LTA include

the SiO2 /Al2 O3, Na2 O/SiO2 , H2 O/Na2 O molar ratio, reaction time and temperature

(De Silva et al., 2007). In addition, a range of operational factors impact the plant

performance and physical characteristics of zeolite LTA such as: identity of

silica/aluminium sources; pre-heating of reactants; order/rate of mixing of reactants;

aging of the gel; vessel configuration/size; heating ramp rate to attain reaction

temperature; impact of mother liquor recycling; cooling rate of zeolite sludge product;

and washing stage to remove unreacted species and reduce solution pH (Collins et al.,

2020). Ultimately, the best synthesis approach and associated manufacturing plant

design should produce zeolites of the highest quality (purity; particle size), greatest

amount (yield), and at the lowest cost (minimise utility and chemical consumption;

optimal plant design). Environmental aspects are also important, especially

wastewater production in the zeolite synthesis process.

The complexity of zeolite LTA synthesis combined with the substantial literature on

this topic suggests that it may be an ideal test case to demonstrate the methodology,

outcomes, and benefits of employing a machine learning approach. At present, there

are some excellent publications that have used machine learning to identify synthesis

conditions to make zeolites (Moliner et al., 2019). Yet, the relationship between

synthesis descriptors and the product zeolite remains a challenge ( Muraoka et al.,

2019). Therefore, the aim of this study was to apply machine learning principles to

zeolite LTA synthesis to enable the pre-evaluation and testing of zeolite synthesis

routes. A key outcome was improved approaches to model development and training,

which enable the development of high performing machine learning models.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
180
The hypothesis was: "If machine learning principles can be applied to zeolite LTA

synthesis, then it may be used to predict the success of laboratory synthesis routes

enabling pre-evaluation and potential increases to zeolite yield and performance of

zeolite products." This study addressed the following research questions to support this

hypothesis: (1) Is the zeolite synthesis data sufficiently robust for machine learning

applications? (2) Which experimental variables should be included when developing

machine learning algorithms? (3) Can a feature transformation technique simplify the

data, or are the transformed factors too intercorrelated? (4) What is the preferred

machine learning approach which can be most accurately applied to zeolite LTA

synthesis? In terms of methodology, a dataset comprising of 25 features and 130

instances was generated at laboratory scale for zeolite LTA synthesis.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
181
Materials and Methods

Machine Learning
The goal of machine learning is to predict the output of the zeolite synthesis system

using the input attributes. The application of machine learning to zeolite LTA synthesis

followed the knowledge discovery process depicted in Figure 50.

Figure 50. Knowledge discovery process.

The application of machine learning used a development pipeline, which followed the

logical progression and application of varying machine learning techniques according

to complexity from PCA, linear regression, tree-based models to ANN. PCA is an

unsupervised machine learning technique. All other techniques applied to the data

formed part of a supervised machine learning approach governed by the relationship

f (x) = y in a continuous space. In f(x) = y, x is a vector populated with input values

corresponding to those of the zeolite synthesis system, and y is the continuous output

of the system, the zeolite yield. The underlying machine learning task is learning the

function f according to the algorithm used based on the training data in the form of

(x, y).

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
182
Progression of Machine Learning Technique Application
A progressive approach to applying machine learning models to the zeolite synthesis

data was taken to better understand the complexity of the zeolite synthesis system. As

a result, various machine learning models were developed following the logical

progression and application of machine learning techniques according to complexity,

as shown in Figure 51.

PCA Linear Regression Regression Tree & Artificial Neural


Random Forest Networks

• Exploratory • Basic machine • Produce a • Enable hidden


statistics and learning models decison tree patterns and
analysis used to based correlations to
technique. determine the predictive be identified.
• Used to baseline error. model. • Ability to train
understand • Enables • The models use and re-train
correlations assessment of cross-validation models. The
and data suitability to optimise models
distributions for further parameters and continuously
present in the regression avoid learn and
data. machine overfitting. improve.
learning. • Application of
gradient
boosting using
XGBoost.

Figure 51. Progression and application of machine learning techniques according to


complexity.

Utilising a progressive approach enabled observations about the progression of

machine learning accuracy with model complexity to be quantified. It is conjectured

that if the machine learning accuracy no longer increases with increasing model

complexity, then at this point, the complexity of the machine learning model and the

zeolite synthesis problem are equal. Significantly, the transparency of each machine

learning model reduces with increasing complexity. Leading up to ANN, the ability

to view model parameters and understand the feature importance of each model greatly

reduces.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
183
Principal Component Analysis
PCA uses the intercorrelation of variables to find a new set of uncorrelated variables

with reduced dimensionality termed principal components (PC) (Corma et al.,2005).

PCA was used to achieve a lower-dimensional representation of the zeolite synthesis

data while preserving relationships and correlations present in the original dataset with

minimal information loss (Jolliffe, 2002).

PCA was applied to gain a greater understanding of multidimensional data by

visualising it in a space of reduced dimensionality. PCA was performed on

standardised data to ensure scaling differences between variables did not influence

observed patterns and trends. The PCA algorithm was implemented by following the

equations derived in the supplementary documentation. MATLAB functions were

used to implement PCA, including standardising data and calculating the singular

value decomposition (SVD). In addition, MATLAB was used to visualise the variance

explained, PC coefficients and projection onto the PC.

Initialising the Data, Training and Test Sets


Before implementing and training the machine learning models, the zeolite synthesis

data was imported into Python. The data was split into training and test datasets using

the scikit-learn (sklearn) package, ensuring the models were always tested and

evaluated on unseen data (Pedregosa et al., 2011). The data was split such that 80 %

was used for training and validation and 20 % for testing. Continuous features in the

dataset were standardised prior to training the models.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
184
Linear, Ridge, Regression Tree , and Random Forest Regression Models
The sklearn package was used to implement the linear regression, ridge regression,

regression tree and random forest models in Python (Pedregosa et al., 2011). Ten-fold

cross-validation was used to train the models and optimise the hyperparameters

through a grid search of the candidate values presented in the supplementary

documentation. The models were trained using the sklearn fit estimator. For ridge

regression, cross-validation was applied to obtain the optimal alpha parameters. The

random forest regressor utilised bagging to create multiple regression trees combined

to give the random forest.

XGBoost
The XGBoost package was used to implement an XGBoost regression model in Python

(T. Chen & Guestrin). Ten-fold cross-validation was used to train the model and

optimise the hyperparameters through a grid search of the candidate values presented

in the supplementary documentation. The model was trained using the fit estimator.

Artificial Neural Network


The ANN models have been implemented in two ways using sklearn and PyTorch.

One model utilised the multilayer perceptron (MLP) regressor through the sklearn

neural network package (Pedregosa et al., 2011). Another model implemented ANN

using Skorch and the Adam optimiser from PyTorch (Paszke et al., 2019), assisted by

various sklearn functions from the model selection and pre-processing packages

(Pedregosa et al., 2011). The MLP regressor was implemented using ten-fold cross-

validation to train the model and optimise the hyperparameters through a grid search

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
185
of the candidate values presented in the supplementary documentation. The network

was trained using the sklearn fit estimator by minimising the cost function.

The manual implementation of ANN in PyTorch provided greater control over the

model architecture. The model utilised 10-fold cross-validation during training,

introducing a third data split for validation. The model was trained and optimised

using the Adam optimiser from PyTorch. For each iteration within a cross-validation

fold, the model carried out forward propagation training by calculating the training and

validation losses and checked for convergence to the tolerance value as part of the

early stopping criteria. Finally, the model applied backpropagation to optimise the

model weights. At the end of cross-validation, the best model and associated weights

were returned for prediction and testing of the neural network.

Approaches to Model Evaluation


After training, the predict method from sklearn was utilised to determine the predicted

feature values ŷ. The true and predicted values were compared to evaluate the model

accuracy. All models were evaluated using the coefficient of determination,

R 2Through the sklearn metrics package (Pedregosa et al., 2011). The R 2 value

represented the proportion of variance from the original data explained by the

independent variables predicted by the machine learning model. The R 2 metric

provided an indication of the model's fit to the original data and, therefore, how well

novel features were likely to be predicted by the model through the proportion of

explained variance. The sklearn metrics package r2_score function used Equation 17,

1
where: y̅ = ∑ni=1 yi.
n

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
186
∑𝐧 ̂ 𝐢) 𝟐
𝐢=𝟏( 𝐲𝐢−𝐲
Equation 17: 𝐑 𝟐 (𝐲, 𝐲̂) = 𝟏 − ∑𝐧 ( ̅)𝟐
𝐢=𝟏 𝐲𝐢−𝐲

The models were trained, tested, and evaluated five times using a different train/test

split such that the average R 2 value with a ± tolerance relative to the minimum and

maximum R 2 value was obtained. The R 2 values for each individual training and

testing run are provided in the supplementary documentation.

Input Variables Selection and Zeolite LTA Synthesis Data


A clean dataset with sufficient instances of data was required to train the machine

learning models successfully. A dataset consisting of 25 features and 130 instances

was obtained from the laboratory synthesis of zeolite LTA from aluminosilicate gels.

An initial gel was created and subsequently reacted hydrothermally at 80 oC. Given

the number of features, the data is multidimensional and requires refining before

training. The "hybrid 10" (see Section “Synthesis of Zeolite LTA from Aluminosilicate

Gels”, below) data was used to train the machine learning models to predict the success

of various zeolite LTA synthesis routes as mass yield or wt% of zeolite LTA. To

prevent false predicting variables from dominating the models, all output product

values were removed from the dataset. The variables included in the input data are

displayed in Table 22.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
187
Table 22. Machine learning input variables.
Variable Units Minimum Maximum Average
Water g 16.19 46.39 29.63
Sodium Silicate g 7.45 12.95 11.46
𝑵𝒂𝑶𝑯 g 2.44 4.89 4.43
Water Temperature ℃ 21.00 77.00 48.81
𝑵𝒂𝟐 𝑶/𝑺𝒊𝑶𝟐 Molar 2.00 2.50 2.41
Ratio
𝑯𝟐 𝑶/𝑵𝒂𝟐 𝑶 Molar 15.00 42.00 26.46
Ratio
Aging Time h 0.00 0.50 0.23
Stirring Speed RPM 0.00 100.00 33.08
Reaction Time h 0.50 6.00 3.06
Zeolite LTA g 0.00 8.93 5.70

Zeolite Characterization

X-Ray Diffraction (XRD)

A Panalytical X'Pert wide-angle X-ray diffractometer equipped with Co Kα radiation

(1.7903 Å) was employed to determine the components of the zeolite product. Every

0.02° XRD traces were recorded over the range of 5 to 90° 2θ (30 seconds per step).

Quantitative XRD analysis was achieved by implementing a 10 wt% corundum

internal standard. Quantitative XRD data was analysed by TOPAS v5 software using

the Rietveld method. Qualitative XRD analysis simply compared the relative intensity

of peaks due to crystalline components present in the zeolite sample.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
188
Scanning Electron Microscopy (SEM)
The presence of zeolite LTA and amorphous material was determined by a JEOL

7001F Scanning Electron Microscope (SEM). Zeolite samples for SEM analysis were

arranged by placement on carbon tape followed by platinum coating. Images were

acquired by using an accelerating voltage of 5 kV.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
189
Results and Discussion

Synthesis of Zeolite LTA from Aluminosilicate Gels


Figure 52 illustrates the general chemistry of zeolite LTA manufacturing and the

impact of synthesis parameters, particularly reaction time. It is noted that the figures

below represent a small subset of the larger database of experimental data used for

machine learning. Using a gel composition of SiO2 /Al2 O3 = 2, Na2 O/SiO2 = 2.5 and

H2 O/Na2 O = 30, zeolite LTA was synthesised at a reaction temperature of 80 ℃

(Figure 52). Quantitative XRD data indicated that there was an initial induction period

where the product was primarily amorphous/unidentified. However, once the reaction

time was increased to two hours, the amount of crystalline zeolite LTA rapidly

increased to approximately 70 wt%. Correspondingly, the amorphous/unidentif ied

content decreased in harmony with the general zeolite growth stages shown in

Supplementary Figure 7

Figure 52. Zeolite synthesis products for gel composition of SiO2/Al2O3= 2,


Na2O/SiO2= 2.5 and H2O/Na2O = 30; reaction temperature 80 oC.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
190
The maximum zeolite LTA content of ca. 70 wt% was significantly less than expected

compared to previous literature, which reported amounts of crystalline zeolite LTA as

high as 99 wt% (Deabriges, 1982). Therefore, the H2 O/SiO2 ratio was varied while

keeping all other parameters constant to determine if the amount of zeolite LTA could

be improved (Figure 53).

Figure 53. Zeolite synthesis products as a function of reaction time for gel composition
of SiO2/Al2O3 = 2, Na2O/SiO2= 2 and H2O/Na2O = 20 to 30; reaction temperature 80
oC.

It was observed that lower H2 O/Na2 O values accelerated the formation of zeolite

LTA, albeit at H2 O/Na2 O = 20, the metastability of zeolite LTA was noted as the

amorphous/unidentified content increased with reaction time. Regardless of the

H2 O/Na2 O molar ratios, the crystalline amount of zeolite LTA did not grow higher

than 75 wt%. Consequently, SEM imaging was performed to investigate the reason

for the limited zeolite LTA formation (Figure 54).

After one hour of synthesis, SEM imaging revealed a mixture of amorphous material

and some zeolite LTA (cubic crystals). This data agreed with the quantitative XRD

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
191
data in that 2.7 wt% zeolite LTA was detected. Increasing the reaction time to two

hours promoted the formation of zeolite LTA as numerous crystals of relatively small

size (< 1 micron), and the quantitative XRD results suggested that the quantity of

crystalline zeolite LTA was 21 wt%. The quantitative XRD results may have under-

reported the zeolite LTA content due to line broadening, which is particularly

prevalent as zeolite particle size decreases (Raic et al., 2020; Strachowski et al., 2021).

In turn, this leads to the Rietveld analysis erroneously predicting the presence of

amorphous and un-identified materials. Accordingly, as shown in Figure 53, optimal

zeolite LTA formation occurred at a reaction time of three hours for a H2 O/Na2 O ratio

of 30. Quantitative XRD suggested that at 3- and 4-hours reaction time, 71 and 72

wt% zeolite LTA was present. However, the presence of significant amounts of

amorphous material was absent in the SEM images.

Consequently, a qualitative XRD approach which is common in the zeolite literature

and does not account for non-diffracting species was applied to the data. Results

indicated a quantity of crystalline zeolite LTA of 99 wt% for the 3- and 4-hour

synthesis times with a H2 O/Na2 O ratio of 30 as shown in Figure 53.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
192
1h synthesis - 2.7 % Zeolite LTA 2h synthesis – 21 % Zeolite LTA

3h synthesis – 71 % Zeolite LTA 4h synthesis – 72 % Zeolite LTA

Figure 54. SEM images of zeolite synthesis products as a function of reaction time
for gel composition of SiO2/Al2O3 = 2, Na2O/SiO2= 2 and H2O/Na2O = 30; reaction
temperature 80 oC.

When applying machine learning, the quality of the input data is important, especially

when using relatively small datasets (Y. Zhang & Ling, 2018). In theory, quantitative

XRD analysis should provide more accurate data for machine learning (Hanson et al.,

2021).

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
193
However, as can be seen from Figure 52, Figure 53 and Figure 54, quantitative and

qualitative XRD analysis were both required to explain the experimental data.

Therefore, a new idea was developed to integrate both quantitative and qualitative

XRD approaches. In the early stages of zeolite LTA synthesis, a quantitative XRD

method is required to account for the presence of amorphous species. On the other

hand, at higher levels of zeolite LTA formation, the use of qualitative XRD seems

appropriate as the presence of amorphous and other non-diffracting material was

expected to be minimal. This new methodology for interpretation of XRD data in

relation to use with machine learning studies has been termed a "hybrid XRD

approach". The remaining challenge is to specify at which point in the zeolite growth

curve quantitative and qualitative XRD methods should be applied to generate the most

accurate dataset. Thus, data analysis was employed in situations where the crystalline

content of zeolite LTA was < 10, < 20 and < 40 wt%. For example, a "hybrid 10"

approach (i.e., crystalline zeolite content < 10 wt%) was initially utilised to provide

data for machine learning techniques to determine the most suitable model for zeolite

LTA synthesis.

Application of Chemometrics to Zeolite LTA Synthesis Data


Following the progression of machine learning technique application, PCA was

conducted and applied to the selected zeolite LTA synthesis data. The correlation

between chosen variables for machine learning is shown in Figure 55.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
194
Figure 55. Correlation of machine learning variables using hierarchical clustering
methods.

From Figure 55, the highest correlation was observed between the following variables:

(1) NaOH and Na2 O/SiO2 with a correlation value of 1; ( Millar, Couperthwaite, et

al., 2016) water and H2 O/Na2 O with a correlation value of 0.95; and water temperature

and aging time with a correlation value of -0.98 (Wen et al., 2018). The relationships

between NaOH, Na2 O/SiO2 and water, H2 O/Na2 O were expected results given the

relationship between input chemicals to the system and the calculated ratios that

influence zeolite synthesis. The similarity of the NaOH and Na2 O/SiO2 correlations in

Figure 55 revealed that one is a redundant variable. Given the importance of the

Na2 O/SiO2 ratio demonstrated in literature (De Silva et al., 2007) NaOH was removed

from the dataset. Other trends of interest included the correlation between the stirring

speed, sodium silicate and aging time.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
195
While there were some clear correlations indicated by correlation values close to one

and negative one, all other variables indicated some level of intercorrelation. Most of

the variables had a correlation value greater than |0.2|. Given the multidimensiona lit y

and collinearity of the zeolite synthesis data, PCA was utilised to obtain a lower-

dimensional representation of the data. This situation enabled simpler data

visualisation to identify trends and develop a greater understanding of the data. The

PCA returned nine components equal to the number of variables in the machine

learning dataset. The importance of each PC was determined by visualising each PC's

variance, as shown in Figure 56a.

(a) (b)

Figure 56. (a) Variance explained by principal components (b) principal component
coefficients.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
196
From Figure 56a, the first five PCs explained greater than 90 % of the variance in the

data. The first four PCs explained greater than 85 %, with PC 1 explaining 40 % of

the variance and PC 2, 3 and 4 explaining 20, 14 and 11 %, respectively. PCs 5 to 10

each explained less than 10 % of the data. This result suggested that only 4 out of 10

PCs significantly influenced the patterns observed in the data. Therefore, PC 1, 2, 3

and 4 were retained for further analysis. The constituents of the first four PC are

visualised in Figure 9b.

Figure 56b showed that each variable contributed at least in part to the first four PCs.

This conclusion affirmed trends observed within the correlation plot and confirmed

the high degree of intercorrelation and multi-collinearity in the data. The coefficients

also confirmed that the variables retained for machine learning all, to a degree,

influenced the result of a zeolite synthesis route.

The relationship between the first four PCs were also visualised in Figure 57 as a

biplot. The biplot displays vectors representing each variable that point away from the

origin to the value assigned by PCA. The plot visualised the values assigned to each

of the PCs for each dataset area. When analysing the biplot, the factors considered

included the length of the vectors and the angle between vectors. The length of the

vector evaluated the representation of the variable by the plot and vice-versa. Longer

vectors indicated a greater representation. Additionally, the size of the angle in

between vectors indicated the correlation. Small angles represented a strong positive

correlation, 90-degree angles indicated no correlation, and 180-degree angles implied

a negative correlation.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
197
Figure 57. Biplot comparison of principal components 1, 2, 3 and 4.

The variables labelled in Figure 57 corresponded to the largest coefficients in the first

four PCs. Visualisation of PC 1 and PC 2 revealed a negative correlation between

water temperature and water mass and no correlation between water temperature and

stirring speed. For zeolite LTA mass, the reaction time and sodium silicate displayed

a near-perfect correlation. Given that the machine learning models predict the zeolite

LTA mass yield, a greater dependency on the input chemicals to the system rather than

the ratios was expected. The amount of water, sodium silicate, and sodium hydroxide

added to the system provided the machine learning model with an indication of the

system's mass balance, which assisted with predicting the resultant zeolite mass yield.

The synthesis routes conducted as part of the data collection have a constant NaAlO2

value of 5 g and SiO2 /Al2 O ratio of 2. In machine learning, variables with constant

values are excluded from the analysis. Therefore, the values of NaAlO2 and

SiO2 /Al2 O variables do not influence the predictions unless they have a varied,

continuous value. Therefore, the machine learning prediction is limited by the sodium

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
198
silicate input synthesis descriptor. Nevertheless, the results from the PCA indicated

that the selected input variables for machine learning were all critical when modelling

the zeolite synthesis system. Furthermore, key relationships identified from past

publications, including reaction time, sodium silicate and the resulting zeolite LTA

mass (Silva et al., 2007), also held for the current dataset, indicating the robustness

and suitability of the data for further machine learning evaluation.

Application of Linear and Ridge Regression to Zeolite LTA Synthesis Data


To evaluate the baseline error of the machine learning, linear and ridge regression

models were trained using the zeolite synthesis data. The results from the training and

testing of the models are displayed numerically in Table 23.

Table 23. Training and testing results for linear and ridge regression.
Model Training 𝐑 𝟐 Testing 𝐑 𝟐

Linear Regression 0.368 ± 0.009 0.339 ± 0.019

Ridge Regression 0.372 ± 0.011 0.340 ± 0.018

The results from Table 23 demonstrated that both the linear and ridge regression

models do not accurately represent the zeolite synthesis system. The training and

testing R 2 values revealed that the models were not overfitting, which must be closely

monitored when modelling systems characterised by small datasets (Y. Zhang & Ling,

2018). Based on these results, it was pertinent to view the feature importance to

identify the variables that the models relied upon to make the predictions.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
199
As shown in Figure 58, the linear and ridge regression results were highly dependent

on the Na2 O/SiO2 ratio. Given the intercorrelation and multi-collinearity observed

between the input synthesis descriptors, the model was expected to depend on multiple

variables from the zeolite synthesis system. However, this result confirmed, along with

the poor model accuracy, that linear and ridge regression models failed to learn the

complex relationship inherent in the zeolite synthesis system.

Figure 58. Linear and ridge regression feature importance from input synthesis
variables.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
200
The accuracy of the linear and ridge regression models (R 2 value) was less than 0.6,

which indicated the need for more advanced machine learning models.

Application of Tree -Based Models to Zeolite LTA Synthesis


More advanced models that can learn non-linear relationships with tree-based

architectures, including regression trees, random forests and XGBoost, were trained

using the zeolite synthesis data.

Table 24. Training and testing results for tree-based models.


Model Training 𝐑 𝟐 Testing 𝐑 𝟐
Regression Tree 0.423 ± 0.012 0.385 ± 0.023

Random Forest 0.686 ± 0.018 0.620 ± 0.019

XGBoost 0.741 ± 0.007 0.700 ± 0.007

As shown in Table 24, as the complexity of the applied machine learning mode l

increased, accuracy improved. The similarity between the training and testing R 2

values confirmed that the models were not overfitting. This result was achieved by

pruning the trees generated in each of the above models, where parts of the tree were

removed to prevent it from growing to its full depth. In tree-based models, overfitting

occurs when the model is designed to perfectly fit all samples in the training dataset

(Praagman, 1985). This situation happens when the tree is allowed to grow without

constraining the number of splits, samples at each tree leaf, and the maximum depth

(Praagman, 1985). The models are most prone to overfitting when a given tree is

particularly deep. Therefore, hyperparameter tuning was conducted.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
201
The models were restricted to depths between 5 and 10, the minimum number of

samples at each split and the minimum number of samples at each leaf as 10 and 5,

respectively. By limiting the tree growth to these parameters, the testing results in

Table 24 were maximised without overfitting the model. The models could achieve

training accuracies of greater than 90%; however, the extent to which the regression

tree, random forest, and XGBoost models were overfitting made those models and

their results invalid. Constraining the tree's growth prevented overfitting;

nevertheless, the accuracy was also significantly reduced.

An example of the regression tree trained for the zeolite synthesis system is displayed

in Supplementary Figure 8. The feature importance is shown in Figure 59. The results

from the regression tree and random forest displayed a greater dependence on a variety

of input zeolite synthesis descriptors. The results were corroborated by trends in the

PCA biplot, which showed the variables with the greatest correlation to zeolite LTA

mass yield as sodium silicate and reaction time (Figure 57).

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
202
Figure 59. Regression tree and random forest feature importance for input synthesis
variables.

The random forest displayed a more balanced feature importance across the dataset.

However, this observation was attributed to the number of estimators in the model.

In addition to applying regression trees and random forests to the zeolite synthesis

data, XGBoost was investigated. XGBoost is a tree-based machine learning technique

that applies gradient boosting to regression trees. The results from the XGBoost model

are shown in Figure 60.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
203
Figure 60. XGBoost testing results using input synthesis variables (results from the
training and testing runs that resulted in the greatest model accuracy).

While the results demonstrated significant improvement from the standard regression

tree model, the XGBoost algorithm did not model the zeolite synthesis system to the

desired accuracy. The balanced feature importance of the XGBoost model confirmed

the multi-collinearity and intercorrelation of all variables within the dataset. The

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
204
results from the feature importance again demonstrated the importance of reaction time

and sodium silicate synthesis descriptors when predicting the output zeolite yield

(Figure 60).

Despite providing a better representation of the entire system, the regression tree,

random forest, and XGBoost models predicted results with reduced accuracy (ranging

between 42 to 74%), confirming the complexity of the zeolite LTA synthesis system.

Application of ANN to Zeolite LTA Synthesis


ANN suffers from significant variation between models due to the dependence on

many weight parameters during learning. Therefore, in this study, two different ANN

models have been developed, one with Python's sklearn package and the other with

the PyTorch package and accompanying hyper-parameter tuning tools from the Skorch

package. The results from the sklearn ANN and PyTorch ANN implementation were

used to validate each other. The sklearn model utilised grid search methodologies for

hyperparameter tuning and implementation of cross-validation during training.

Alternatively, the PyTorch model implemented an ANN model constructed using

PyTorch descriptors. For hyperparameter tuning, the Skorch library was employed,

enabling the use of grid search methodologies. Both the models had the same

architecture and very similar hyperparameters. However, the PyTorch implementation

provided the added advantage of greater control over the model architecture and cross-

validation results.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
205
Given the complexity of the data and ANN techniques combined with the limited

dataset, additional steps were taken to prevent the model from overfitting. These

included limiting the number of layers and neurons in the model, limiting the number

of iterations, controlling the learning rate, and conducting cross-validation. Both

models were constructed with an input layer, two hidden layers and an output layer,

all using the ReLU activation function. The results from training and testing the ANN

models are displayed numerically in Table 25.

Table 25. Training and testing results for ANN models (Ackley, Rege, & Saxena,
2003).
ANN Model Training 𝑹𝟐 Testing 𝑹𝟐

sklearn 0.887 ± 0.010 0.840 ± 0.016

PyTorch 0.847 ± 0.012 0.835 ± 0.012

The results in Table 25 confirmed that the ANN models were not overfitting, as

confirmed by the similarity between the model training and testing R 2 values. The

results from model training and testing of the ANN models are shown in Figure 61

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
206
Figure 61. ANN model results using input synthesis variables (results from the
training and testing runs that resulted in the greatest model accuracy).

Both the PyTorch and sklearn ANN models predicted the zeolite yield with the highest

level of accuracy observed in this study (84 - 85 %). The convergence of the training

and validation curves confirmed that the model was not overfitting (Supplementary

Figure 9). In addition, the hyperparameter tuning of the models reduced overfitting

such that the training accuracy exhibited a high degree of similarity to the testing

accuracy. Unlike previous models, the hidden element in the ANN model architecture

meant the feature importance was not obtainable. Thus, there was a trade-off between

the accuracy and transparency of the ANN model. However, from the previous

visualisations of feature importance, it was clear that as the model complexity

increased, the model's dependence on the input synthesis descriptors became more

balanced. Thus, it was expected that this trend was true for the ANN models, given

the higher accuracy obtained by the models. Therefore, it is assumed that the feature

importance obtained for the XGBoost model in Figure 60 may be similar for the ANN

models.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
207
Discussion: Model Comparison

Figure 62a shows a trend of increasing training and testing accuracy with model

complexity. The reappearance of sodium silicate and reaction time synthesis variables

from the PCA through to the feature importance reinforced the machine learning

results and data quality. The lesser performance of XGBoost in comparison to ANN

models was of interest as recent developments in gradient boosting algorithms often

return accuracies greater than those obtained through ANN (T. Chen & Guestrin).

However, it is critical to understand the data structure and relationships when

comparing the accuracy of XGBoost to ANN. XGBoost outperforms ANN, where

sequential relationships exist within a dataset (Pedregosa et al., 2011). In contrast,

given the interconnected nodes and layers, ANNs more accurately model multiple

independent relationships and data exhibiting multi-collinearity (Van Dyke Parunak,

1998). As shown through PCA, the multi-collinearity of the zeolite synthesis data

supported the existence of parallel relationships in the dataset. Hence, this situation

explained why the ANN modelled the system with greater accuracy than XGBoost

(and other tree-based model architectures).

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
208
Figure 62. (a) Comparison of machine learning model results (average R^2 value) (b)
Comparison of approaches to zeolite product analysis through ANN models (average
R^2 value)

From this study, the ANN modelled the zeolite synthesis system with the highest

accuracy. To confirm the validity of the hybrid data approach and the selection of the

hybrid 10 dataset, the ANN models were trained using various qualitative, quantitative,

hybrid 10, hybrid 20 and hybrid 40 datasets. The sklearn model was trained using the

varying datasets but with the same hyperparameters and training/testing split to

facilitate comparison (Figure 62b).

When modelled using ANN, the quantitative XRD data analysis approach provided

the least accurate representation of the zeolite synthesis system. This outcome was in

accord with the SEM images, which showed that the extent of zeolite crystals present

was not accounted for in the quantitative analysis approach. Moreover, the

quantitative dataset returned overfitted results when training the ANN model. The

qualitative XRD approach was computationally the most accurate despite errors

previously identified. However, further analysis of the zeolite LTA wt% distribution

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
209
within the qualitative data revealed the problems with the qualitative XRD method.

The distribution of data between 0 and 100 wt% was heavily skewed in the negative

direction, with 80% of the data lying between 80 and 100 wt%, of which 77% of the

qualitative data fell between 90 and 100 wt% zeolite LTA. Therefore, when predicting

the amount of zeolite LTA produced using the qualitative data, most of the predicted

results were between 90 and 100 wt%. Thus, the model incorrectly learned that most

of the zeolite synthesis routes in the qualitative data were successful.

Of the hybrid approaches, "hybrid 10" produced the most accurate model, thus

validating the understanding that the ambiguous cases of zeolite product analysis

occurred at the beginning of zeolite growth. Notably, the "hybrid 40" dataset also

overfitted the training model data. The new hybrid data strategy has developed a better

dataset for zeolite synthesis evaluation. Notably, theoretical understanding of the

zeolite growth curve combined with machine learning has proven to deliver new levels

of accuracy in practical applications. The results from the machine learning

application demonstrate that without a robust and high-quality dataset, the machine

learning models cannot be expected to accurately model complex systems.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
210
Conclusions

This study met the primary aim of this study by successfully applying statistics and

machine learning principles to zeolite LTA synthesis. In addition, the hypothesis that

machine learning may predict preferred zeolite synthesis strategies was supported.

Indeed, the new modelling strategies have the potential to pre-evaluate synthesis routes

for a wide range of zeolite synthesis problems to optimise yield and performance.

Creation of a robust dataset was of primary importance as there were not only many

variables to consider but also a high degree of intercorrelation and multi-collinearit y

in the data. Management of the dataset was consequently of importance. For example,

a combination of quantitative and qualitative X-ray diffraction data was required to

determine the presence of crystalline and amorphous/non-diffracting material.

Quantitative XRD was able to identify amorphous/non-diffracting material. Whereas

qualitative XRD analysis was necessary when the synthesis was more progressed.

Consequently, a "hybrid" XRD analysis method was preferred. This dataset

interpretation strategy was innovative and resulted from the machine learning

approach. However, it was critical to only include the input synthesis variables.

Inclusion of output synthesis descriptors other than the zeolite yield resulted in false

predicting of the machine learning results. The application of PCA enabled

visualization of the higher dimensional data. Nevertheless, the component coefficients

revealed that each variable contributed in some way to the first four PCs.

The linear regression, ridge regression and tree-based models all returned R 2 values

less than 0.7 indicating that these machine learning models cannot accurately model

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
211
the zeolite synthesis system. The ANN approach modelled the system to the highest

accuracy of R2 of 0.84. The similarity of the training and testing results confirmed that

the models were not overfitting, an issue that must be closely monitored when

modelling systems characterised by small datasets. In terms of ANN performance, the

"hybrid 10" dataset produced the most accurate model; thus, validating that the

ambiguous cases of zeolite product analysis occurred at the beginning of the zeolite

growth curve. Thus, for this investigation the preferred model was the Artificial

Neural Network.

Future research should focus on adding complexity to the data, including modelling

the zeolite mother liquor, which would enable material economies to be optimised.

The additional complexity may require Deep Neural Networks (DNN) to achieve

higher testing accuracy. Alternate advanced machine learning techniques, including

Generative Adversarial Networks (GAN) and physics-informed GAN (PI-GAN), may

also be beneficial.

Chapter 7: Evaluation and application of machine learning principles to Zeolite LTA synthesis
212
Chapter 8: Conclusions and Future Work

Conclusions

The aim of this research was to establish a low cost, scalable, and environmentally

sustainable process for converting aluminosilicate waste materials into high-value

zeolite products. As such, various waste activation methodologies using a combination

of alkali dissolution, temperature, and time of exposure were investigated. The

resultant active material was then used to make synthetic zeolite LTA as this is the

largest volume zeolite used today. Integration of machine learning was completed in

order to further optimize the conversion of waste aluminosilicate materials to zeolite

LTA. In particular, the need for clean and robust datasets was demonstrated.

The first research approach involved activation of aluminosilicate waste from the

lithium industry (spodumene leachate residue (SLR)). Two products were

subsequently made from the activated material: (1) cancrinite (which could be further

modified to make zeolite LTA by acid dissolution, neutralization, and hydrothermal

reaction) and (2) zeolite LTA made from hydrothermal reaction of mother liquor to

which sodium aluminate was added to create an Si/Al ratio = 1. The preferred

activation temperature was 240°C and the time required for activation was one hour.

Notably these conditions were probably more cost effective than previously reported

conditions (such as 150°C for 24 hours) as the overall power consumption was

substantially lower. It is essential for both economic and environmental reasons that

cancrinite is either employed, as for example, a sorbent for water treatment, or

converted to another material. For example, conversion of cancrinite to zeolite LTA

Chapter 8: Conclusions and Future Work 213


via acid dissolution, NaOH neuralization was not deemed to be feasible due to

additional cost to recover or remove sodium chloride generated in this process.

Chapter 4 presented the findings of research objective 2, wherein the potential of

natural zeolite waste conversion to two products (cancrinite and zeolite LTA) was

studied. As natural zeolite waste contains a higher silica content than SLR (ca. 4.8:1

Si/Al ratio compared to Si/Al ratio =2 for SLR) this material should produce relatively

more zeolite LTA compared to cancrinite (thus limiting the amount of this material

formed). The natural zeolite waste was again activated at 240°C for one-hour. A key

finding was the fact that 6M NaOH was necessary to promote dissolution of active Si

in the mother liquor. As predicted, a smaller amount of cancrinite was created, and

significantly, a range of impurities in the initial natural zeolite were mixed in with the

cancrinite as they were not soluble under the activation conditions. The resultant

zeolite LTA exhibited 95.54 wt% crystallinity. Moreover, lanthanum impregnated

cancrinite was shown to remove phosphates from water (achieving 99% removal

efficiency).

Despite the success of the two-product approach, it was still desirable to develop a

method which activated wastes and only produced one product. Thus, the third

research objective concerned discovery of a novel Quench Method for transforming

SLR into zeolite LTA. The activation of the waste material was carried out at low

temperature (100°C), using 10 M NaOH for 60 minutes.

The key insight was that the activation solution had to be quenched by rapid addition

of water and control of temperature to change the reaction conditions to promote

Chapter 8: Conclusions and Future Work 214


zeolite LTA synthesis. Consequently, prevention of the formation of undesirable by-

products, such as cancrinite or sodalite was minimised. The Quench approach appears

relatively low cost due to use of only 100 oC to activate the SLR waste.

The fourth research objective was addressed in Chapters 6 and 7, which focused on

the application of machine learning methods to advance zeolite synthesis. The

implementation of machine learning approaches for zeolite synthesis is not trivial as

inherently small data sets are employed which makes statistical analysis difficult.

Consequently, “web crawlers” to extract data to increase the data set size are often

used. However, this study revealed the pitfalls of using data from other research

groups. For example, it was found that utilizing identical recipes in different reactors

resulted in significant differences in the final product. Other criteria investigated

included pre-heated samples, use of stirred or static conditions, availability of

comprehensive process data (including mass balance) and zeolite yields. In all cases,

these latter criteria were not commonly addressed in the literature. Hence, use of data

from other researchers to facilitate machine learning was probably quality deficient.

Consequently, this study created a clean data set of information collected only in our

laboratory. Despite the small size (ca. 85 tests) of the data set, the use of statistical

and machine learning principles was successful in identifying that artificial neural

networks were at least 85 % accurate in predicting zeolite synthesis outputs.

In conclusion, this study pioneered the development of both a two-product and a

Quench process to activate various aluminosilicate wastes more effectively.

Chapter 8: Conclusions and Future Work 215


Both processes significantly reduce the cost of activation based upon the power

consumed during the activation procedure. In addition, both methods concomitantly

removed the majority of impurity phases present. Integration of machine learning

principles with the novel activation processes may further improve zeolite LTA

synthesis.

Future Work

Despite the success of the study, further work is required to enhance the technology

readiness level (TRL).

Firstly, the use of cancrinite for sorption applications needs to be developed further.

Key information required includes: (1) Coating of the cancrinite with either lanthanum

salts or alternate species such as iron which have been shown to sorb phosphate species

from water; (2) Determination of doped cancrinite performance (kinetics; equilibr ia;

column studies); (3) Extrusion and beading of cancrinite powders into industry

relevant forms; (4) Physical robustness (crush strength, attrition rates).

Secondly, the use of a continuous reactor to activate the aluminosilicate is sensible in

terms of controlling the quenching rate when using the Quench Method. A prototype

of the Quench Methods needs to be bult and tested to ensure performance is optimal.

(Appendix A).

Lastly, it is recommended to continue increasing the complexity of the database for

machine learning activities to enable a more efficient process. While the Artificial

Chapter 8: Conclusions and Future Work 216


Neural Network (ANN) was the preferred model for this research, increasing data

complexity may enable higher testing accuracy through Deep Neural Networks (DNN)

or more advanced machine learning techniques.

Chapter 8: Conclusions and Future Work 217


Bibliography

Ackley, M. W., Rege, S. U., & Saxena, H. (2003). Application of natural zeolites in
the purification and separation of gases. Microporous and Mesoporous
Materials, 61(1), 25-42. doi:10.1016/S1387-1811(03)00353-6
Adjiman, C. S., Sahinidis, N. V., Vlachos, D. G., Bakshi, B., Maravelias, C. T., &
Georgakis, C. (2021). Process systems engineering perspective on the design
of materials and molecules. Industrial & Engineering Chemistry Research,
60(14), 5194-5206. doi:10.1021/acs.iecr.0c05399
Ahmadon, A., Nazir, L. S. M., Yeong, Y. F., & Sabdin, S. (2018). Formation of pure
NaX zeolite: Effect of ageing and hydrothermal synthesis parameters. Paper
presented at the IOP Conference Series: Materials Science and Engineering.
Al-Shayji, K. A., & Liu, Y. A. (2002). Predictive modeling of large-scale commercial
water desalination plants: data-based neural network and model-based process
simulation. Industrial & Engineering Chemistry Research, 41(25), 6460-6474.
Alkhlel, A., & de Lasa, H. (2018). Catalytic cracking of hydrocarbons in a CREC Riser
Simulator using a Y‑Zeolite-based catalyst: Assessing the Catalyst/Oil Ratio
Effect. Industrial & Engineering Chemistry Research, 57(41), 13627-13638.
doi:10.1021/acs.iecr.8b02427
Ameh, A. E., Fatoba, O. O., Musyoka, N. M., & Petrik, L. F. (2017). Influence of
aluminium source on the crystal structure and framework coordination of Al
and Si in fly ash-based zeolite NaA. Powder Technology, 306, 17-25.
doi:10.1016/j.powtec.2016.11.003
Andaç, Ö., Tatlıer, M., Sirkecioğlu, A., Ece, I., & Erdem-Şenatalar, A. (2005). Effects
of ultrasound on zeolite A synthesis. Microporous and Mesoporous Materials,
79(1), 225-233. doi:10.1016/j.micromeso.2004.11.007
Australian-Government. (2020). Industry 4.0. Retrieved from
https://www.industry.gov.au/funding-and-incentives/industry-40
Ayele, L., Pérez-Pariente, J., Chebude, Y., & Diaz, I. (2016a). Synthesis of zeolite A
using kaolin from Ethiopia and its application in detergents. New Journal of
Chemistry, 4(4), 344-3446. doi:10.1039/c5nj03097h
Ayele, L., Pérez-Pariente, J., Chebude, Y., & Díaz, I. (2016b). Conventional versus
alkali fusion synthesis of zeolite A from low grade kaolin. Applied Clay
Science, 132-133, 485-490. doi:10.1016/j.clay.2016.07.019
Basaldella, E. I., Kikot, A., & Tara, J. C. (1997). Effect of aluminum concentration on
crystal size and morphology in the synthesis of a NaAl zeolite. Materials
Letters, 31, 83-86.
Bauer, H. P. (2008). Production of silicates and zeolites for detergent industry. In
Handbook of Detergents, Part F: Production (pp. 387-419).
Bertini, G., Elia, S., Ceciarini, F., & Dani, C. (2013). Reduction of catheter-related
bloodstream infections in preterm infants by the use of catheters with the
AgION antimicrobial system. Early Human Development, 89(1), 21-25.
doi:10.1016/j.earlhumdev.2012.07.003
Beushausen, H., Alexander, M., & Ballim, Y. (2012). Early-age properties, strength
development and heat of hydration of concrete containing various South
African slags at different replacement ratios. Construction and Building
Materials, 29, 533-540. doi:10.1016/j.conbuildmat.2011.06.018

Bibliography 218
Bingre, R., Louis, B., & Nguyen, P. (2018). An Overview on Zeolite Shaping
Technology and Solutions to Overcome Diffusion Limitations. Catalysts, 8.
doi:10.3390/catal8040163
Breck, D. W., Eversole, W. G., Milton, R. M., Reed, T. B., & Thomas, T. L. (1956).
Crystalline Zeolites. I. The Properties of a New Synthetic Zeolite, Type A.
Journal of the American Chemical Society, 78(23), 5963-5972.
doi:10.1021/ja01604a001
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5-32.
doi:10.1023/A:1010933404324
Bronić, J., Palčić, A., Subotić, B., Itani, L., & Valtchev, V. (2012). Influence of
alkalinity of the starting system on size and morphology of the zeolite A
crystals. Materials Chemistry and Physics, 132(2-3), 973-976.
doi:10.1016/j.matchemphys.2011.12.043
Buhl, J. C. (1991). Synthesis and characterization of the basic and non-basic members
of the cancrinite-natrodavyne family. Thermochimica Acta, 178, 19-31.
doi:10.1016/0040-6031(91)80294-S
Buhl, J. C. (2016). On the autothermal synthesis of zeolites. Microporous and
Mesoporous Materials, 222, 73-79. doi:10.1016/j.micromeso.2015.10.004
Buhl, J. C., & Petrov, V. (2021). Synthesis and properties of phosphate cancrinite
(PO4‐CAN) a synthetic counterpart of depmeierite. Zeitschrift für
Anorganische und Allgemeine Chemie (1950), 647(16-17), 1647-1654.
doi:10.1002/zaac.202100160
Burriesci, N., Crisafulli, M. L., Giordano, N., & Bart, J. (1984). Factors affecting
formation of zeolite a from alumino-silicate gels. Materials Letters, 2(5), 401-
406.
Butler, K. T., Davies, D. W., Cartwright, H., Isayev, O., & Walsh, A. (2018). Machine
learning for molecular and materials science. Nature, 559(7715), 547-555.
doi:10.1038/s41586-018-0337-2
Caballero, L., Jojoa, M., & Percybrooks, W. S. (2020). Optimized neural networks in
industrial data analysis. SN Applied Sciences, 2(2). doi:10.1007/s42452- 020-
2060-5
Cabrera, P., Carta, J. A., González, J., & Melián, G. (2017). Artificial neural networks
applied to manage the variable operation of a simple seawater reverse osmosis
plant. Desalination, 416, 140-156. doi:10.1016/j.desal.2017.04.032
Carr, D. A., Lach-hab, M., Yang, S., Vaisman, I. I., & Blaisten-Barojas, E. (2009).
Machine learning approach for structure-based zeolite classification.
Microporous and Mesoporous Materials, 117(1-2), 339-349.
doi:10.1016/j.micromeso.2008.07.027
Casci, J. L. (2005). Zeolite molecular sieves: preparation and scale-up. Microporous
and Mesoporous Materials, 82(3), 217-226.
doi:10.1016/j.micromeso.2005.01.035
Chen, D., Hu, X., Shi, L., Cui, Q., Wang, H., & Yao, H. (2012). Synthesis and
characterization of zeolite X from lithium slag. Applied Clay Science, 59-60,
148-151. doi:10.1016/j.clay.2012.02.017
Chen, T., & Guestrin, C. (2016, August). Xgboost: A scalable tree boosting system. In
Proceedings of the 22nd ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining (pp. 785-794).
Chen, Y., Tian, Q., Chen, B., Shi, X., & Liao, T. (2011). Preparation of lithium
carbonate from spodumene by a sodium carbonate autoclave process.
Hydrometallurgy, 109(1-2), 43-46. doi:10.1016/j.hydromet.2011.05.006

Bibliography 219
Choubey, P. K., Kim, M.-s., Srivastava, R. R., Lee, J.-c., & Lee, J.-Y. (2016). Advance
review on the exploitation of the prominent energy-storage element: Lithium.
Part I: From mineral and brine resources. Minerals Engineering, 89, 119-137.
doi:10.1016/j.mineng.2016.01.010
Chu, K. H., & Hashim, M. A. (2023). Modeling of aqueous phase adsorption: Is it time
to bid adieu to the Harkins–Jura isotherm? Journal of molecular liquids, 371,
121122. doi:10.1016/j.molliq.2022.121122
Collins, F., Rozhkovskaya, A., Outram, J. G., & Millar, G. J. (2020). A critical review
of waste resources, synthesis, and applications for Zeolite LTA. Microporous
and Mesoporous Materials, 291. doi:10.1016/j.micromeso.2019.109667
Conroy, B., Nayak, R., Hidalgo, A. L. R., & Millar, G. J. (2022). Evaluation and
application of machine learning principles to Zeolite LTA synthesis.
Microporous and Mesoporous Materials, 335.
doi:10.1016/j.micromeso.2022.111802
Conterosito, E., Lopresti, M., & Palin, L. (2020). In Situ X-ray Diffraction Study of
Xe and CO2 Adsorption in Y Zeolite: Comparison between Rietveld and PCA-
Based Analysis. Crystals (Basel), 10(6), 483. doi:10.3390/cryst10060483
Corma, A., Serra, J. M., Serna, P., & Moliner, M. (2005). Integrating high-throughput
characterization into combinatorial heterogeneous catalysis: unsupervised
construction of quantitative structure/property relationship models. Journal of
Catalysis, 232(2), 335-341. doi:10.1016/j.jcat.2005.03.019
Cortez, P., Cerdeira, A., Almeida, F., Matos, T., & Reis, J. (2009). Modeling wine
preferences by data mining from physicochemical properties. Decision Support
Systems, 47(4), 547-553. doi:10.1016/j.dss.2009.05.016
Courtney, P., Murphy, S., Chambers, B., Green, S., & Christensen, P. (2017). Mining
wetlands: Retrieved from
https://www.abc.net.au/news/rural/programs/landline/2017-07-08/mining-
wetlands:-abandoned-mine-pits-evolve-into/8691008
Cundy, C. S., & Cox, P. A. (2005). The hydrothermal synthesis of zeolites: Precursors,
intermediates and reaction mechanism. Microporous and Mesoporous
Materials, 82(1-2), 1-78. doi:10.1016/j.micromeso.2005.02.016
Dahal, K. R., Dahal, J. N., Banjade, H., & Gaire, S. (2021). Prediction of wine quality
using machine learning algorithms. Open Journal of Statistics, 11(2), 278-289
Đặng, T.-H., Nguyễn, X.-H., Chou, C.-L., & Chen, B.-H. (2021). Preparation of
cancrinite-type zeolite from diatomaceous earth as transesterification catalysts
for biodiesel production. Renewable Energy, 174, 347-358.
doi:10.1016/j.renene.2021.04.068
Daramola, M. O., Oloye, O., & Yaya, A. (2017). Nanocomposite sodalite/ceramic
membrane for pre-combustion CO<inf>2</inf> capture: synthesis and
morphological characterization. International Journal of Coal Science and
Technology, 4(1), 60-66. doi:10.1007/s40789-016-0124-3
Davis, M. E. (1991). Zeolites and molecular sieves: not just ordinary catalysts.
Industrial & Engineering Chemistry Research, 30(8), 1675-1683.
doi:10.1021/ie00056a001
De Silva, P., Sagoe-Crenstil, K., & Sirivivatnanon, V. (2007). Kinetics of
geopolymerization : Role of Al2O3 and SiO2. Cement and Concrete Research,
37(4), 512-518. doi:10.1016/j.cemconres.2007.01.003
Deabriges, J. (1982) Industrial process for continuous production of zeolite A. U.S.
Patent No. 4,314,979. 9 Feb. 1982. Washington, DC: U.S. Patent and
Trademark Office.

Bibliography 220
Deneyer, A., Ke, Q., Devos, J., & Dusselier, M. (2020). Zeolite synthesis under
nonconventional conditions: reagents, reactors, and modi operandi. Chemistry
of Materials, 32(12), 4884-4919. doi:10.1021/acs.chemmater.9b04741
Deng, Y., Flury, M., Harsh, J. B., Felmy, A. R., & Qafoku, O. (2006). Cancrinite and
sodalite formation in the presence of cesium, potassium, magnesium, calcium
and strontium in Hanford tank waste simulants. Applied Geochemistry, 21(12),
2049-2063. doi:10.1016/j.apgeochem.2006.06.019
Derouane, E. G. (Ed.). (1992). Zeolite microporous solids : synthesis, structure, and
reactivity (1st ed. 1992.). Springer Science Business Media B. V.
https://doi.org/10.1007/978-94-011-2604-5
Dessemond, C., Lajoie-Leroux, F., Soucy, G., Laroche, N., & Magnan, J. F. (2019).
Spodumene: the lithium market, resources and processes. Minerals, 9(6).
doi:10.3390/min9060334
D’Elia, A., Pinto, D., Eramo, G., Giannossa, L. C., Ventruti, G., & Laviano, R. (2018).
Effects of processing on the mineralogy and solubility of carbonate-rich clays
for alkaline activation purpose: mechanical, thermal activation in red/ox
atmosphere and their combination. Applied Clay Science, 152, 9–21.
https://doi.org/10.1016/j.clay.2017.11.036
Dong, S., Wang, Y., Zhao, Y., Zhou, X., & Zheng, H. (2017). La3+/La(OH)3 loaded
magnetic cationic hydrogel composites for phosphate removal: Effect of
lanthanum species and mechanistic study. Water Research, 126, 433-441.
doi:https://doi.org/10.1016/j.watres.2017.09.050
Dong, Y., Lin, H., & He, Y. (2017). Correlation between physicochemical properties
of modified clinoptilolite and its performance in the removal of ammonia -
nitrogen. Environmental Monitoring and Assessment, 189(3), 107-107.
doi:10.1007/s10661-017-5806-9
Drioli, E., & Giorno, L. (2016). Encyclopedia of membranes, Berlin-Heidelberg:
Springer-Verlag
El-Nahas, S., Osman, A. I., Arafat, A. S., Al-Muhtaseb, A. a. H., & Salman, H. M.
(2020). Facile and affordable synthetic route of nano powder zeolite and its
application in fast softening of water hardness. Journal of Water Process
Engineering, 33, 101104. doi:10.1016/j.jwpe.2019.101104
Eroglu, N., Emekci, M., & Athanassiou, C. G. (2017). Applications of natural zeolites
on agriculture and food production. Journal of the Science of Food and
Agriculture, 97(11), 3487-3499. doi:10.1002/jsfa.8312
Esaifan, M., Warr, L. N., Grathoff, G., Meyer, T., Schafmeister, M. T., Kruth, A., &
Testrich, H. (2019). Synthesis of hydroxy-sodalite/cancrinite zeolites from
calcite-bearing kaolin for the removal of heavy metal ions in aqueous media.
Minerals, 9(8). doi:10.3390/min9080484
Evans, J.D. and Coudert, F.X. (2017) Predicting the Mechanical Properties of Zeolite
Frameworks by Machine Learning. Chemistry of Materials, 2017. 29(18)
7833-7839.
Fakin, T., Ristić, A., Mavrodinova, V., & Zabukovec Logar, N. (2015). Highly
crystalline binder-free ZSM-5 granules preparation. Microporous and
Mesoporous Materials, 213, 108-117. doi:10.1016/j.micromeso.2015.04.010
Fawer, M., Postlethwaite, D., & Klüppel, H. J. (1998). Life cycle inventory for the
production of Zeolite A for detergents. International Journal of Life Cycle
Assessment, 3(2), 71-74. doi:10.1007/BF02978490
Flexer, V., Baspineiro, C. F., & Galli, C. I. (2018). Lithium recovery from brines: A
vital raw material for green energies with a potential environmental impact in

Bibliography 221
its mining and processing. Science of the Total Environment, 639, 1188-1204.
doi:10.1016/j.scitotenv.2018.05.223
Frank, A. G., Dalenogare, L. S., & Ayala, N. F. (2019). Industry 4.0 technologies:
Implementation patterns in manufacturing companies. International Journal of
Production Economics, 210, 15-26. doi:10.1016/j.ijpe.2019.01.004
Fruijtier-Pölloth, C. (2009). The safety of synthetic zeolites used in detergents.
Archives of Toxicology, 83(1), 23-35. doi:10.1007/s00204-008-0327-5
Gaillac, R., Chibani, S., & Coudert, F. X. (2020). Speeding up discovery of auxetic
zeolite frameworks by machine learning. Chemistry of Materials, 32(6), 2653-
2663. doi:10.1021/acs.chemmater.0c00434
García, G., Aguilar-Mamani, W., Carabante, I., Cabrera, S., Hedlund, J., & Mouzon,
J. (2015). Preparation of zeolite A with excellent optical properties from clay.
Journal of Alloys and Compounds, 619, 771-777.
doi:10.1016/j.jallcom.2014.09.080
Garcia, G., Cabrera, S., Hedlund, J., & Mouzon, J. (2018). Selective synthesis of FAU-
type zeolites. Journal of Crystal Growth, 489, 36-41.
doi:10.1016/j.jcrysgro.2018.02.022
Ghrear, T. M. A., Ng, E.-P., Vaulot, C., Daou, T. J., Ling, T. C., Tan, S. H., . . .
Mintova, S. (2020). Recyclable synthesis of Cs-ABW zeolite nanocrystals
from non-reacted mother liquors with excellent catalytic henry reaction
performance. Journal of Environmental Chemical engineering, 8(1), 103579.
doi:10.1016/j.jece.2019.103579
Gougazeh, M., & Buhl, J. C. (2014). Synthesis and characterization of zeolite A by
hydrothermal transformation of natural Jordanian kaolin. Journal of the
Association of Arab Universities for Basic and Applied Sciences, 15(1), 35-42.
doi:10.1016/j.jaubas.2013.03.007
Guarino Bertholini, M. (2016). Innovative applications of natural zeolites. Doctoral
dissertation, Queensland University of Technology,
Gupta, Y. (2018). Selection of important features and predicting wine quality using
machine learning techniques. Procedia Computer Science, 125, 305-312.
doi:10.1016/j.procs.2017.12.041
Hackbarth, K., Gesing, T. M., Fechtelkord, M., Stief, F., & Buhl, J. C. (1999).
Synthesis and crystal structure of carbonate cancrinite
Na8[AlSiO4]6CO3(H2O)3.4, grown under low-temperature hydrothermal
conditions. Microporous and Mesoporous Materials, 30(2), 347-358.
doi:https://doi.org/10.1016/S1387-1811(99)00046-3
Hagio, T., Park, J. H., Lin, Y., Tian, Y., Hu, Y. H., Li, X., . . . Ichino, R. (2021). Facile
hydrothermal synthesis of EAB-Type zeolite under static synthesis conditions.
Crystal Research and Technology, 56(4). doi:10.1002/crat.202000163
Han, G., Gu, D., Lin, G., Cui, Q., & Wang, H. (2018). Recovery of lithium from a
synthetic solution using spodumene leach residue. Hydrometallurgy, 177, 109-
115. doi:10.1016/j.hydromet.2018.01.004
Han, Y., Jeppesen, E., Lürling, M., Zhang, Y., Ma, T., Li, W., . . . Li, K. (2022).
Combining lanthanum-modified bentonite (LMB) and submerged
macrophytes alleviates water quality deterioration in the presence of omni-
benthivorous fish. Journal of Environmental Management, 314, 115036-
115036. doi:10.1016/j.jenvman.2022.115036
Hang Chau, J. L., Tellez, C., Yeung, K. L., & Ho, K. (2000). The role of surface
chemistry in zeolite membrane formation. Journal of Membrane Science,
164(1), 257-275. doi:10.1016/S0376-7388(99)00214-8

Bibliography 222
Hanson, A. B., Schwerdt, I. J., Nizinski, C. A., Lee, R. N., Mecham, N. J., Abbott, E.
C., . . . McDonald, L. W. (2021). Impact of controlled storage conditions on
the hydrolysis and surface morphology of amorphous-UO3. ACS Omega,
6(12), 8605-8615. doi:10.1021/acsomega.1c00435
Hao, H., Liu, Z., Zhao, F., Geng, Y., & Sarkis, J. (2017). Material flow analysis of
lithium in China. Resources Policy, 51, 100-106.
doi:10.1016/j.resourpol.2016.12.005
He, J., Wang, W., Sun, F., Shi, W., Qi, D., Wang, K., . . . Chen, X. (2015). Highly
efficient phosphate scavenger based on well-dispersed La(OH)3 nanorods in
polyacrylonitrile nanofibers for nutrient-starvation antibacteria. ACS Nano,
9(9), 9292-9302. doi:10.1021/acsnano.5b04236
He, Z-h., Li, L.-y., & Du, S.-g. (2017). Mechanical properties, drying shrinkage, and
creep of concrete containing lithium slag. Construction & Building Materials,
147, 296-304. doi:10.1016/j.conbuildmat.2017.04.166
Heng, S. (2020). Industry 4.0 : challenges, trends, and solutions in management and
engineering. Boca Raton, FL: CRC Press/Taylor & Francis Group.
Hermeler, G., Buhl, J. C., & Hoffmann, W. (1991). The influence of carbonate on the
synthesis of an intermediate phase between sodalite and cancrinite. Catalysis
Today, 8(4), 415-426. doi:10.1016/0920-5861(91)87020-N
Hinesly, T. D., & Jones, R. L. (1990). Phosphorus in waters from sewage sludge
amended lysimeters. Environmental pollution (1987), 65(4), 293-309.
doi:10.1016/0269-7491(90)90122-S
Hnizdo, E., & Vallyathan, V. (2003). Chronic obstructive pulmonary disease due to
occupational exposure to silica dust: a review of epidemiological and
pathological evidence. Occupational and Environmental Medicine, 60(4), 237-
243. doi:10.1136/oem.60.4.237
Hong, J. L. X., Maneerung, T., Koh, S. N., Kawi, S., & Wang, C. H. (2017).
Conversion of Coal Fly Ash into Zeolite Materials: Synthesis and
Characterizations, Process Design, and Its Cost-Benefit Analysis. Industrial
and Engineering Chemistry Research, 56(40), 11565-11574.
doi:10.1021/acs.iecr.7b02885
Howard, W. R. (2007). Pattern Recognition and Machine LearningPattern Recognition
and Machine Learning. Kybernetes, 36(2), 275-275.
doi:10.1108/03684920710743466
Huang, S., Wang, Y., Zhang, L., Zhu, S., Ma, Z., Cui, Q., & Wang, H. (2023). Insight
into the kinetic behavior of microwave-assisted synthesis of NaX zeolite from
lithium slag. New Journal of Chemistry, 47(3), 14335–14343.
https://doi.org/10.1039/d3nj02260a
Hwang, E.-Y., Kim, J.-R., Choi, J.-K., Woo, H.-C., & Park, D.-W. (2002).
Performance of acid treated natural zeolites in catalytic degradation of
polypropylene. Journal of Analytical and Applied Pyrolysis, 62(2), 351-364.
doi:10.1016/S0165-2370(01)00134-6
Ivankovic, T., Dikic, J., Du Roscoat, S. R., Dekic, S., Hrenovic, J., & Ganjto, M.
(2019). Removal of emerging pathogenic bacteria using metal-exchanged
natural zeolite bead filter. Water Science and Technology, 80(6), 1085-1098.
doi:10.2166/wst.2019.348
IZA-SC ( 2007) Database of Zeolite Structures, Structure Commission of the
International Zeolite Association (IZA-SC) https://asia.iza-structure.org/IZA-
SC/framework.php?STC=LTA

Bibliography 223
Izidoro, J. C., Kim, M. C., Bellelli, V. F., Pane, M. C., Botelho Junior, A. B., Espinosa,
D. C. R., & Tenório, J. A. S. (2019). Synthesis of zeolite A using the waste of
iron mine tailings dam and its application for industrial effluent treatment.
Journal of Sustainable Mining. doi:10.1016/j.jsm.2019.11.001
Jafari, M., Mohammadi, T., & Kazemimoghadam, M. (2014). Synthesis and
characterization of ultrafine sub-micron Na-LTA zeolite particles prepared via
hydrothermal template-free method. Ceramics International, 40(8 PART A),
12075-12080. doi:10.1016/j.ceramint.2014.04.047
Jamil, T. S., Ibrahim, H. S., Abd El-Maksoud, I. H., & El-Wakeel, S. T. (2010).
Application of zeolite prepared from Egyptian kaolin for removal of heavy
metals: I. Optimum conditions. Desalination, 258(1-3), 34-40.
doi:10.1016/j.desal.2010.03.052
Jawad, J., Hawari, A. H., & Zaidi, S. (2020). Modeling of forward osmosis process
using artificial neural networks (ANN) to predict the permeate flux.
Desalination, 484, 114427. Jensen, Z., Kim, E., Kwon, S., Gani, T. Z. H.,
Roman-Leshkov, Y., Moliner, M., . . . Olivetti, E. (2019). A Machine Learning
Approach to Zeolite Synthesis Enabled by Automatic Literature Data
Extraction. ACS Central Science, 5(5), 892-899.
doi:10.1021/acscentsci.9b00193
Ji, X., Jia, H., Wang, Y., & Yang, X. (2018). One-pot synthesis of IM-5 zeolite.
Journal of Porous Materials, 26(2), 343-351. doi:10.1007/s10934-018-0606- 3
Ji, Y., Wang, Y., Xie, B., & Xiao, F.-S. (2015). Zeolite Seeds: Third Type of Structure
Directing Agents in the Synthesis of Zeolites. Comments on Inorganic
Chemistry, 36(1), 1-16. doi:10.1080/02603594.2015.1031375
Jiang, J., Feng, L., Gu, X., Qian, Y., Gu, Y., & Duanmu, C. (2012). Synthesis of zeolite
A from palygorskite via acid activation. Applied Clay Science, 55, 108-113.
doi:10.1016/j.clay.2011.10.014
Johnson, E. B. G., & Arshad, S. E. (2014). Hydrothermally synthesized zeolites based
on kaolinite: A review. Applied Clay Science, 97-98, 215-221.
doi:10.1016/j.clay.2014.06.005
Jolliffe, I. T. (2002). Principal Component Analysis (2nd ed. ed.). Secaucus: Springer.
Jordan, M. I., & Mitchell, T. M. (2015). Machine learning: Trends, perspectives, and
prospects. Science, 349(6245), 255-260..
Ju, J., Zeng, C., Zhang, L., & Xu, N. (2006). Continuous synthesis of zeolite NaA in a
microchannel reactor. Chemical Engineering Journal, 116(2), 115-121.
doi:10.1016/j.cej.2005.11.006
Karrech, A., Azadi, M. R., Elchalakani, M., Shahin, M. A., & Seibi, A. C. (2020). A
review on methods for liberating lithium from pegmatities. Minerals
Engineering, 145, 106085. doi:10.1016/j.mineng.2019.106085
Kassem, A. H., Ayoub, G. M., & Zayyat, R. (2022). Advances in nanomaterials for
phosphates removal from water and wastewater: a review. Nanotechnology for
Environmental Engineering, 7(3), 609-634. doi:10.1007/s41204-022-00258-w
Katsuki, H., Furuta, S., Watari, T., & Komarneni, S. (2005). ZSM-5 zeolite/porous
carbon composite: Conventional- and microwave-hydrothermal synthesis from
carbonized rice husk. Microporous and Mesoporous materials, 86(1), 145-151.
doi:10.1016/j.micromeso.2005.07.010
Kettinger, F. R., Laudone, J. A., & Pierce, R. H. (1979). Preparing Zeolite NAA. US
4150100. PQ Corporation
.Khan, H., Yerramilli, A. S., D'Oliveira, A., Alford, T. L., Boffito, D. C., & Patience,
G. S. (2020). Experimental methods in chemical engineering: X‐ray diffraction

Bibliography 224
spectroscopy—XRD. Canadian Journal of Chemical Engineering, 98(6),
1255-1266. doi:10.1002/cjce.23747
Khan, M. A., Soh, J. E., Maenner, M., Thompson, W. W., & Nelson, N. P. (2019). A
machine-learning algorithm to identify hepatitis C in health insurance claims
data. Online Journal of Public Health Informatics, 11(1).
doi:10.5210/ojphi.v11i1.9685
Kim, E., Huang, K., Tomala, A., Matthews, S., Strubell, E., Saunders, A., . . . Olivetti,
E. (2017). Machine-learned and codified synthesis parameters of oxide
materials. Scientific Data, 4, 170127. doi:10.1038/sdata.2017.127
Klumpp, M., Zeng, L., Al-Thabaiti, S. A., Weber, A. P., & Schwieger, W. (2016).
Building concept inspired by raspberries: From microporous zeolite
nanocrystals to hierarchically porous assemblies. Microporous and
Mesoporous Materials, 229, 155-165. doi:10.1016/j.micromeso.2016.04.012
Koshy, N., & Singh, D. N. (2016). Fly ash zeolites for water treatment applications.
Journal of Environmental Chemical Engineering, 4(2), 1460-1472.
doi:10.1016/j.jece.2016.02.002
Kosinov, N., Gascon, J., Kapteijn, F., & Hensen, E. J. M. (2016). Recent developments
in zeolite membranes for gas separation. Journal of Membrane Science, 499,
65-79. doi:10.1016/j.memsci.2015.10.049
Kostinko, J. A. (1982). Factors influencing the synthesis of zeolites a, x, and y. Paper
presented at the American Chemical Society, Division of Petroleum Chemistry,
Preprints.
Kostinko, J. A. (1985). Preparation of sodium silicate solutions. In: US Patent
4539191.
Krauskopf, K. B. (1956). Dissolution and precipitation of silica at low temperatures.
Geochimica et Cosmochimica Acta, 10(1), 1-26.
doi:https://doi.org/10.1016/0016-7037(56)90009-6
Krongkrachang, P., Thungngern, P., Asawaworarit, P., Houngkamhang, N., & Eiad-
Ua, A. (2019). Synthesis of zeolite Y from kaolin via hydrothermal method.
Materials Today : Proceedings, 17, 1431-1436.
doi:10.1016/j.matpr.2019.06.164
Kubota, Y., Tawada, S., Nakagawa, K., Naitoh, C., Sugimoto, N., Fukushima, Y., . . .
Sugi, Y. (2000). Synthetic investigation of CIT-5 catalyst. Microporous and
Mesoporous Materials, 37(3), 291-301. doi:10.1016/S1387-1811(99)00272-3
Kuenzel, C., & Ranjbar, N. (2019). Dissolution mechanism of fly ash to quantify the
reactive aluminosilicates in geopolymerisation. Resources, Conservation and
Recycling, 150, 104421. doi:10.1016/j.resconrec.2019.104421
Kuroki, S., Hashishin, T., Morikawa, T., Yamashita, K., & Matsuda, M. (2019).
Selective synthesis of zeolites A and X from two industrial wastes: Crushed
stone powder and aluminum ash. Journal of Environmental Management, 231,
749-756. doi:10.1016/j.jenvman.2018.10.082
Längauer, D., Čablík, V., Hredzák, S., Zubrik, A., Matik, M., & Danková, Z. (2021).
Preparation of synthetic zeolites from coal fly ash by hydrothermal synthesis.
Materials, 14(5), 1-25. doi:10.3390/ma14051267
Li, H., Yang, S., Riisager, A., Pandey, A., Sangwan, R. S., Saravanamurugan, S., &
Luque, R. (2016). Zeolite and zeotype-catalysed transformations of biofuranic
compounds. Green Chemistry, 18(21), 571-5735. doi:10.1039/c6gc02415g
Li, T., Liu, H., Fan, Y., Yuan, P., Shi, G., Bi, X. T., & Bao, X. (2012). Synthesis of
zeolite Y from natural aluminosilicate minerals for fluid catalytic cracking
application. Green Chemistry, 14(12). doi:10.1039/c2gc36101a

Bibliography 225
Li, Y., Li, H., Xiao, L., Zhou, L., Shentu, J., Zhang, X., & Fan, J. (2012). Hemostatic
efficiency and wound healing properties of natural zeolite granules in a lethal
rabbit model of complex groin injury. Materials, 5(12), 2586-2596.
doi:10.3390/ma5122586
Li, Y., Li, L., & Yu, J. (2017). Applications of zeolites in sustainable chemistry.
Chemistry, 3(6), 928-949. doi:10.1016/j.chempr.2017.10.009
Lima, C. G. S., Moreira, N. M., Paixão, M. W., & Corrêa, A. G. (2019). Heterogenous
green catalysis: Application of zeolites on multicomponent reactions. Current
Opinion in Green and Sustainable Chemistry, 15, 7-12.
doi:10.1016/j.cogsc.2018.07.006
Limlamthong, M., Lee, M., Jongsomjit, B., Ogino, I., Pang, S., Choi, J., & Yip, A. C.
K. (2021). Solution-mediated transformation of natural zeolite to ANA and
CAN topological structures with altered active sites for ethanol conversion.
Advanced Powder Technology, 32(11), 4155-4166.
doi:10.1016/j.apt.2021.09.018
Lin, G., Zhuang, Q., Cui, Q., Wang, H., & Yao, H. (2015). Synthesis and adsorption
property of zeolite FAU/LTA from lithium slag with utilization of mother
liquid. Chinese Journal of Chemical Engineering, 23(11), 1768-1773.
doi:10.1016/j.cjche.2015.10.001
Liu, C., Gao, X., Ma, Y., Pan, Z., & Tang, R. (1998). Study on the mechanism of
zeolite Y formation in the process of liquor recycling. Microporous and
Mesoporous materials, 25(1-3), 1-6. doi:10.1016/S1387-1811(98)00149-8
Liu, H., Shen, T., Li, T., Yuan, P., Shi, G., & Bao, X. (2014). Green synthesis of
zeolites from a natural aluminosilicate mineral rectorite: Effects of thermal
treatment temperature. Applied Clay Science, 90, 53-60.
doi:10.1016/j.clay.2014.01.006
Liu, H., Shen, T., Wang, W., Li, T., Yue, Y., & Bao, X. (2015). From natural
aluminosilicate minerals to zeolites: synthesis of ZSM-5 from rectorites
activated via different methods. Applied Clay Science, 115, 201-211.
doi:10.1016/j.clay.2015.07.040
Liu, H., Yue, Y., Shen, T., Wang, W., Li, T., & Bao, X. (2019). Transformation and
crystallization behaviors of titanium species in synthesizing Ti-ZSM-5 zeolites
from natural rectorite mineral. Industrial & Engineering Chemistry Research,
58(27), 11861-11870. doi:10.1021/acs.iecr.9b01826
Liu, X. D., Wang, Y. P., Cui, X. M., He, Y., & Mao, J. (2013). Influence of synthesis
parameters on NaA zeolite crystals. Powder Technology, 243, 184-193.
doi:10.1016/j.powtec.2013.03.048
Liu, Y., Liu, S., Li, Y., & Cao, J. (2021). Rapid zeolization of potassic rocks by means
of a microwave‐assisted hydrothermal method. ChemistrySelect (Weinheim),
6(25), 6333-6338. doi:10.1002/slct.202100381
Liu, Z., Wang, J., Jiang, Q., Cheng, G., Li, L., Kang, Y., & Wang, D. (2019). A green
route to sustainable alkali-activated materials by heat and chemical activation
of lithium slag. Journal of Cleaner Production, 225, 1184-1193.
doi:10.1016/j.jclepro.2019.04.018
Liu, Z., Zhu, J., Peng, C., Wakihara, T., & Okubo, T. (2019). Continuous flow
synthesis of ordered porous materials: from zeolites to metal-organic
frameworks and mesoporous silica. Reaction Chemistry & Engineering, 4(1),
1699-1172. doi:10.1039/c9re00142e

Bibliography 226
Lu, Y. (2017). Industry 4.0: A survey on technologies, applications and open research
issues. Journal of Industrial Information Integration, 6, 1-10.
doi:10.1016/j.jii.2017.04.005
Luscombe, J. H. (2018). Thermodynamics. Boca Raton ;: CRC Press, Taylor & Francis
Group.
Ma, H., Yao, Q., Fu, Y., Ma, C., & Dong, X. (2010). Synthesis of Zeolite of Type A
from Bentonite by Alkali Fusion Activation Using Na2CO3. ACS Publications.
Ma, S., Shang, C., Wang, C. M., & Liu, Z. P. (2020). Thermodynamic rules for zeolite
formation from machine learning based global optimization. Chemical
Science, 11(37), 10113-10118. doi:10.1039/d0sc03918g
Machado, C., & Paulo, D. (2020). Industry 4.0 : challenges, trends, and solutions in
management and engineering. Boca Raton, FL: CRC Press/Taylor & Francis
Group.
Mahima Kumar, M., Senthilvadivu, R., Brahmaji Rao, J. S., Neelamegam, M., Ashok
Kumar, G. V. S., Kumar, R., & Jena, H. (2020). Characterization of fly ash by
ED-XRF and INAA for the synthesis of low silica zeolites. Journal of
Radioanalytical and Nuclear Chemistry, 325(3), 941-947.
doi:10.1007/s10967-020-07243-0
Mahima, U. G., Yatindra Patidar, Abhishek Agarwal & Kushall Pal Singh (2020).
Wine quality analysis using machine learning algorithms. Micro-Electronics
and Telecommunication Engineering, 106.
Majdinasab, A. R., Manna, P. K., Wroczynskyj, Y., van Lierop, J., Cicek, N., Tranmer,
G. K., & Yuan, Q. (2019). Cost-effective zeolite synthesis from waste glass
cullet using energy efficient microwave radiation. Materials Chemistry and
Physics, 221, 272-287. doi:10.1016/j.matchemphys.2018.09.057
Malekpour, A., Millani, M. R., & Kheirkhah, M. (2008). Synthesis and
characterization of a NaA zeolite membrane and its applications for
desalination of radioactive solutions. Desalination, 225(1-3), 199-208.
doi:10.1016/j.desal.2007.02.096
Maxwell, P., & Mora, M. (2019). Lithium and Chile: looking back and looking
forward. Mineral Economics, 33(1-2), 57-71. doi:10.1007/s13563-019-00181-
8
Meng, F., McNeice, J., Zadeh, S. S., & Ghahreman, A. (2021). Review of Lithium
Production and Recovery from Minerals, Brines, and Lithium-Ion Batteries.
Mineral Processing and Extractive metallurgy Review, 42(2), 123-141.
doi:10.1080/08827508.2019.1668387
Meshram, P., Pandey, B. D., & Mankhand, T. R. (2014). Extraction of lithium from
primary and secondary sources by pre-treatment, leaching and separation: A
comprehensive review. Hydrometallurgy, 150, 192-208.
doi:10.1016/j.hydromet.2014.10.012
Meshram, S. U., Khandekar, U. R., Mane, S. M., & Mohan, A. (2014). Novel route of
producing zeolite A resin for quality-improved detergents. Journal of
Surfactants and Detergents, 18(2), 259-266. doi:10.1007/s11743-014-1656-4
Miao, Q., Zhao, B., Liu, S., Guo, J., Tong, Y., & Cao, J. (2016). Decomposition of the
potassic rocks by sub-molten salt method and synthesis of low silica X zeolite.
Asia-Pacific Journal of Chemical Engineering, 11(4), 558-566.
doi:10.1002/apj.1980
Millar, G. J., Couperthwaite, S. J., & Alyuz, K. (2016). Behaviour of natural zeolites
used for the treatment of simulated and actual coal seam gas water. Journal of

Bibliography 227
Environmental Chemical Engineering, 4(2), 1918-1928.
doi:10.1016/j.jece.2016.03.014
Millar, G. J., Couperthwaite, S. J., Dawes, L. A., Thompson, S., & Spencer, J. (2017).
Activated alumina for the removal of fluoride ions from high alkalinit y
groundwater: New insights from equilibrium and column studies with
multicomponent solutions. Separation and Purification Technology, 187, 14-
24. doi:10.1016/j.seppur.2017.06.042
Millar, G. J., Winnett, A., Thompson, T., & Couperthwaite, S. J. (2016). Equilibr ium
studies of ammonium exchange with Australian natural zeolites. Journal of
Water Process Engineering, 9, 47-57. doi:10.1016/j.jwpe.2015.11.008
Moisés, M. P., da Silva, C. T. P., Meneguin, J. G., Girotto, E. M., & Radovanovic, E.
(2013). Synthesis of zeolite NaA from sugarcane bagasse ash. Materials
Letters, 108, 243-246. doi:10.1016/j.matlet.2013.06.086
Moliner, M., Román-Leshkov, Y., & Corma, A. (2019). Machine Learning Applied to
Zeolite Synthesis: The Missing Link for Realizing High-Throughput
Discovery. Accounts of Chemical Research, 52(10), 2971-2980.
doi:10.1021/acs.accounts.9b00399
Müller, P., Russell, A., & Tomas, J. (2015). Influence of binder and moisture content
on the strength of zeolite 4A granules. Chemical Engineering Science, 126,
204-215. doi:10.1016/j.ces.2014.12.031
Muraoka, K., Sada, Y., Miyazaki, D., Chaikittisilp, W., & Okubo, T. (2019). Linking
synthesis and structure descriptors from a large collection of synthetic records
of zeolite materials. Nature Communications, 10(1), 4459-4411.
doi:10.1038/s41467-019-12394-0
Murukutti, M. K., & Jena, H. (2022). Synthesis of nano-crystalline zeolite-A and
zeolite-X from Indian coal fly ash, its characterization and performance
evaluation for the removal of Cs+ and Sr2+ from simulated nuclear waste.
Journal of Hazardous Materials, 423. doi:10.1016/j.jhazmat.2021.127085
Nakhli, S. A. A., Delkash, M., Bakhshayesh, B. E., & Kazemian, H. (2017).
Application of Zeolites for Sustainable Agriculture: a Review on Water and
Nutrient Retention. Water, Air, and Soil pollution, 228(12), 1-34.
doi:10.1007/s11270-017-3649-1
Ojumu, T. V., Du Plessis, P. W., & Petrik, L. F. (2016). Synthesis of zeolite A from
coal fly ash using ultrasonic treatment - A replacement for fusion step.
Ultrasonics Sonochemistry, 31, 342-349. doi:10.1016/j.ultsonch.2016.01.016
Palcic, A., Sekovanic, L., Subotic, B., & Bronic, J. (2012). Zeolite a synthesis under
dynamic conditions, after hydrogel ageing. Croatica Chemica Acta, 85(3),
297-301. doi:10.5562/cca2055
Pan, T., Wu, Z., & Yip, A. (2019). Advances in the green synthesis of microporous
and hierarchical zeolites: a short review. Catalysts, 9(3).
doi:10.3390/catal9030274
Pastorello, F., & Troglia, C. (1987). Process for the manufacture of zeolites 4A having
a high crystallinity and a fine granulometry and being particularly suitable for
the formulation of detergent compositions. US Patent 4649036.
Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., . . . Chintala, S.
(2019). PyTorch: An Imperative Style, High-Performance Deep Learning
Library. Advances in neural information processing systems, 32.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., . . .
Duchesnay, E. (2011). Scikit-learn: Machine learning in Python. Journal of
Machine Learning Research. doi:10.5555/1953048.2078195

Bibliography 228
Peng, H., Seneviratne, D., & Vaughan, J. (2018). Role of the amorphous phase during
sodium aluminosilicate precipitation. Industrial and Engineering Chemistry
Research, 57(5), 1408-1416. doi:10.1021/acs.iecr.7b04538
Pina, M.O., Arruebo, M., Felipe, M., Fleta, F., Bernal, M.P., Coronas,J., Menéndez,
M., & Santamaría, J. (2004) A semi-continuous method for the synthesis of
NaA zeolite membranes on tubular supports, Journal of Membrane Science,
244 (1–2), 141-150. https://doi.org/10.1016/j.memsci.2004.06.049 Pineda, F.
J. (1987). Generalization of back-propagation to recurrent neural networks.
Physical Review Letters, 59(19), 2229-2232.
doi:10.1103/PhysRevLett.59.2229
Poole, C., Prijatama, H., & Rice, N. M. (2000). Synthesis of zeolite adsorbents by
hydrothermal treatment of PFA wastes: A comparative study. Minerals
Engineering, 13(8), 831-842. doi:10.1016/S0892-6875(00)00072-8
Poursaeidesfahani, A., Andres-Garcia, E., de Lange, M., Torres-Knoop, A., Rigutto,
M., Nair, N., . . . Vlugt, T. J. H. (2019). Prediction of adsorption isotherms
from breakthrough curves. Microporous and Mesoporous Materials, 277, 237-
244. doi:10.1016/j.micromeso.2018.10.037
Praagman, J. (1985). Classification and regression trees: Leo BREIMAN, Jerome H.
FRIEDMAN, Richard A. OLSHEN and Charles J. STONE The Wadsworth
Statistics/Probability Series, Wadsworth, Belmont, 1984, x + 358 pages. In
(Vol. 19, pp. 144-144): Elsevier B.V.
Probst, J., Couperthwaite, S. J., Millar, G. J., & Kaparaju, P. (2022a). Ammoniacal
nitrogen removal and reuse: Process engineering design and technoeconomics
of zeolite N synthesis. Journal of Environmental Chemical Engineering, 10(3).
doi:10.1016/j.jece.2022.107942
Probst, J., Couperthwaite, S. J., Millar, G. J., & Kaparaju, P. (2022b). Critical
evaluation of zeolite N synthesis parameters which influence process design.
Journal of Environmental Chemical Engineering, 10(5).
doi:10.1016/j.jece.2022.108347
Probst, J., Outram, J. G., Couperthwaite, S. J., Millar, G. J., & Kaparaju, P. (2021).
Sustainable ammonium recovery from wastewater: Improved synthesis and
performance of zeolite N made from kaolin. Microporous and Mesoporous
Materials, 316, 110918. doi:10.1016/j.micromeso.2021.110918
Raccuglia, P., Elbert, K. C., Adler, P. D. F., Falk, C., Wenny, M. B., Mollo, A., . . .
Norquist, A. J. (2016). Machine-learning-assisted materials discovery using
failed experiments. Nature, 533(7601), 73-76. doi:10.1038/nature17439
Raić, M., Mikac, L., Maric, I., Stefanic, G., Skrabic, M., Gotic, M., & Ivanda, M.
(2020). Nanostructured silicon as potential anode material for Li-Ion batteries.
Molecules 25(4), 891. doi:10.3390/molecules25040891
Raić, M., Mikac, L., Marić, I., Štefanić, G., Škrabić, M., Gotić, M., & Ivanda, M.
(2020). Nanostructured silicon as potential anode material for Li-ion batteries.
Molecules, 25(4), 891.
Rayalu, S., Udhoji, J. S., Munshi, K. N., & Hasan, M. Z. (2001). Highly crystalline
zeolite — a from flyash ofbituminous and lignite coal combustion. Journal of
Hazardous Materials, 88(1), 107-121.88(1), 107-121.
Reinoso, D., Adrover, M., & Pedernera, M. (2018). Green synthesis of nanocrystalline
faujasite zeolite. Ultrasonics Sonochemistry, 42, 303-309.
doi:10.1016/j.ultsonch.2017.11.034

Bibliography 229
Rios, C., Williams, C., & Fullen, M. (2009). Nucleation and growth history of zeolite
LTA synthesized from kaolinite by two different methods. Applied Clay
Science, 42(3-4), 446-454. doi:10.1016/j.clay.2008.05.006
Rodríguez-Iznaga, I., Rodríguez-Fuentes, G., & Petranovskii, V. (2018). Ammonium
modified natural clinoptilolite to remove manganese, cobalt and nickel ions
from wastewater: Favorable conditions to the modification and selectivity to
the cations. Microporous and Mesoporous Materials, 255, 200-210.
doi:10.1016/j.micromeso.2017.07.034
Roland, E. (1989). Industrial production of zeolites. In Studies in Surface Science and
Catalysis (Vol. 46, pp. 645-659): Elsevier.
Rosales, G., Ruiz, M., & Rodriguez, M. (2016). Study of the extraction kinetics of
lithium by leaching β-Spodumene with hydrofluoric acid. Minerals, 6(4).
doi:10.3390/min6040098
Rozhkovskaya, A., Rajapakse, J., & Millar, G. J. (2021). Synthesis of high-qualit y
zeolite LTA from alum sludge generated in drinking water treatment plants.
Journal of Environmental Chemical Engineering, 9(2), 104751.
doi:10.1016/j.jece.2020.104751
Sato, K., Aoki, K., Sugimoto, K., Izumi, K., Inoue, S., Saito, J., . . . Nakane, T. (2008).
Dehydrating performance of commercial LTA zeolite membranes and
application to fuel grade bio-ethanol production by hybrid distillation/vapor
permeation process. Microporous and Mesoporous Materials, 115(1-2), 184-
188. doi:10.1016/j.micromeso.2007.10.053
Selim, A. Q., Mohamed, E. A., Seliem, M. K., & Zayed, A. M. (2018). Synthesis of
sole cancrinite phase from raw muscovite: characterization and optimizat ion.
Journal of Alloys and Compounds, 762, 653-667.
doi:10.1016/j.jallcom.2018.05.195
Selvaratnam, B., & Koodali, R. T. (2020). Machine learning in experimental materials
chemistry. Catalysis Today. 371, 77-84
doi:https://doi.org/10.1016/j.cattod.2020.07.074
Sen, S., Datta, L., & Mitra, S. (Eds.). (2018). Machine learning and IoT: a biological
perspective. CRC Press.
Serra, J. M., Baumes, L. A., Moliner, M., Serna, P., & Corma, A. (2007). Zeolite
synthesis modelling with support vector machines: A combinatorial approach.
Combinatorial Chemistry and High Throughput Screening, 10 (1), 13-24.
doi:10.2174/138620707779802779
Setoudeh, N., Nosrati, A., & Welham, N. J. (2020). Lithium extraction from
mechanically activated of petalite-Na2SO4 mixtures after isothermal heating.
Minerals Engineering, 151. doi:10.1016/j.mineng.2020.106294
Shahzad, M. W., Burhan, M., & Ng, K. C. (2017). Pushing desalination recovery to
the maximum limit: Membrane and thermal processes integration.
Desalination, 416, 54-64. doi:10.1016/j.desal.2017.04.024
Shams, T., Schober, G., Heinz, D., & Seifert, S. (2021). Production of autoclaved
aerated concrete with silica raw materials of a higher solubility than quartz part
I: Influence of calcined diatomaceous earth. Construction and Building
Materials, 272. doi:10.1016/j.conbuildmat.2020.122014
Shan, Y., Sui, Z., Zhu, Y., Zhou, J., Zhou, X., & Chen, D. (2018). Boosting Size-
Selective Hydrogen Combustion in the Presence of Propene Using
Controllable Metal Clusters Encapsulated in Zeolite. Angewandte Chemie -
International Edition, 57(31), 9770-9774. doi:10.1002/anie.201805150

Bibliography 230
Shirazian, S., & Ashrafizadeh, S. N. (2015). Synthesis of substrate-modified LTA
zeolite membranes for dehydration of natural gas. Fuel, 148, 112-119.
doi:10.1016/j.fuel.2015.01.086
Silva, P.D., Sagoe-Crenstil, K., & Sirivivatnanon, V. (2007) Kinetics of
geopolymerization: Role of Al2O3 and SiO2, Cement and Concrete Research,
37, (512-518.
Smith, J. V. (1985). Book Review: Minerals and rocks 18. Natural zeolites. Glauco
Gottardi and Ermanno Galli, 1985. Springer-Verlag, Berlin, 409 pp. and 218
figures. $59.00. Geochimica et Cosmochimica Acta, 49(11), 2502-2502.
Steenland, K., & Ward, E. (2014). Silica: Aalung carcinogen. CA: a Cancer Journal
for Clinicians, 64(1), 63-69. doi:10.3322/caac.21214
Stewart, F. H. (1941). On sulphatic cancrinite and analcime (eudnophite) from Loch
Borolan, Assynt. Mineralogical magazine and journal of the Mineralogical
Society, 26(172), 1-8. etrieved from
https://rruff.info/doclib/MinMag/Volume_26/26-172-1.pdf
Strachowski, T., Grzanka, E., Mizeracki, J., Chlanda, A., Baran, M., Małek, M., &
Niedziałek, M. (2021). Microwave-assisted yydrothermal synthesis of zinc-
aluminum spinel ZnAl2O4. Materials, 15(1), 245. doi:10.3390/ma15010245
Retrieved from https://www.mdpi.com/1996-1944/15/1/245
Sun, S.-Y., Xiao, J.-L., Wang, J., Song, X., & Yu, J.-G. (2014). Synthesis and
adsorption properties of Li1.6Mn1.6O4 by a combination of redox
precipitation and solid-phase reaction. Industrial & Engineering Chemistry
Research, 53(40), 15517-15521. doi:10.1021/ie5004625
USGS Survey (2020). Lithium Statistics and Information. Retrieved from
https://www.usgs.gov/centers/nmic/lithium-statistics-and-information
Valpotic, H. (2017). Zeolite clinoptilolite nanoporous feed additive for animals of
veterinary importance: potentials and limitations. Periodicum Biologorum,
119(3), 159-172. doi:10.18054/pb.v119i3.5434
Van Dyke Parunak, H. (1998). Book review: Neural networks for pattern recognition
by Christopher M. Bishop (Clarendon Press, 1995). In (Vol. 9, pp. 41-43).
Van Houwelingen, H. C. (2004). The Elements of Statistical Learning, Data Mining,
Inference, and Prediction. Trevor Hastie, Robert Tibshirani and Jerome
Friedman, Springer, New York, 2001. No. of pages: xvi+533. ISBN 0-387-
95284-5. Statistics in. Medicine 23 (3) 527-528).
Victoria State Government (2023) Geology and exploration; zeolite.
https://earthresources.vic.gov.au/geology-exploration/minerals/industrial-
minerals/zeolite.
Vikström, H., Davidsson, S., & Höök, M. (2013). Lithium availability and future
production outlooks. Applied Energy, 110, 252-266.
doi:10.1016/j.apenergy.2013.04.005
Waajen, G., van Oosterhout, F., Douglas, G., & Lürling, M. (2016). Management of
eutrophication in Lake De Kuil (The Netherlands) using combined flocculant
– Lanthanum modified bentonite treatment. Water Research (Oxford), 97, 83-
95. doi:10.1016/j.watres.2015.11.034
Wang, B., Li, J., Zhou, X., Hao, W., Zhang, S., Chang, L., Wang, X., Wang, Z., Xu,
J., Jia-Nan, Z., Li, X., & Wenfu Yan. (2022). Facile activation of lithium slag
for the hydrothermal synthesis of zeolite A with commercial quality and high
removal efficiency for the isotope of radioactive 90Sr. Inorganic Chemistry
Frontiers, 9(3), 468–477. https://doi.org/10.1039/d1qi01492g

Bibliography 231
Wang, C., Boithias, L., Ning, Z., Han, Y., Sauvage, S., Sánchez-Pérez, J.-M., . . .
Hatano, R. (2017). Comparison of Langmuir and Freundlich adsorption
equations within the SWAT-K model for assessing potassium environmental
losses at basin scale. Agricultural Water Management, 180, 205-211.
doi:10.1016/j.agwat.2016.08.001
Wang, J.-Q., Huang, Y.-X., Pan, Y., & Mi, J.-X. (2014). Hydrothermal synthesis of
high purity zeolite A from natural kaolin without calcination. Microporous and
Mesoporous Materials, 199, 50-56.
doi:https://doi.org/10.1016/j.micromeso.2014.08.002
Wang, J.-Q., Huang, Y.-X., Pan, Y., & Mi, J.-X. (2016). New hydrothermal route for
the synthesis of high purity nanoparticles of zeolite Y from kaolin and quartz.
Microporous and Mesoporous Materials, 232, 77-85.
doi:10.1016/j.micromeso.2016.06.010
Wang, S., & Peng, Y. (2010). Natural zeolites as effective adsorbents in water and
wastewater treatment. Chemical Engineering Journal, 156(1), 11-24.
doi:10.1016/j.cej.2009.10.029
Wang, X., Hu, H., Liu, M., Li, Y., Tang, Y., Zhuang, L., & Tian, B. (2021).
Comprehensive utilization of waste residue from lithium extraction process of
spodumene. Minerals Engineering, 170. doi:10.1016/j.mineng.2021.106986
Wang, X., & Nguyen, A. V. (2016). Characterisation of electrokinetic properties of
clinoptilolite before and after activation by sulphuric acid for treating CSG
water. Microporous and Mesoporous Materials, 220, 175-182.
doi:10.1016/j.micromeso.2015.09.003
Wang, Y., Wu, Q., Meng, X., & Xiao, F.-S. (2017). Insights into the Organotemplate -
Free Synthesis of Zeolite Catalysts. Engineering, 3(4), 567-574.
doi:10.1016/j.Eng.2017.03.029
Weitkamp, J. (2000). Zeolites and catalysis. Solid State Ionics, 131(1), 175-188.
doi:10.1016/S0167-2738(00)00632-9
Wen, J., Dong, H., & Zeng, G. (2018). Application of zeolite in removing
salinity/sodicity from wastewater: A review of mechanisms, challenges and
opportunities. Journal of Cleaner Production, 197, 1435-1446.
doi:10.1016/j.jclepro.2018.06.270
Wernert, V., Schaef, O., Aloui, L., Chassigneux, C., Ayari, F., Ben Hassen Chehimi,
D., & Denoyel, R. (2020). Cancrinite synthesis from natural kaolinite by high
pressure hydrothermal method: Application to the removal of Cd2+ and Pb2+
from water. Microporous and Mesoporous Materials, 301.
doi:10.1016/j.micromeso.2020.110209
Whiting, G. T., Chung, S.-H., Stosic, D., Chowdhury, A. D., van der Wal, L. I., Fu,
D., . . . Weckhuysen, B. M. (2019). Multiscale Mechanistic Insights of Shaped
Catalyst Body Formulations and Their Impact on Catalytic Properties. ACS
Catalysis, 9(6), 4792-4803. doi:10.1021/acscatal.9b00151
Wruck, K., Millar, G. J., & Wang, T. (2021). Transformation of heulandite type natural
zeolites into synthetic zeolite LTA. Environmental Technology & Innovation,
21, 101371. doi:10.1016/j.eti.2021.101371
Wu, Q., Meng, X., Gao, X., & Xiao, F. S. (2018). Solvent-Free Synthesis of Zeolites:
Mechanism and Utility. Accounts of Chemical Research, 51(6), 1396-1403.
doi:10.1021/acs.accounts.8b00057
Xiao-Yong, Q., Qian, W., Hui-Jun, X., Jian-Chun, Z., & Qing-Yang, D. (2016). Effect
of the aging temperature on the synthesis of small crystal LTA zeolites.

Bibliography 232
Materials Research Innovations, 20(3), 161-164.
doi:10.1179/1433075X15Y.0000000039
Xiao, J., Li, C., Fang, L., Böwer, P., Wark, M., Bénard, P., & Chahine, R. (2020).
Machine learning–based optimization for hydrogen purification performance
of layered bed pressure swing adsorption. International Journal of Energy
Research, 44(6), 4475-4492. doi:10.1002/er.5225
Xiaotian, Z., Li, L., Yueqin, Q., & Wanshuang, L. (2022). Adsorption of phosphate by
cancrinite in red mud: a first-principles study. Materials Research Express,
9(4). doi:10.1088/2053-1591/ac5e20
Xu, M., Chen, S., Seo, D.-K., & Deng, S. (2019). Evaluation and optimization of
VPSA processes with nanostructured zeolite NaX for post-combustion CO2
capture. Chemical Engineering Journal, 371, 693-705.
doi:10.1016/j.cej.2019.03.275
Yan, Q., Li, X., Yin, Z., Wang, Z., Guo, H., Peng, W., & Hu, Q. (2012). A novel
process for extracting lithium from lepidolite. Hydrometallurgy, 121-124, 54-
59. doi:10.1016/j.hydromet.2012.04.006
Yang, J., Liu, H., Diao, H., Li, B., Yue, Y., & Bao, X. (2017). A quasi-solid-phase
approach to activate natural minerals for zeolite synthesis. ACS Sustainable
Chemistry & Engineering, 5(4), 3233-3242.
doi:10.1021/acssuschemeng.6b03031
Yang, S., Lach-hab, M., Blaisten-Barojas, E., Li, X., & Karen, V. L. (2010). Machine
learning study of the heulandite family of zeolites. Microporous and
Mesoporous Materials, 130(1-3), 309-313.
doi:10.1016/j.micromeso.2009.11.027
Yang, Y., Zhu, H., Xu, X., Bao, L., Wang, Y., Lin, H., & Zheng, C. (2021).
Construction of a novel lanthanum carbonate-grafted ZSM-5 zeolite for
effective highly selective phosphate removal from wastewater. Microporous
and Mesoporous Materials, 324. doi:10.1016/j.micromeso.2021.111289
Yiren, W., Dongmin, W., Yong, C., Dapeng, Z., & Ze, L. (2019). Micro-morphology
and phase composition of lithium slag from lithium carbonate production by
sulphuric acid process. Construction and Building Materials, 203, 304-313.
doi:10.1016/j.conbuildmat.2019.01.099
Yu, S., Kwon, S., & Na, K. (2018). Synthesis of LTA zeolites with controlled crystal
sizes by variation of synthetic parameters: Effect of Na+ concentration, aging
time, and hydrothermal conditions. Journal of Sol-Gel Science and
Technology. doi:10.1007/s10971-018-4850-4
Yue, Y., Hu, Y., Dong, P., Li, X., Liu, H., Bao, J., . . . Bao, X. (2020). Mesoscale
depolymerization of natural rectorite mineral via a quasi-solid-phase approach
for zeolite synthesis. Chemical Engineering Science, 220.
doi:10.1016/j.ces.2020.115635
Yue, Y., Liu, H., Yuan, P., Li, T., Yu, C., Bi, H., & Bao, X. (2014). From natural
aluminosilicate minerals to hierarchical ZSM-5 zeolites: A nanoscale
depolymerization–reorganization approach. Journal of Catalysis, 319, 200-
210. doi:10.1016/j.jcat.2014.08.009
Yue, Y., Liu, H., Yuan, P., Yu, C., & Bao, X. (2015). One-pot synthesis of hierarchical
FeZSM-5 zeolites from natural aluminosilicates for selective catalytic
reduction of NO by NH3. Scientific Reports, 5, 9270. doi:10.1038/srep09270
Yusupov, T. S., Shumskaya, L. G., & Kirillova, Y. A. (2000). State and perspectives
of natural zeolite beneficiation. Journal of Mining Science, 36(3), 299-304.
doi:10.1007/BF02562534

Bibliography 233
Zampori, L., Dotelli, G., Gallo Stampino, P., Cristiani, C., Zorzi, F., & Finocchio, E.
(2012). Thermal characterization of a montmorillonite, modified with
polyethylene-glycols (PEG1500 and PEG4000), by in situ HT-XRD and FT
IR: Formation of a high-temperature phase. Applied Clay Science, 59-60, 140-
147. doi:10.1016/j.clay.2012.02.015
Zhan, Y., Li, X., Zhang, Y., Han, L., & Chen, Y. (2013). Phase and morphology
control of LTA/FAU zeolites by adding trace amounts of inorganic ions.
Ceramics International, 39(5), 5997-6003.
doi:10.1016/j.ceramint.2013.01.005
Zhang, J., Mao, Y., Li, J., Wang, X., Xie, J., Zhou, Y., & Wang, J. (2015). Ultrahigh
mechanically stable hierarchical mordenite zeolite monolith: Direct binder -
/template-free hydrothermal synthesis. Chemical Engineering Science, 138,
473-481. doi:10.1016/j.ces.2015.08.016
Zhang, L., Wan, L., Chang, N., Liu, J., Duan, C., Zhou, Q., . . . Wang, X. (2011).
Removal of phosphate from water by activated carbon fiber loaded with
lanthanum oxide. Journal of Hazardous Materials, 190(1), 848-855.
doi:https://doi.org/10.1016/j.jhazmat.2011.04.021
Zhang, P., Li, S., & Zhang, C. (2019). Solvent-free synthesis of nano-cancrinite from
rice husk ash. Biomass Conversion and Biorefinery, 9(3), 641-649.
doi:10.1007/s13399-019-00375-8
Zhang, T., Ding, L., Ren, H., Guo, Z., & Tan, J. (2010). Thermodynamic modeling of
ferric phosphate precipitation for phosphorus removal and recovery from
wastewater. Journal of Hazardous Materials, 176(1), 444-450.
doi:10.1016/j.jhazmat.2009.11.049
Zhang, X., Tang, D., Zhang, M., & Yang, R. (2013). Synthesis of NaX zeolite:
Influence of crystallization time, temperature and batch molar ratio
SiO2/Al2O3 on the particulate properties of zeolite crystals. Powder
Technology, 235, 322-328. doi:10.1016/j.powtec.2012.10.046
Zhang, Y., & Ling, C. (2018). A strategy to apply machine learning to small datasets
in materials science. npj Computational Materials, 4(1). doi:10.1038/s41524-
018-0081-z
Zhang, Y., Zhong, S., Zhang, M., & Lin, Y. (2009). Antibacterial activity of silver -
loaded zeolite A prepared by a fast microwave-loading method. Journal of
Materials Science, 44(2), 457-462. doi:10.1007/s10853-008-3129-5
Zheng, R., Feng, X., Zou, W., Wang, R., Yang, D., Wei, W., . . . Chen, H. (2021).
Converting loess into zeolite for heavy metal polluted soil remediation based
on “soil for soil-remediation” strategy. Journal of Hazardous Materials, 412.
doi:10.1016/j.jhazmat.2021.125199
Zhuang, Q., Lin, G., Cui, Q., & Wang, H. (2014). Characterization and performance
of FAU/LTA co-crystalline zeolite synthesized by lithium slag. Shiyou
Xuebao, Shiyou Jiagong/Acta Petrolei Sinica (Petroleum Processing Section),
30(2), 348-352. doi:10.3969/j.issn.1001-8719.2014.02.024
Zijun, Z., Effeney, G., Millar, G. J., & Stephen, M. (2021). Synthesis and cation
exchange capacity of zeolite W from ultra-fine natural zeolite waste.
Environmental Technology & Innovation, 23, 101595.
doi:10.1016/j.eti.2021.101595

Bibliography 234
Appendices

Appendix A

Supplementary data for Chapter 3: Feasibility of Zeolite LTA Synthesis from


Spodumene Leachate Residue via Cancrinite Formation

Supplementary Table 1 XRF analysis of SLR


Oxide Mass % (Dry Basis) Oxide Mass % (Dry Basis)

SiO2 66.64 P2O5 0.04

Al 2O3 23.7 SO3 0.30

Fe 2O3 1.61 Cr2O3 0.03

Na2O 0.52 NiO 0.01

K2O 0.56 ZnO 0.01

MgO 0.99 PbO 0.01

CaO 0.74 BaO 0.06

TiO2 0.04 Mn3O4 0.17

LOI 5.12

Appendices 235
Supplementary Table 2 Quantitative XRD analysis of SLR
Phase SLR wt%

Hydrogen Aluminium Silicon Oxide 46.91

Quartz 3.93

Lithium Sulfate 0.30

Non-diffracting/unidentified 48.86

Sum 100

Supplementary Table 3 Crystalline cancrinite content of the products formed after


hydrothermal treatment of SLR material at 100℃
Phase wt% Crystalline

Quartz 3.97

Cancrinite 10.37

Sodalite 9.83

Zeolite LTA 1.69

Non-diffracting 74.24

Appendices 236
Supplementary Table 4 Crystalline cancrinite content of the products formed after
hydrothermal treatment of SLR material at 150℃
Reaction NaOH Quartz Non-
Cancrinite% Sodalite% ZeoliteA%
time (hrs) molarity % diffracting%

1 4 4.82 14.44 5.33 2.57 72.84

1 5 5.18 14.06 7.38 1.92 71.46

1 7 3.48 18.75 8.58 2.08 67.11

1 10 2.25 26.38 7.73 2.04 61.60

24 4 0.30 63.39 0.47 3.25 32.59

24 5 0.16 65.26 0.54 3.15 30.89

24 7 0.14 68.34 0.71 3.29 27.52

24 10 0.07 78.63 8.79 3.09 9.42

Appendices 237
Supplementary Table 5 Crystalline cancrinite content of the products formed after
hydrothermal treatment of SLR material at 200℃
Reaction NaOH Quartz Non-
Cancrinite% Sodalite% ZeoliteA%
time (hrs) molarity % diffracting%

1 5 0.72 37.93 1.25 3.16 56.94

1 7 0.30 43.89 1.56 4.30 49.96

1 10 0.00 50.97 5.51 2.69 40.84

1 14 0.00 14.61 10.48 1.89 73.03

15 5 0.34 42.42 1.19 3.63 52.42

15 7 0.21 41.49 1.48 3.62 53.20

15 10 0.38 37.72 2.49 3.16 56.26

15 14 0.00 21.11 15.76 0.11 63.01

24 4 0.13 67.78 0.15 3.57 28.38

24 5 0.08 70.69 0.00 3.68 25.55

24 7 0.03 66.91 0.13 3.38 29.55

24 10 0.16 70.59 0.14 3.40 25.72

Appendices 238
Supplementary Table 6 XRF analysis of Cancrinite
Oxide Mass % (Dry Basis) Oxide Mass % (Dry Basis)

SiO2 38.822 P2O5 0.035

Al 2O3 25.802 SO3 0.354

Fe 2O3 1.59 Cr2O3 0.034

Na2O 23.572 NiO 0.024

K2O 0.049 ZnO 0.007

MgO 1.024 PbO 0.013

CaO 0.767 BaO 0.051

TiO2 0.051 Mn3O4 0.146

LOI 9.48

Supplementary Table 7 XRF analysis of Zeolite LTA, activated at 2M HCl and


synthetised at 4M NaOH, 4 hours synthesis at 80℃
Oxide Mass % (Dry Basis) Oxide Mass % (Dry Basis)

SiO2 32.963 P2O5 0.012

Al 2O3 28.787 SO3 0.112

Fe 2O3 1.236 Cr2O3 0.022

Na2O 17.311 NiO 0.021

K2O 0.024 ZnO 0.006

MgO 0.77 Mn3O4 0.112

CaO 0.595 Cl 0.089

TiO2 0.023

LOI 17.9

Appendices 239
5

Volume Density (%)


3

0
0 100 200 300 400 500
Size Classes (um)

a) SEM Image b) Particle Size Distribution

Supplementary Figure 1 a) SEM imaging and b) particle size of SLR

100 100
1h Synthesis 1h Syntesis
24h Synthesis 24h Synthesis
wt% Crystalline Cancrinite

80 80
wt% Non-Diffracting

60 60

40 40

20 20

0 0
4 6 8 10 4 6 8 10
NaOH Molarity NaOH Molarity

wt% Crystalline Cancrinite (a) wt% Non-Diffracting/unidentified (b)

100 100
1h Synthesis 1h Synthesis
24h Synthesis 24h Synthesis
wt% Crystalline Sodalite

wt% Crystalline Quartz

80 80

60 60

40
40

20
20

0
0 4 6 8 10
4 6 8 10
NaOH Molarity
NaOH Molarity

wt% Crystalline Sodalite (c) wt% Crystalline Quartz (d)

Supplementary Figure 2 Influence of time and NaOH molarity in Cancrinite


synthesis from SLR at 150℃ a) Crystalline cancrinite, b) non-Diffracting material, c)
Crystalline Sodalite and d) Crystalline Quartz

Appendices 240
100 100
1h Synthesis 1h Syntesis
15hrs Synthesis 15 hrs Syntesis

wt% Crystalline Cancrinite


24hrs Synthesis 24 hrs Syntesis
80 80

wt% Non-Diffracting
60 60

40 40

20 20

0 0
4 6 8 10 12 14 4 6 8 10 12 14
NaOH Molarity NaOH Molarity

wt% Crystalline Cancrinite (a) wt% non-Diffracting/unidentified (b)

100 100
1h Synthesis 1h Synthesis
15hrs Synthesis 15hrs Synthesis
24hrs Synthesis
wt% Crystalline Sodalite

24hrs Synthesis

wt% Crystalline Quartz


80 80

60 60

40
40

20
20

0
0 4 6 8 10
4 6 8 10 12 14
NaOH Molarity
NaOH Molarity

wt% Crystalline Sodalite wt% Crystalline Quartz

Supplementary Figure 3 Influence of time and NaOH molarity in Cancrinite


synthesis from SLR at 200C, a) Crystalline cancrinite, b) non-Diffracting material, c)
Crystalline Sodalite, d) Crystalline Quartz

Appendices 241
Supplementary Figure 4 Energy required to produce cancrinite at 240°C and 1 hour
synthesis (Reactor1) vs energy required to produce cancrinite at 150 °C and 24 hours
synthesis.

Appendices 242
2 hours synthesis

4,000x 8,000x
3 hours synthesis

4,000x 8,000x
4 hours synthesis

4,000x 8,000x
5 hours synthesis

Appendices 243
4,000x 8,000x
6 hours synthesis

4,000x 8,000x
Supplementary Figure 5. SEM images of zeolite LTA after hydrothermal reaction
from dissolved cancrinite, reaction temperature 80℃, time= 2 to 6 hours

Appendices 244
Appendix B

Supplementary data for Chapter 5: Quench Method

Supplementary Figure 6. Flow diagram of Continuous Reactor Prototype

Appendices 245
Appendix C

Supplementary data for Chapter 7: Evaluation and application of machine


learning principles to Zeolite LTA synthesis

Supplementary Table 8. Candidate values for linear regression hyperparameters.


Fit Intercept True, False 2 candidates
Normalise True, False 2 candidates
Copy X True, False 2 candidates

Supplementary Table 9. Candidate values for ridge regression hyperparameters.


Fit Intercept True, False 2 candidates
Normalise True, False 2 candidates
Copy X True, False 2 candidates
Alpha −10
1 × 10 , … , 1 400 candidates

Supplementary Table 10. Candidate values for regression tree hyperparameters.


Maximum Tree Depth 5, 6, 7, 8, 9, 10 6 candidates
Minimum Samples Required to Split
5, …, 15 10 candidates
Internal Node
Minimum Samples Required at a Leaf Node 5, …, 15 10 candidates
Maximum Number of Features to Consider
Auto, Sqrt, Log2 3 candidates
when Evaluating Best Split
5, 10, 15, 20, 25,
Maximum Number of Leaf Nodes 6 candidates
30

Supplementary Table 11. Candidate values for random forest hyperparameters.


Number of Trees 10, 50, 100 3 candidates
Maximum Tree Depth None, 1, …, 10 11 candidates
Minimum Samples Required to Split
5, …, 15 10 candidates
Internal Node
Minimum Samples Required at a Leaf Node 5, …, 15 10 candidates
Maximum Number of Features to Consider
Auto, Sqrt, Log2 3 candidates
when Evaluating Best Split
5, 10, 15, 20, 25,
Maximum Number of Leaf Nodes 6 candidates
30

Supplementary Table 12. Candidate values for XGBoost hyperparameters.


Number of Trees 10, 50, 100 3 candidates

Appendices 246
Maximum Tree Depth None, 1, …, 10 11 candidates
0.0001, 0.001,
Learning Rate 3 candidates
0.01
Gamma (minimum loss reduction required to
0, 0.5, 1, 5, 15,
make a further partition on a leaf node of the 6 candidates
20
tree)
Minimum Child Weight (minimum sum of
1, …, 10 10 candidates
instance weight needed in a child)
Subsample (subsample ratio of the training
0.5, …, 1.0 5 candidates
instances)
Column Sample by Tree (subsample ratio of
0.1, …, 0.5 5 candidates
columns when constructing each tree)
Alpha (regularisation term) 0.1, …, 1.0 10 candidates
Lambda (regularisation term) 0.1, …, 1.0 10 candidates

Supplementary Table 13. Candidate values for sci-kit learn ANN hyperparameters.
Hidden Layer 1 5, …, 30 25 candidates
Hidden Layer 2 5, …, 30 25 candidates
Activation Identity, Logistic, Tanh, ReLU 4 candidates
Solver IBFGS, SGD, Adam 3 candidates
Alpha 0.0001, 0.001, 0.01 3 candidates
Learning Rate 0.0001, …, 0.01 Continuous
Number of Iterations 100, 500, 1000, 1500, 2000 5 candidates

Supplementary Table 14. Candidate values for PyTorch ANN hyperparameters.


Hidden Layer 1 5, …, 30 25 candidates
Hidden Layer 2 5, …, 30 25 candidates
Learning Rate 0.0001, …, 0.01 Continuous
100, 500, 1000, 1500,
Number of Iterations 5 candidates
2000

Appendices 247
Supplementary Table 15. Complete training and testing results reported as R^2
values. The results from each trial were obtained from a different training and testing
data split.
Average
Trial 1 Trial 2 Trial 3 Trial 4 Trial 5
Value
𝟎. 𝟑𝟔𝟖
Train 0.377 0.367 0.359 0.373 0.364
Linear ± 𝟎. 𝟎𝟎𝟗
Regression 𝟎. 𝟑𝟑𝟗
Test 0.320 0.359 0.323 0.343 0.350
± 𝟎. 𝟎𝟏𝟗
𝟎. 𝟑𝟕𝟐
Train 0.371 0.361 0.363 0.383 0.381
Ridge ± 𝟎. 𝟎𝟏𝟏
Regression 𝟎. 𝟑𝟒𝟎
Test 0.322 0.358 0.347 0.326 0.345
± 𝟎. 𝟎𝟏𝟖
𝟎. 𝟒𝟐𝟑
Train 0.411 0.417 0.435 0.422 0.431
Regression ± 𝟎. 𝟎𝟏𝟐
Tree 𝟎. 𝟑𝟖𝟓
Test 0.408 0.367 0.405 0.362 0.381
± 𝟎. 𝟎𝟐𝟑
𝟎. 𝟔𝟖𝟔
Train 0.677 0.681 0.698 0.704 0.668
Random ± 𝟎. 𝟎𝟏𝟖
Forest 𝟎. 𝟔𝟐𝟎
Test 0.639 0.632 0.608 0.601 0.618
± 𝟎. 𝟎𝟏𝟗
𝟎. 𝟕𝟒𝟏
Train 0.748 0.740 0.737 0.734 0.745
± 𝟎. 𝟎𝟎𝟕
XGBoost
𝟎. 𝟕𝟎𝟎
Test 0.707 0.697 0.693 0.700 0.705
± 𝟎. 𝟎𝟎𝟕
𝟎. 𝟖𝟖𝟕
Train 0.897 0.877 0.880 0.886 0.894
ANN ± 𝟎. 𝟎𝟏𝟎
(sklearn) 𝟎. 𝟖𝟒𝟎
Test 0.856 0.825 0.824 0.852 0.843
± 𝟎. 𝟎𝟏𝟔
𝟎. 𝟖𝟒𝟕
Train 0.859 0.848 0.842 0.835 0.852
ANN ± 𝟎. 𝟎𝟏𝟐
(PyTorch) 𝟎. 𝟖𝟑𝟓
Test 0.847 0.842 0.824 0.823 0.837
± 𝟎. 𝟎𝟏𝟐

Appendices 248
Supplementary Figure 7. Zeolite growth curve displaying the relationship between
nucleation and crystal growth rates and their contribution to the overall crystal
growth curve.

Supplementary Figure 8. Example of a regression tree trained for the zeolite LTA
synthesis system where “value” represents the predicted variable, the zeolite yield.

Appendices 249
Supplementary Figure 9. Learning curves for PyTorch ANN. The convergence of the
training and validation curves confirmed that the model is not overfitting.

Background Theory

The multidisciplinary nature of this study necessitates the need for background theory

on machine learning. The goal is to ensure all readers understand the machine learning

algorithms applied in the study. The following section provides derivations and

analysis of machine learning algorithms before the methodology describes the

application of the algorithms.

Principal Component Analysis

In a standard setting, PCA is applied to a space of N observations or instances of data

x1 , x2 , … , xN ∈ ℝM where each observation consists of M features or dimensions. PCA

selects a number n and obtains an n-dimensional representation such that

b1 , b2 , … , bN ∈ ℝn where n ≤ M and 𝐛i represents the observation 𝐱 i in a reduced

dimensionality space (M. A. Khan et al., 2019).

Appendices 250
To transform the vector 𝐱 to an n ≤ M-dimensional space, each observation

represented by 𝐛i is defined as the projection of 𝐱 i onto 𝐕 where 𝐕 is an orthonormal

basis v1 , … , v n of an n-dimensional subspace. The mean is subtracted from each

observation to ensure the projection is invariant under constant addition (M. A. Khan

et al., 2019). Supplementary Algorithm 1 provides the general format of the PCA

algorithm.

Supplementary Algorithm 1. PCA general algorithm


1
Mean calculation: 𝒎 = ∑𝑁𝑖=1 𝒙 𝑖
𝑁

̃ 𝑖 = 𝒙𝑖 − 𝒎 where 𝑥̃𝑖 is stored in an 𝑁 × 𝑀 matrix 𝑋̃


Subtract the mean: 𝒙

Projection: 𝒃𝑇𝑖 = ̃
𝒙𝑇𝑖 𝑽 where 𝑽 = (𝑣1 𝑣2 … 𝑣𝑛 )

When selecting the projection matrix 𝐕, the degree to which the projected observations

are spread out is measured by the variance of b1 , … , bN scaled by a factor of N −1 where

1
b̅ = ∑N
i=1 bi (Jolliffe, 2002).
N

𝐍 𝐍
𝟏 𝟐 𝟐
𝐖 = 𝐍 × 𝐕𝐚𝐫𝐢𝐚𝐧𝐜𝐞 [𝐛𝟏 , … , 𝐛𝐧 ] = 𝐍 ∑(𝐛𝐢 − 𝐛̅ ) = ∑(𝐛𝐢 − 𝐛̅)
𝐍
𝐢=𝟏 𝐢=𝟏
Supplementary Equation 1

The goal of PCA is to select 𝐯1 as the vector that maximises the spread of the projected

data: 𝐯1 = arg max


T
W where 𝐯1 is normalised. The solution to the maximisation
v v=1

̃T 𝐗
problem comes about by defining the matrix 𝐒 = 𝐗 ̃ and introducing the Lagrangian

multiplier λ to maximise the Lagrangian2. The solution to the maximisation problem

involves the selection of 𝐯1 , … , 𝐯n as the n orthonormal eigenvectors of 𝐒 with the n

largest eigenvalues using singular value decomposition (SVD) (Jolliffe, 2002).

Appendices 251
SVD computes the following three matrices for an N × M input:

σ1 0 ⋯ 0
0 σ2 0
⋮ ⋱ ⋮
∑= 0 0 ⋯ σM , 𝐔 = [u1 u2 ⋯ uN ], 𝐕 = [v1 v2 ⋯ vM ]
0 0 ⋯ 0
⋮ ⋮ ⋮
[0 0 ⋯ 0 ]

Supplementary Equation 2

SVD computes the matrices in Supplementary Equation 2 such that 𝐔 ∑ 𝐕 T = 𝐗 and

𝐕 T 𝐕 = 𝐈, 𝐔 T 𝐔 = 𝐈 and σ1 ≥ σ2 ≥ ⋯ ≥ σM are the singular values of 𝐗. The above

conditions reveal that the columns of 𝐕 are orthonormal (Jolliffe, 2002). Computing

𝐕 T 𝐯i reveals the following:

v1T v1T vi 0
⋮ ⋮ ⋮
𝐕 T 𝐯i = v iT v i = viT vi = 1 = ei
⋮ ⋮ ⋮
T
[v M ] T
[vM vi] [0 ]

Supplementary Equation 3

Supplementary Equation 4 demonstrates the following:

T T T
(𝐗 T 𝐗)𝐯i = (𝐕 ∑ 𝐔 T 𝐔 ∑ 𝐕 T ) 𝐯i = 𝐕 ∑ ∑ 𝐞i = 𝐕σ2i ei = σ2i 𝐯i

Supplementary Equation 4

Inferred from Supplementary Equation 4 is the relationship of 𝐯i . Each value of 𝐯i is

an eigenvector of 𝐗 T 𝐗 with an eigenvalue of σ2i . Relating this to the previous definition

of 𝐕n reveals that the first n columns of 𝐕 are the first n eigenvectors. The projection

Appendices 252
of a single observation 𝐱 Ti onto a subspace spanned by the first n PC is written as 𝐛Ti =

̃𝐕n.
𝐱 Ti 𝐕n or alternatively 𝐁 = 𝐗

The vectors 𝐛i in matrix, 𝐁 correspond to the coordination of the vector 𝐱 i when

projected onto an n-dimensional subspace. Variance explained by the PC is utilised to

determine the correspondence of 𝐛i = [bi1 bi2 ⋯ bin ]T to vectors in the original

space. The projected coordinates in the original space are 𝐱 ′i = bi1 𝐯1 + ⋯ + bin 𝐯n =

̃𝐕n 𝐕nT for all observations (Jolliffe, 2002).


𝐕n 𝐛i or 𝐗 ′ = (𝐕n 𝐁T )T = 𝐗

̃𝐕𝐕 T = 𝐗
Exploring the case of n = M and 𝐕n = 𝐕 obtains the relationship 𝐗 ′ = 𝐗 ̃𝐈 =

̃. This relationship dictates that no information is lost if all M principal directions are
𝐗

̃
applied to the projection. In this case, the projected matrix 𝐁 is still different from 𝐗

but corresponds to a direct rotation of all observations in M directions. When

̃ is rotated back without any information loss (Jolliffe, 2002).


translating back to 𝐗′, 𝐗

̃ will lose information if n < M. All


However, in practice, the reconstructed matrix 𝐗

̃, which is orthogonal to the space 𝐕n, is lost because those directions


variability of 𝐗

are projected away (Jolliffe, 2002). The variance explained is computed from the

Frobenius norm to determine how much information is retained in a reconstruction of

n PC.

‖𝐗 ′ ‖2F
Variance Explained = 2
̃‖
‖𝐗 F

Supplementary Equation 5

Appendices 253
Using the relationship ‖𝐗‖2F = trace(𝐗 T 𝐗) and exploiting relationships in SVD

provides the formula for the variance explained by the n first PC.

∑ni=1 σ 2i
Variance Explained = M 2
∑i=1 σ i

Supplementary Equation 6

When applying the SVD and computing the variance explained, the PCA algorithm

becomes:

Supplementary Algorithm 2. PCA with SVD and variance explained.


1
Mean calculation: 𝒎 = ∑𝑁𝑖=1 𝒙 𝑖
𝑁

Subtract the Mean: 𝒙


̃ 𝑖 = 𝒙𝑖 − 𝒎

𝑥̃ 𝑖𝑗 1
Standardise the data: 𝑥̃ 𝑖𝑗 = , where 𝑠𝑘 = √ ∑𝑁 ̃ 2𝑖𝑘
𝑖=1 𝑥
𝑠𝑘 𝑁−1

Compute the SVD: 𝑼 ∑ 𝑽𝑇 = 𝑿

̃ 𝑽𝑛
Project the observations onto a subspace as 𝒃𝑇𝑖 = 𝒙𝑇𝑖 𝑽𝑛 or 𝑩 = 𝑿

∑𝑛 𝜎𝑖2
Calculate the variance explained: ∑𝑖=1
𝑀 2
𝑖=1 𝜎𝑖

Linear Regression

Linear regression models the output prediction as a linear combination of all input

features where the M features are parameterised by weight coefficients, w and input

variables 𝐱 = ( x1 , … , xM ) T (Howard, 2007).

Appendices 254
y (𝐱, 𝐰) = w0 + w1 x1 + ⋯ + wM xM

Supplementary Equation 7

The regression model extends to problems with multiple input features by considering

linear combinations of fixed nonlinear functions where ϕj (𝐱) is defined as a basis

function. Limiting the index j to a value of M − 1, the total number of parameters

considered in the model will be equal to M (Howard, 2007).

M−1

y (𝐱, 𝐰) = w0 + ∑ wj ϕj (𝐱)
j=0

Supplementary Equation 8

The term w0 considers the fixed offset of a dataset as the bias parameter (Howard,

2007). The inclusion of an additional basis function, ϕ0 (𝐱) = 1, is convenient in that

the model becomes:

M−1

y (𝐱, 𝐰) = ∑ wj ϕj (𝐱)
j=0

Supplementary Equation 9

Supplementary Equation 9 is simplified by the addition of the following vectorized

terms.

w0 1
w1 ϕ1 (𝐱)
𝐰 = [ ⋮ ], 𝛟 = [ ]

wM−1 ϕM−1 (𝐱)

Supplementary Equation 10

Appendices 255
Combining Supplementary Equation 9 and Supplementary Equation 10, the

model becomes:

y (𝐱, 𝐰) = 𝐰 T 𝛟(𝐱)

Supplementary Equation 11

To solve the linear regression problem, the least-squares approach is applied. The

resulting cost function becomes an optimisation problem to minimise the function with

respect to model parameters (Howard, 2007; Van Houwelingen, 2004).

N N
1 2 1 2
ℰ M (w) = ∑ (y n − y(𝐱 n ,𝐰)) = ∑(y n − 𝐰 T 𝛟(𝐱n ))
N N
n=1 n=1

Supplementary Equation 12

To solve the cost function, the partial derivative of ℰ is obtained with respect to w

(Howard, 2007; Van Houwelingen, 2004).

N N
∂ℰ M 2 2
= ∑ yn 𝛟(𝐱n) T − [∑ 𝛟(𝐱 n)𝛟(𝐱 n)T ]𝐰 T
∂wn N N
n=1 n=1

Supplementary Equation 13

The gradient of the function is then:


∂ℰ
∂w1 2 2
(
∇ℰ M w = ) ⋮ = yn𝛟(𝐱 n)T − (𝛟(𝐱 n )𝛟(𝐱n )T )𝐰 T
∂ℰ N N
(∂wM )

Supplementary Equation 14

The optimal weights are found by setting the gradient function to zero, which returns

Supplementary Equation 15 (Howard, 2007; Van Houwelingen, 2004).

Appendices 256
−1
̃T 𝐗
𝐰 ∗ = (𝐗 ̃) ̃
𝐗T 𝐲

Supplementary Equation 15

̃ is defined as the design matrix where the elements


In Supplementary Equation 15, 𝐗

̃ nj = ϕj (𝐱 n) resulting in the construction of the following matrix


are given by X

(Howard, 2007).

ϕ0 (𝐱1 ) ϕ1 ( 𝐱1 ) ⋯ ϕM−1(𝐱1 )
̃ = ( ϕ0 (𝐱 2 ) ϕ1 (𝐱 2 ) ⋯
𝐗
ϕM−1 (𝐱 2)
)
⋮ ⋮ ⋱ ⋮
ϕ0 (𝐱 N ) ϕ1 (𝐱 N ) ⋯ ϕM−1 (𝐱 N)

Supplementary Equation 16

Ridge Regression

The application of ridge regression builds upon the standard linear model in

Supplementary Equation 9. However, it introduces regularisation as a general

approach for managing model complexity. Ridge regression utilises least-squares

regularisation with the standard linear model, which takes the following form

(Howard, 2007).

ℰ M(𝐰) + λℰ W (𝐰)

Supplementary Equation 17

Within linear regression, two potential issues can occur:

̃ T𝐗
𝐗 ̃ is unable to be inverted as a result of N ≤ M or the presence of many

̃.
linearly dependent rows in 𝐗

M is large relative to N, and the model overfits.

Appendices 257
To solve these issues, regularisation alters the cost function to have a stronger

preference for small weights by applying parameter shrinkage (Howard, 2007; Van

Houwelingen, 2004). To ensure the constant term in the regression model is not

affected by the regularisation, the cost function is considered as:

N
1 2 λ
ℰ λ (𝐰) = ∑(yn − 𝐰 T 𝛟(𝐱 n)) + 𝐰 T 𝐰
N N
n=1

Supplementary Equation 18

The magnitude of the weights is affected by the scaling of X, and therefore,

standardisation is applied according to Supplementary Equation 19.

N N
Xij − uj 1 1 2
̂
X ij = , uj = ∑ Xkj , ŝj = √ ∑(Xij − uj )
ŝ j N N−1
i=1 i=1

Supplementary Equation 19

Supplementary Equation 20 displays the resulting cost function where λ ‖𝐰‖2 is the

regularisation term and λ is the regularisation constant (Howard, 2007).

2
̂𝐰‖ + λ‖𝐰‖2 , λ ≥ 0
ℰ λ (𝐰) = ‖𝐲 − 𝐗

Supplementary Equation 20

To solve the ridge regression problem, the derivative is set to zero, which then provides

the formula for the optimal weights, 𝐰 ∗ (Howard, 2007).

dℰ λ 2
̂ T (𝐲 − 𝐗
= −𝐗 ̂ 𝐰) + 2𝐰
dw

Supplementary Equation 21

Appendices 258
𝐰 ∗ = arg min ℰ λ (𝐰) = (𝐗 ̂ + λ𝐈)−1 𝐗
̂T𝐗 ̂T𝐲
w

Supplementary Equation 22

6.4 Regression Trees

The goal of regression trees is the same as linear and ridge regression, predicting a

continuous output value y i based on input features xi . Hunt’s algorithm, displayed in

Supplementary Algorithm 3, is used to construct regression trees. Hunt’s algorithm

requires the following parameters (Praagman, 1985):

Initial tree T containing the root node.

Dr , the data associated with the current tree branch, initially the complete dataset.

Supplementary Algorithm 3. Regression tree.


if the stop criterion is met:

Add a leaf node to the tree, which assigns every observation the mean value

of the node in Dr :

1
y(r) = ∑ yi
N (r)
i∈r

else

Try different splits of Dr , compute the impurity for each split using the sum

of squares impurity measure and select the split Dr = Dv1 , ⋯ , Dvk with the

highest purity gain. Recursively call the method on Dv1 , ⋯ , Dvk.

end

Appendices 259
The model evaluates each new split by computing the average sum of squares error
∑i∈vyi
between the observation y i and the mean value y(v ) = (5).
N(v)

1 2
I(v) = ∑(yi − y(v))
N(v )
i∈v

Supplementary Equation 23

The model then applies Hunt’s algorithm. The stopping criteria occurs when the

impurity falls below a specified value.

Random Forest

Before introducing random forests, it is important to understand the concept of

bagging. Bagging begins with a dataset D of size N, which selects T new datasets

D1 , ⋯ , DT of size N ′ ≤ N by randomly subsampling D. When N ′ = N, each Dt is

sampled by N randomly selected points from the original data D with replacement. The

same points may occur many times in each Dt . The regression model is trained on each

T dataset to produce T different regressors combined to give a single regression model

(Breiman, 2001). This concept is explained visually in Supplementary Figure 10

Appendices 260
Supplementary Figure 10. Combining regression trees to form a random forest using
bagging.

Random forests are an application of bagging to regression trees. The original data is

used to produce T datasets. The standard regression tree algorithm is trained on each

sampled dataset to produce combined predictors using Supplementary Equation

24(Breiman, 2001).

T
1
y = ∑ f t (x)
T
t=1

Supplementary Equation 24

Regression trees often select the same split at each iteration creating trees of high

correlation. To avoid this, each step of Hunt’s algorithm should only consider splits

from m < M of the features selected randomly from all M features to produce less

correlated trees (Breiman, 2001).

XGBoost

XGBoost is a tree-based model that applies gradient boosting to regression trees to

achieve greater accuracy and faster computation time. XGBoost uses optimised

Appendices 261
gradient boosting through parallel processing, tree-pruning, and regularisation to avoid

overfitting and bias within machine learning models (T. Chen & Guestrin).

XGBoost build trees by minimising the following loss function:

1
ℒ (ϕ) = ∑ l(ŷ i ,yi ) + ∑ Ω (f k ) where Ω(f) = γT + λ‖w‖2
2
i k

Supplementary Equation 25

The term ∑i l(ŷi ,y i ) represents the loss function, commonly mean squared error,

calculated from the true and predicted values. The term ∑k Ω(f k ) contains the

regularisation function used to reduce the predictions insensitivity to individua l

observations which, can be adjusted to prevent overfitting. The omega term also

includes the leaf weights w, terminal nodes or leaves in a tree T and the user parameter,

γ, used to encourage tree pruning (T. Chen & Guestrin).

The goal is to find an optimised output value for the leaf to minimise Supplementary

Equation 25. The equation for weight optimisation is written below with the

corresponding loss function.

− ∑i∈Ij g i
wj∗ =
− ∑i∈Ij hi + λ

Supplementary Equation 26

2
T
1 (− ∑i∈Ij g i )
ℒ̃ ( t)(q) = − ∑ + γT
2 − ∑i∈Ij hi + λ
j=1

Supplementary Equation 27

Appendices 262
The computational time to enumerate through all the possible tree structures q

necessitates an alternative algorithm. A greedy algorithm starts from a single leaf and

adds branches to the tree over several iterations. In this algorithm, IL and IR are the

instance sets of the left and right nodes after the split. The loss reduction after the split

is given by the following equation by letting I = IL ∪ IR (T. Chen & Guestrin). This

formula is used in practice to evaluate the split candidates (T. Chen & Guestrin).

2 2
1 (− ∑i∈IL g i ) (− ∑i∈IR g i ) (− ∑i∈I g i )2
ℒsplit = [ + − ]−γ
2 − ∑i∈IL hi + λ − ∑i∈IR hi + λ − ∑i∈I hi + λ

Supplementary Equation 28

Artificial Neural Network

ANNs consist of information processing units called neurons connected to other

neurons by weighted connections. The neurons are organised in layers that feed to each

other. Information in the network is processed sequentially. First, the input vector is

given to the input layer and passed through the neural network layers and activation of

each neuron. Next, the activation is propagated through one or more hidden layers to

the output layer, consisting of neurons corresponding to coordinates of the output

vector. This process completes a forward pass through the network (Van Dyke

Parunak, 1998).

The forward pass of a neural network builds upon the standard linear model in

Supplementary Equation 9; however, where the function f (∙) is the identity for

regression.

Appendices 263
M

y (𝐱, 𝐰) = f ( ∑ wj ϕj (𝐱))
j=1

Supplementary Equation 29

The model is extended so the basis functions ϕj (x) are dependent on model parameters

which are adjusted with the weight coefficients throughout the model training (Van

Dyke Parunak, 1998). Like linear regression, neural networks use basis functions that

are a linear combination of all inputs where the combination coefficients are adaptive

parameters (Van Dyke Parunak, 1998). The basic neural network model is described

as a series of functional transformations. In this model, M is the number of neurons in

the input layer, and D is the neurons in the output layer (Van Dyke Parunak, 1998).

The neural network becomes a mapping of f: ℝM → ℝD which maps x to y = f(x). The

input processing unit or neuron for a neural network is described in Supplementary

Equation 30 for M linear combinations of input variables x1 , … , xD .

D
(1) ( 1)
aj = ∑ wij xi + wj0
i=1

Supplementary Equation 30

The subscript (M. A. Khan et al., 2019) indicates the current layer of the network and

(1)
the parameters associated with that layer. Here, the term wij corresponds to the

(1)
layer’s weights during the term wj0 corresponds to the biases of the neural network

layer (Howard, 2007). The quantity aj is a unit of activation. In a neural network, aj is

transformed by a differentiable nonlinear activation function h(∙) (Howard, 2007).

Appendices 264
zj = h (aj )

Supplementary Equation 31

The values from Supplementary Equation 31 correspond to the output of the hidden

units within a network. Similarly, the general linear model can be applied to give the

output unit activation of a neural network where k = 1, … , K and K are the number of

outputs (Van Dyke Parunak, 1998).

M
(2) ( 2)
ak = ∑ wkj zj + wk0
j=1

Supplementary Equation 32

As represented by the subscripts, Supplementary Equation 32 corresponds to the

second layer of a neural network. The output unit activations are transformed to give

the output of the network, yk (Van Dyke Parunak, 1998). The activation function for

the output unit is the identity such that:

yk = ak

Supplementary Equation 33

When combining the layers of the network from Supplementary Equation 30and

Supplementary Equation 32, the overall function becomes (Van Dyke Parunak, 1998):

M D
(2) ( ) ( ) ( )
y k (𝐱, 𝐰) = (∑ wkj h (∑ wij1 xi 1 2
+ wj0 ) + wk0 )
j=1 i=1

Supplementary Equation 34

Supplementary Equation 34 is evaluated through forward propagation. However,

similar to linear regression, an additional input variable defined as x0 = 1 is included

Appendices 265
so the bias parameters become integrated into the weight parameters such that the

overall function simplifies to (Van Dyke Parunak, 1998):

M D
( 2) ( 1)
yk (𝐱, 𝐰) = (∑ wkj h (∑ wij xi ))
j=0 i=0

Supplementary Equation 35

The solution to the neural network is very similar to both linear and ridge regression.

Given a training set of input vectors 𝐱 n and a corresponding set of target vectors 𝐭 n for

n = 1, … , N the network is trained by minimising the cost function (Van Dyke

Parunak, 1998).

N
1
ℰ (𝐰) = ∑‖𝐲 (𝐱 N ,𝐰) − 𝐭 n‖2
N
n=1

Supplementary Equation 36

𝐰 ∗ = arg min ℰ (𝐰)


w

Supplementary Equation 37

Once again, the optimal weights are found by obtaining the gradient function and

setting it to zero. This can be achieved through backpropagation, the message passing

scheme that sends information back and forth through the network (Van Dyke

Parunak, 1998). The backpropagation algorithm consists of two stages, the feed-

forward stage discussed previously and the backpropagation stage, where the

prediction error is computed, and the error is propagated backwards to adjust the

weights. The backpropagation algorithm uses deterministic optimisation to minimise

the squared error sum through the gradient descent method which, requires the

Appendices 266
computation of the activation function’s partial derivatives with respect to the network

weights (Van Dyke Parunak, 1998).

Supplementary Equation 38shows the evaluation of the cost function for each point

within the training set:

ℰ (𝐰) = ∑ ∇ℰ n(𝐰)
n=1

Supplementary Equation 38

The problem then becomes evaluating the gradient ∇ℰ n (𝐰). The derivative of ℰ (𝐰)

is dependent on the weight wji via the summed input aj to unit j (Pineda, 1987; Van

Dyke Parunak, 1998)

∂En ∂En ∂aj


=
∂wji ∂ai ∂wji

Supplementary Equation 39

∂En ∂aj
Where = δj referred to as errors and, knowing that aj = ∑i wji zi, = zi (Pineda,
∂ai ∂w ji

1987; Van Dyke Parunak, 1998)With this understanding, the equations are combined

to obtain:

∂En
= δj zi
∂wji

Supplementary Equation 40

Supplementary Equation 40 explains that the derivative required to solve the

minimisation problem is obtained by multiplying the value for δ at the output of the

Appendices 267
weight by z for the input of the weight (Pineda, 1987; Van Dyke Parunak, 1998). The

calculation of δ at the output is simply:

δk = y k − t k

Supplementary Equation 41

To obtain δ for the hidden units, partial derivatives are applied again (Pineda, 1987;

Van Dyke Parunak, 1998).

∂En ∂En ∂ak


δj ≡ =∑
∂ai ∂ak ∂aj
k

Supplementary Equation 42

This equation passes over all units k to which information and connections are

received from unit j. The backpropagation formula is then obtained through

substitution (Pineda, 1987; Van Dyke Parunak, 1998).

δj = h′ (aj ) ∑ wkj δk
k

Supplementary Equation 43

The backpropagation formula demonstrates the ability to obtain the value for δ from a

particular hidden unit by the propagation of δ values backwards from values higher

ranked in the network (Pineda, 1987; Van Dyke Parunak, 1998).

The general training method for a neural network with backpropagation is displa yed

in Supplementary Algorithm 4

Appendices 268
Supplementary Algorithm 4. General neural network training method.
Provide an input vector 𝐱 n to the network and apply forward propagation

using aj = ∑i wji zi to find the activation of all hidden and output units.

Evaluate δk for all output units using δk = y k − t k.

Apply backpropagation using δj = h′ (aj ) ∑k wkj δk to obtain the δj value

for each hidden unit present in the neural network.

∂En
Evaluate the derivative by = δj zi and utilise to solve the minimisat ion
∂w ji

problem for the network weights.

Various activation functions can be utilised in model training and development for a

neural network. The common activation functions for neural networks are shown in

Supplementary Figure 11.

Supplementary Figure 11. Neural network activation functions.

For each function in Supplementary Figure 11, the basic information processing ability

is similar; however, when training the model, the choice of activation function

becomes important as the gradients will differ in magnitude (Pineda, 1987; Van Dyke

Appendices 269
Parunak, 1998). Additionally, as part of the weight optimisation, different optimisers

are available, including:

Limited-memory Broyden–Fletcher–Goldfarb–Shanno Optimiser

Stochastic Gradient Descent

Adam Optimiser

Cross-Validation and Hyperparameter Tuning

To train the machine learning models, cross-validation and hyperparameter tuning are

used to increase the model accuracy. Cross-validation uses part of the data to fit the

model and part to test the model. This is desirable in cases like the zeolite synthesis

problem, with a small dataset limiting the ability to set aside a validation dataset (Van

Houwelingen, 2004).

K-fold cross-validation splits the data into K roughly equal-sized sets of data. For the

kth partition, the model fits the other K − 1 parts of the data. The prediction error is

calculated for the fitted model when predicting the kth part of the data. Cross-

validation calculates the estimated prediction error where κ: [1, … , N] → (1, … K) is an

indexing function that indicates the partition to which observation i is allocated by

randomisation and ̂f −k (x) is the fitted function computed without the kth partition

(Van Houwelingen, 2004).

N
1
CV (̂f ) = ∑ L(yi , ̂f −κ(i)(xi))
N
i=1

Supplementary Equation 44

Appendices 270
For a model characterised by a tuning parameter, α, the cross-validation prediction

error is:

N
1
CV (̂f ,α) = ∑ L(y i, ̂f −κ(i) (xi , α))
N
i=1

Supplementary Equation 45

Validation of Machine Learning Models

Before applying machine learning to the zeolite synthesis data, the machine learning

models were validated against a validation dataset to prove the accuracy and reliabilit y

of the model results.

The machine learning models developed for application in zeolite synthesis use

regression algorithms. The goal of machine learning is to predict the zeolite yield based

on synthesis descriptors and conditions. Therefore, a dataset with similar machine

learning goals was obtained to validate the models. The wine quality dataset has been

widely used in regression and classification machine learning (Cortez et al., 2009).

The data consists of physicochemical information about red and white wine samples.

When used in regression-based machine learning, the physicochemical parameters and

processing conditions are used to predict the quality of a wine sample (Cortez et al.,

2009). The wine quality dataset has similar machine learning goals to the zeolite

synthesis data; therefore, it is the most suitable dataset for validating the models.

The models were validated by comparing the developed model results with published

literature values. The results from the model validation are presented numerically in

Appendices 271
Supplementary Table 16. While the accuracy of the models differed slightly from the

published values, this was attributed to differing model architectures and expected

variation in the results from individual training runs, testing runs, and hyperparameter

tuning.

Supplementary Table 16. Model validation results.


Developed Published Model Error
Model Type Source
Model Accuracy Accuracy Metric
Linear (Gupta,
0.3630 0.3606 𝑅2
Regression 2018)
(Dahal et al,
Ridge Regression 0.3949 0.3869 MSE
2021)
( Dahal et al
Regression Tree 0.3790 0.3741 MSE
2021)
(Mahima et
Random Forest 0.3716 0.4000 MSE
al, 2020)
( Dahal et al
0.3990 0.4000 MSE
2021)
(Caballero,
sklearn ANN
Jojoa, &
0.4938 0.51 MAE
Percybrooks,
2020)
( Dahal et al
PyTorch ANN 0.3968 0.4000 MSE
2021)
(Caballero
0.4989 0.51 MAE
et al., 2020)

The results from the model validation are presented visually in Supplementary Figure

12. Published literature models reported both training and testing results using a

variety of error and accuracy metrics. While it is ideal to use the R 2 value for regression

problems, not all published studies reported this value. The error is presented in terms

of MSE for the testing data for consistency and comparison. Additionally, for the ANN

models, the mean absolute error (MAE) is provided in Supplementary Table 9.

Comparable values in the literature for the linear regression model were only available

Appendices 272
as an R 2 value and therefore are reported in Supplementary Table 9 but not in

Supplementary Figure 12.

Supplementary Figure 12. Machine learning error comparison with published model
data.

The results in Supplementary Table 16 and Supplementary Figure 12 prove the

accuracy and validity of the machine learning models developed in this study. In

addition, there is a high level of correlation between the developed and the published

model error values, therefore, validating the machine learning models for use on the

zeolite LTA synthesis data.

Appendices 273

You might also like