Volume 18 Number 21 7 June 2016 Pages 3737–3950

CrystEngComm
www.rsc.org/crystengcomm

PAPER
Peter G. Boyd and Tom K. Woo
A generalized method for constructing hypothetical nanoporous materials
of any net topology from graph theory
 CrystEngComm
View Article Online
PAPER View Journal | View Issue

A generalized method for constructing

hypothetical nanoporous materials of any net
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.

Cite this: CrystEngComm, 2016, 18,

3777 topology from graph theory†
Peter G. Boyd and Tom K. Woo*

Received 19th February 2016,
Accepted 21st March 2016
DOI: 10.1039/c6ce00407e
www.rsc.org/crystengcomm
Here we present a method for constructing hypothetical crystalline nanoporous materials, such as metal–
organic frameworks (MOFs), using a graph theoretical approach. The method takes as input the discrete
secondary (or structural) building units (SBUs) with defined connection points, and a desired 3-dimensional
net topology in the form of a labelled quotient graph. The hypothetical materials are constructed based on
the principle that using a labelled quotient graph obtained, for example from the reticular chemistry struc-
ture resource (RCSR), one can construct a net embedding in 3-D Euclidean space with an infinite number
of different representations. Thus, crystalline structures can be realized by manipulating a net's embedding
such that vertices of the net match the geometries of the desired SBUs. To demonstrate the methodology,
46 different network topologies (i.e. tbo, pcu), are used to build MOFs from the same pair of 4-coordinate
and 3-coordinate SBUs. We further show that the method can be used to generate hypothetical MOFs
where the most common realization of a net, called the barycentric representation, will not produce a via-
ble structure. When combined with a robust force field based geometry optimizer, the method can be
used to generate large and structurally diverse hypothetical databases for virtual screening purposes.

Introduction surface area. However in order to robustly identify these

Metal organic frameworks (MOF)s are a class of nanoporous
materials which are being considered for use in many practi-
cal applications. Their versatility in terms of structure and
chemistry can be traced back to their modular nature; build-
ing blocks composed of inorganic and organic clusters, typi-
cally referred to as secondary (or structural) building units
(SBU)s, assemble to form beautiful 2 or 3-dimensional highly
symmetric structures. MOFs can exhibit pores which possess
various shapes and chemistry giving rise to vastly different
properties. The near infinite possible combinations of these
SBUs and their organization in three-dimensional space gen-
erate an incredibly large search space for candidate materials.
In this light, computational methods have recently began tak-
ing the lead in terms of identifying candidate porous mate-
rials, in particular for gas sorption applications.1–3
The general strategy employed in these computational
screening studies is to identify trends in material activity,
based on materials descriptors such as pore size and internal
Centre for Catalysis Research and Innovation, Department of Chemistry and
Biomolecular Sciences, University of Ottawa, Ottawa, ON, K1N 6N5, Canada.
E-mail: twoo@uottawa.ca
† Electronic supplementary information (ESI) available: Details of TOPOS calcu-
lations, derivations of net embeddings for dia and qtz and hypothetical struc-
tures discussed in the text are provided in the CIF file format. See DOI: 10.1039/
c6ce00407e
trends, a huge number of diverse materials need to be sam-
pled. With the limited number of synthesized MOF materials,
efforts have been focused on enumerating a large number of
MOFs computationally. To do so requires sampling a seem-
ingly infinite search space of SBU combinations and orienta-
tions in silico which join to form a periodic lattice. If one can
abstract the SBUs in a MOF to discrete vertices connected to
each other by edges, the problem can be re-stated: how many
different possible infinite arrays of vertices of finite local con-
nectivity can be realized? The answer, is of course, infinite.
However despite the scale of the problem it was found that
in general, many realized crystal structures can be classified
in to a relatively small collection of vertices and edges
possessing high symmetry. This abstraction of chemical in-
formation in crystalline materials to a collection of vertices
and edges is not new, in fact it was pioneered in the 70's and
80's by Wells4 and O'Keeffe.5
These collections of vertices and edges are special types of
graphs called nets. It should be stressed that nets do not
specify a particular geometry, rather they describe an underly-
ing connectivity between vertices. As such a single net can en-
compass a number of different representations in Euclidean
space, making the concept useful as a classification tool for
crystalline materials.6 The number of possible unique nets,
i.e. ones which cannot be superimposed on each other by
some minor adjustment of edges and vertices, is a subject of

This journal is © The Royal Society of Chemistry 2016 CrystEngComm, 2016, 18, 3777–3792 | 3777

Paper
mathematical enumeration and is not upper bounded. How-
ever, databases of nets have sprung up such as the reticular
chemistry structure resource (RCSR), which contain many of
the nets important in classifying crystal structures.7 On its
surface, the database provides three dimensional representa-
tions of nets. The vertices and edges of each net are placed to
obtain the maximum possible symmetry achievable, which
was shown to be when vertices are at the centre of its edge-
connected neighbours.8 This is known as the net's equilib-
rium or barycentric placement. Fig. 1 demonstrates this for
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.
the diamond net, where a) shows the barycentric form and b)
shows a distortion in the node geometry, which lowers the
maximum achievable symmetry.
There have recently been two examples of computational
algorithms which exploit crystallographic nets to generate pe-
riodic materials. Wilmer et al.1 constructed a database of
137 000 hypothetical materials by combinatorially snapping
SBUs together, as the authors put it, like tinker toys. We have
implemented this methodology9,10 and found that one of the
limitations of this approach is that the when connecting
SBUs together, their relative orientations are pre-defined geo-
metric parameters and the approach can be very sensitive to
these parameters. This requires, in some cases, extensive
trial-and-error adjustment of orientation parameters to
achieve a periodic structure. Additionally, the structures gen-
erated by the program will always possess the same underly-
ing topology. This was evidenced in a recent study, which
demonstrated that of the 137 000 MOFs in the hypothetical
MOF database, the connectivity of each could be classified in
to only one of 6 underlying nets.11
More recently Martin and Haranczyk reported an algo-
rithm which assembles SBUs together using the barycentric
placement of nets as ‘blueprints’.12 In such manner, SBUs
Fig. 1 The diamond (dia) net in it's a) barycentric representation
where all edge lengths are equal, and b) a non-barycentric representa-
tion. The different edge lengths result in different geometries of the
nodes in the net.
3778 | CrystEngComm, 2016, 18, 3777–3792
View Article Online
CrystEngComm
with the correct connectivities are aligned to vertices within
the net representation. Following this orientation, either the
ideal symmetry operations are applied to the SBUs to attempt
to form the structure, or the SBUs are connected together it-
eratively until a unit cell is formed. This algorithm is an in-
novation in the field of hypothetical structure generation as
it circumvents the difficulty of parameterizing SBUs by hand,
and also affords the generation of potentially many more
MOFs of varying topology. However, to fully realize all possi-
ble structures a given set of SBUs can achieve, it would be
more advantageous to start from a more abstract definition
of the net.
Eon formally demonstrated that one can obtain a unique
representation of a net with deviations from its barycentric
placement using a series of distortion vectors called colattice
vectors.13 These vectors, along with a net's lattice vectors can
fully describe a three-dimensional embedding. Herein we re-
port a graph theoretical approach for obtaining hypothetical
MOFs from underlying nets. Our algorithm requires the same
information as described in the method of Martin and
Haranczyk,12 that is to say, the user must input a set of SBUs
with arbitrarily defined connection points and a desired net
topology. However our method utilizes the underlying net,
not the barycentric representation, to obtain hypothetical
MOF structures. This maximizes the number of achievable
hypothetical MOF structures from a given set of SBUs.
The paper is outlined as follows, the next section will in-
troduce some of the graph theoretic terminology used in the
remainder of this text, as well as define nets and labelled
quotient graphs which are fundamental concepts used in the
construction of hypothetical MOFs in this work. The Methods
section will discuss how a dimensionless labelled quotient
graph is embedded in Euclidean space, initially in its
barycentric form, and then in a form which supports SBUs of
arbitrary geometry. In the Results section we will demon-
strate the method in a number of applications; i) embedding
nets to fit SBUs with varying geometries, ii) construction of a
wide variety of structures from using only two SBUs and la-
belled quotient graphs from the RCSR, iii) building a MOF
using the labelled quotient graph of an unstable net, which
possess vertex collisions in its barycentric form and finally iv)
we construct hypothetical versions of MOF-210 and NU-110
to demonstrate the complexity of the problem and the poten-
tial of this method to build real structures. In addition, we
demonstrate the construction of record high surface area hy-
pothetical MOFs using SBUs taken from experimentally deter-
mined high surface area MOFs.
Background
As is the case in several studies presenting graph theory in
context of studying the topologies of crystal structures,13–17
these authors find it necessary to define some basic graph
theoretical terms to aid the reader in navigating the remain-
der of this work.
This journal is © The Royal Society of Chemistry 2016

CrystEngComm
A graph, G(V, E), consists of a set of vertices V and edges
E. Each edge, e ∈ E is terminated by two vertices u, v ∈ V or
one vertex in V in the case of loops. If the graph has a finite
number of vertices in the set V, then the graph is called fi-
nite. If each vertex u ∈ V contains a finite number of edges
incident upon it, the vertices are locally finite.
A walk in a graph G(V, E) is a traversal of edges {e1, e2,⋯, en}
∈ E from u to v in V.
A path between two vertices u, v ∈ V is a walk over a se-
quence of vertices and edges in G(V, E) such that no edge is
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.
visited twice.
Cycles are paths which the starting and ending vertex is
the same, however all other edges and vertices visited are
unique.
Connected graphs are graphs which there exists a path be-
tween any two vertices u, v ∈ V. All graphs discussed herein
are connected graphs.
Co-cycles are sets of edges in E, which upon removal, will
fragment a connected graph. Co-cycles are also called cuts of
the graph.
Trees are connected graphs but contain no cycles.
Automorphisms on a graph are permutations of the
graphs edges and vertices which preserve incidence. They can
be likened to symmetry operations in crystalline structures,
whose operation (rotation, reflection, translation, or a combi-
nation thereof) preserves atomic structure.
An isomorphsim between two graphs G and G′ describes a
one-to-one correspondence of their vertex and edge sets.
There exists a mapping which sends each vertex and edge in
G to a vertex and edge in G′.
The degree of a graph with locally finite vertices is the
number of edges incident upon each vertex. If a graph is of
degree four, then all vertices possess four neighbours.
A net, as described by Delgado-Friedrichs and O'Keeffe,14
is a ‘special kind of graph’. Nets are connected infinite
graphs with locally finite vertices. Moreover, nets are simple,
meaning they possess no edge loops or multiple edges
connecting two vertices. Nets are dimensionless, being fully
described by an infinite set of vertices and edges. A net is
considered n-periodic if there exists a group of automor-
phisms which act freely on the net (leave no vertex fixed) with
rank n. In other words, the net has n unique translational
symmetry elements. A discussion of these automorphsims,
called the translational group, can be found in Eon.13 In that
work, he demonstrates the method of embedding nets in the
general n-periodic case, however in most of the cases we will
discuss, the 3-periodic nets obtained from the RCSR7 will be
used. Thus, for the sake of simplicity we will focus our dis-
cussion on 3-dimensional embeddings of 3-periodic nets, and
the term net and 3-periodic net will be synonymous.
It is this concept of a net which is frequently used in stud-
ies to classify a particular material. For example, the iso-
reticular series of MOFs introduced by Yaghi and coworkers18
can all be classified into the net pcu. This is accomplished by
abstracting the Zn4O cluster as a single 6-connected vertex
and the linear organic dicarboxylate molecules as edges be-
This journal is © The Royal Society of Chemistry 2016
View Article Online
Paper
tween vertices. The definition of the net above infers an im-
portant concept; the vertex geometries and spatial distances
between neighbouring vertices are irrelevant when classifying
MOFs into a net. Topological analysis programs such as
Systre8 and TOPOS19 perform analyses on the incidence struc-
ture of the MOF's underlying vertices and edges to determine
isomorphism with a known net, ignoring the spatial geome-
tries that arise from chemical bonding information. Thus it
can be said there are an infinite number of possible
3-dimensional representations, or embeddings, of a net. Here
we refer to embeddings as what Delgado-Friedrichs and
O'Keeffe14 call faithful embeddings, that is, the net vertices
are assigned to nodes in Euclidean space, and the edges
assigned to lines connecting nodes such that there are no
line crossings nor overlapping nodes.
We will work with a finite graph called the labelled quo-
tient graph, which contains all the translationally equivalent
vertices and edges of a 3-periodic net. This is conceptually
analogous to the unit cell of a crystalline material. To ensure
a one-to-one relationship between quotient graphs and their
nets, it is necessary to assign labels and directions to the
edges of the quotient graph.15 Labels represent the transla-
tional periodicity of edges, with assignments in three dimen-
sional integer space (3). For example in Fig. 2a, the edge la-
beled e2 is associated with the label ‘100’ which indicates
that e2 connects vertex A to vertex B via by crossing the peri-
odic vector 100. Since 3-D vectors possess an end point and
an origin by definition, each edge in the labeled quotient
graph must be assigned a “+” or “−” direction to ensure a
one-to-one relationship between the edge incidence structure
and the direction of a vector embedded in Euclidean space.
We will differentiate an edge from a directed (or oriented)
edge by calling edges with orientation, arcs. An orientation of
the graph is a set of directions assigned to the edges of the
graph. For example the orientation of the graph in Fig. 2a de-
scribes the four edges as arcs, whose positive directions point
towards vertex B from vertex A. Like the nets they describe,
quotient graphs do not have a particular n-dimensional repre-
sentation; they are purely topological constructs. For example
a quotient graph of the 3-periodic diamond net can be fully
described by the set of directed edges from vertex A to vertex
B: (e1, 100), (e2, 100), (e3, 010), (e4, 001). A 2-dimensional
Fig. 2 a) The labelled quotient graph of dia, b) the labelled quotient
graph of qtz. By convention the arrows on each edge indicate the
positive orientation of the arc.
CrystEngComm, 2016, 18, 3777–3792 | 3779

Paper
representation of the quotient graph of dia is provided in
Fig. 2a.
Methods
The aim of this work is to construct a 3-D net from an ab-
stract labelled quotient graph such that when the 3-D net is
realized, the geometry of it's nodes will fit exactly with the ge-
ometry of the SBUs, whereupon superposition, will generate a
novel hypothetical crystalline material. For the sake of clarity,
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.
we will say vertices and arcs (or edges) are dimensionless ele-
ments of the labelled quotient graph, and their images in 3-D
Euclidean space are points and lines, respectively.
There are two main aspects to constructing a unit-cell rep-
resentation of an abstract labelled quotient graph i) construc-
tion of the metric tensor, which defines the translational
symmetry of the net (i.e. the unit cell), and ii) sending each
vertex and arc from the labelled quotient graph to a 3-D point
and line in the unit cell. The information required to perform
these two operations is all contained within the labelled quo-
tient graph and the following sections will demonstrate how
this information is extracted.
Periodic lattice representations of the arcs
It was shown13 that one can construct a matrix representa-
tion, B*, of the arcs in a labelled quotient graph such that,
Ω + = B*−1α (1)
where Ω + are the 3-D lines represented in the fractional coor-
dinates of the unit cell and α is a matrix corresponding to
the sum of arc labels in the matrix B*.
To construct B* and the corresponding α matrix for an ar-
bitrary labelled quotient graph with m arcs and n vertices,
one must first establish the basis of the graphs' cycle and co-
cycle (or cut) space. The cycle basis of a graph is an irreduc-
ible representation of all of its possible cycles. For connected
graphs such as all of those representing crystalline nets, the
basis of the cycle space contains a total of m − n + 1 vectors,
where each vector is a sum of the arcs (taking orientation
into consideration) which form that cycle. An example can be
seen for the graph of qtz in Fig. 2b, where a cycle is
highlighted in orange. Note, to be described as a cycle, the
highlighted orange path must ‘flip’ the orientation of the arc
e2, since it is pointing in the wrong direction. Thus the basis
cycle vector, represented in terms of the edges (e1, e2, e3, e4,
e5, e6) is (1, −1, 1, 0, 0, 0) since it only contains the first three
arcs in its path and the second entry is negated to indicate
the reversed orientation of that arc in the cycle.
Many of the labelled quotient graphs used to construct
MOFs are much larger than the examples shown in Fig. 2
(some with over 100 arcs!), so to systematically construct the
cycle basis in these graphs, we use a minimum spanning tree
algorithm20–23 from the SAGE mathematical package.24 Upon
forming a minimum spanning tree, which defines the mini-
mum number of edges that connect all vertices in a graph, a
3780 | CrystEngComm, 2016, 18, 3777–3792
View Article Online
CrystEngComm
cycle basis can be realized for a labelled quotient graph by
adding the edges not included in the minimum spanning tree.
We are only half done, to construct a 3-D representation
of a labelled quotient graph, B* must unambiguously map all
of the arcs to 3-D lines, and is therefore an invertible m × m
matrix. The cycle basis is only of rank m − n + 1, thus the
remaining n − 1 rows of the matrix consist of the graphs' co-
cycle space. Here we will say that the co-cycle space consists
of all outward oriented arcs from the first n − 1 vertices of the
graph. An example of a co-cycle basis vector for qtz in Fig. 2b
is to represent all the outward oriented arcs of vertex ‘A’, if
necessary ‘flipping’ orientations in the same fashion as was
done for the cycle bases. Thus one co-cycle vector for the la-
belled quotient graph of qtz is (1, 0, −1, −1, 0, 1) representing
the outgoing arcs from ‘A’; e1, ‘flipped’ e3, ‘flipped’ e4, and
e6. We construct the co-cycle basis in a relatively straight for-
ward manner, by iterating over the first n − 1 vertices and tak-
ing their outward oriented arcs as the co-cycle basis vectors.
We map the basis vectors in B* to the 3-D lattice space by
defining their lattice orientations in the m × 3 matrix α. For
the cycle basis vectors, this is accomplished by summing the
labels associated with the arcs in that vector. Thus the first
m − n + 1 rows of the matrix α are called ‘lattice vectors’ and
contain information about what periodic image the corre-
sponding cycle in B* projects into. For example the lattice
vector associated with the orange cycle of qtz in Fig. 2b would
be a summation of the labels of +e1, −e2, and e3 giving.
. Conceptually, this cycle mapping can be
described by a walk of the net starting at vertex ‘A’, traversing
e1 to vertex ‘B’, traversing –e2 to vertex ‘C’, and finally tra-
versing e3 to arrive at an image of vertex ‘A’ in the periodic
direction 001.
We define the co-lattice vectors as the remaining n − 1 en-
tries of α. These correspondingly map the co-cycle basis vec-
tors in B* to the lattice space. Importantly, these co-lattice vec-
tors can be considered a deviation of an embedded point from
the centre of mass of its neighbouring points. Eon13 demon-
strated that if one maps the cycle basis to the lattice vectors
and the co-cycle basis to a set of zero vectors (L* = {000}), one
obtains the unique barycentric representation of a net, a repre-
sentation seen in the RCSR7 as well as publications discussing
nets in the context of MOF classification.6,25 We leave the de-
tails of identifying the cycle, co-cycle basis and mapping the
two graphs shown in Fig. 2 to their barycentric representations
in the ESI.† It was also demonstrated that one can obtain an
arbitrary, unique, representation of a net by setting the co-
cycle basis to a set of non-zero vectors (L* ≠ {000}), which is
the general strategy employed in this work to obtain net em-
beddings supporting the geometries of a set of SBUs. To our
knowledge, no other program is this generalized in
constructing hypothetical MOFs or other materials.
The metric tensor
We have defined the arcs as lines and the vertices as points
in the lattice space. To construct a representation of a net in
This journal is © The Royal Society of Chemistry 2016

CrystEngComm
Cartesian space for the purposes of building hypothetical
MOFs, we must construct a representation of the lattice in
3. The metric tensor, Z, thus completes this description,
representing a mapping from the lattice coordinate system
defined by the integer vectors (100), (010), and (001) to some
other basis in three dimensional space. It is characterized by
the dot products of the lattice basis vectors,
(2)
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.
One can assign arbitrary values to the vectors a b and c to
yield isomorphic representations of the net in 3, however, it
was shown13 that a metric tensor of the net can be
constructed from the labelled quotient graph, implicitly as-
suming a barycentric placement of the vertices and arcs in
3. Let be a 3 × m matrix containing the set of cycle vec-
tors which form a lattice basis of the 3-periodic net. The met-
ric tensor can be expressed by creating an orthogonal projec-
tion in the edge space, , whose kernel is a basis of the
edge space with labels that sum to (000) such that,
(3)
The projection matrix ensures that upon constructing
the metric tensor from the edge space, the cycles whose edge
labels sum to zero are indeed projected to the null vector.
The kernel, K, of the projection contains a basis of the cycles
whose edge labels sum to (000) as well as the co-cycle vectors.
The orthogonal projection is then constructed according to
Godsil and Royle:26
(4)
In cases of quotient graphs whose 3-periodic net is mini-
mal (their cycle basis is of rank 3), the projection is reduced
to the identity, Im. For example, the dia net has 3 vectors as a
basis of its cycle space, as described above and is thus mini-
mal in 3-periodic representations. We can therefore construct
its metric tensor (Z) using the 3 cycles forming the basis, and
eqn (2)
(5)
This indicates 3 lattice vectors of equal length, , and
angles between vectors of 60° which describes the primitive
cell of dia. For nets not minimal in 3-periodic representations
such as qtz, which has a cycle basis of rank 4, K is
constructed as follows. From the matrix α we discover which
This journal is © The Royal Society of Chemistry 2016
View Article Online
Paper
cycles correspond to the lattice directions (100), (010), and
(001). In the qtz graph these are L1 = (e3 − e4), L2 = (e1 − e6),
and L3 = (e4 + e6 + e5). The lattice vectors obtained from the
remaining cycles which form the basis of the cycle space can
then be expressed as a linear combination of L1, L2, and L3.
For qtz, a single remaining cycle L4 = (e2 − e5), whose sum of
edge labels is can be expressed as a linear combination
of the first two lattice vectors, L4 = (−L1 − L2). Thus one can
construct a cycle vector for the kernel K by summing the
above cycles = L4 + L1 + L2. The cycle, expressed in the
edge space, (e1, e2, e3, −e4, −e5, −e6) along with the two co-
cycle vectors, complete the kernel of the projection.
(6)
Where the first row represents the cycle , and the last two
rows are the co-cycle vectors, representing outward oriented
arcs incident on vertices A and B. Using eqn (3) the projec-
tion is then,
(7)
Using the projection matrix above and eqn (3), the
metric tensor for the qtz net is
(8)
Using the metric tensor and the lattice representations of
the lines (Ω +) a Cartesian barycentric representation of these
nets can be defined. The Cartesian representations for the la-
belled quotient graphs for dia and qtz are shown in Fig. 3
where the a lattice vector is placed along the x-axis and the b
vector is fixed in the x–y plane.
This section has demonstrated the method in which one
can obtain a representation of a labelled quotient graph as a
net in Cartesian space. We generated the barycentric embed-
ding of the nets dia and qtz could be found from knowing
the lattice (L) vectors and setting the co-lattice (L*) vectors
equal to zero. However we are interested in using these nets
as structure directing, subject to the local geometries
governed by the chemistry in SBUs. This will require isomor-
phic adjustments of the metric tensor and deviations of co-
lattice vectors outside of their barycentric values. Subsequent
sections will show how to obtain non-barycentric representations
CrystEngComm, 2016, 18, 3777–3792 | 3781

Paper
Fig. 3 a) A barycentric embedding of the labelled quotient graph of
dia, b) a barycentric embedding of the labelled quotient graph of qtz.
Each was constructed with the algorithm.
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.
from labelled quotient graphs, such that local geometries of
embedded nets match those of desired SBUs.
Constructing non-barycentric net embeddings based on SBU
geometries
Given a set of SBUs, the challenge is to decide upon the pos-
sible crystalline structures they can self-assemble into. This
will be governed by the particular choice of net, which dic-
tates the relative orientations of these SBUs in the periodic
lattice. Nets are chosen if there is a match between vertex de-
gree and the number of connection sites on each SBU. How-
ever, once the net is chosen, the geometries of the SBUs will
dictate the final embedding of the net. To accomplish this,
we construct the necessary matrices as prescribed above; de-
fining the cycles, cocycles and metric tensor of its labelled
quotient graph. We now define a bond between two adjacent
SBUs as a single arc in the labelled quotient graph. Embed-
ded arcs (lines) can now be compared with the SBU connec-
tion sites in terms of their length and the angles they make
with other lines. In general, the set of inner products of an
SBU with n connection sites can be represented by the set
SSBU = {vp·vq|(1 ≤ p ≤ n), (q ≤ p)} (9)
To compare these values with a specific Euclidean embed-
ding of a net, one can construct the m × m inner product ma-
trix of its lines using the following equation,
g = Ω +·Z·(Ω +)t (10)
which is a symmetric matrix whose diagonal entries are the
squared lengths of the lines, gii = ‖i‖2 and off-diagonal entries
can be considered the angular component between two dif-
ferent lines, i and j gij = ‖i‖‖ j‖ cos θ. This affords a representa-
tion of the line geometries without the need to represent the
vectors in a Cartesian coordinate system. It is then trivial to
discern which elements of the net's inner product matrix cor-
respond to the local geometry of the points and lines associ-
ated with each SBU.
We formally define an objective function; for each SBU
with k inner products in its set, SSBU, we minimize,
(11)
3782 | CrystEngComm, 2016, 18, 3777–3792
View Article Online
CrystEngComm
which is the sum squared differences between the inner
products of the SBU geometry and the associated inner prod-
ucts (gjk) of the lines in the embedded net. At each step of
the optimization, we construct an embedding of the net by
combining two sets of variables, the first is a set of co-lattice
vectors, L*(X), in 3.
L*(X) = {(xi1, xi2, xi3)|1 ≤ i ≤ n − 1} (12)
Recall the co-lattice vectors are included in the last n − 1
entries of the matrix, α. This affords a unique embedding of
the net in the lattice space. From eqn (1);
Ω +(X) = B *−1α(X) (13)
As an example, consider the simple case of the dia net,
discussed in the previous section. There is a single co-lattice
vector, which we define a set of three variables x1 = (x11, x12, x13),
giving
(14)
The second set of variables are the parameters of the unit
cell P = (a, b, c, α, β, γ) used to define the metric tensor such
that;
(15)
Upon initialization, the program takes as input the vari-
ables in X and P corresponding to the barycentric representa-
tion. That is, X = {xi = (0,0,0)|1 ≤ i ≤ n − 1} and the values of
the cell parameters are extracted from the barycentric deriva-
tion of the metric tensor. However, in cases where an unsta-
ble net is used (here, the barycentric placement can have
overlapping nodes) a small random perturbation from the
barycentric placement of vertices can be introduced to pre-
vent undefined inner products in the matrix g.
Because this is a non-linear optimization problem, the
best general approach is to select a global optimization rou-
tine to optimize the objective function in eqn (11). However
in many cases, the SBU geometries do not deviate signifi-
cantly from the barycentric representation of the net, thus for
the purposes of computational expedience, a local optimiza-
tion was implemented using an L-BFGS algorithm.27,28 Both
a L-BFGS local optimizer and a global controlled random
search (CRS)29 were tested. However, we did not find any in-
stances where a global optimizer was required, and in all
cases the resulting net embeddings were the same when ap-
plying the global and local optimization routines. Since the
local optimizer was found to be at least 5 times faster, we
This journal is © The Royal Society of Chemistry 2016

CrystEngComm
opted for the L-BFGS optimizer. In favour of using a global
optimizer; one can perceive cases where the net is very large,
such that significant local geometric errors in the net geome-
try would not impact the relative value of eqn (11). In these
cases a local minima could result in a net embedding with
distorted nodes, yielding a potentially non-physical structure.
As an example, consider the two SBUs in Fig. 4, an
octatopic Cd2+ inorganic SBU abbreviated as Cd4(O2CR)8 (ref.
30) and a ditopic diphenyl organic SBU. The net bcu contains
a single vertex of degree 8, which matches the degree of the
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.
inorganic SBU in Fig. 4a. Likewise, the linear ditopic organic
SBU shown in Fig. 4b can be assigned to the edges of the
graph. Here we will demonstrate how the method described
above will embed the net bcu based on the geometries of the
SBUs in Fig. 4.
The initial step is to assign SBUs to specific vertices within
the labelled quotient graph. This can be accomplished by
matching vertex and SBU coordination numbers. In the case
of bcu, this task is straight forward; there is a single node
with eight edges incident upon it. Thus the 8-coordinate Cd2+
SBU will be assigned to the vertex. In order to account for
particular bonding geometries of SBUs, the labelled quotient
graphs are augmented with more vertices and edges. Specifi-
cally, edges are cut between each original vertex to include an
additional two vertices of degree 2. The extra edge between
these vertices represents the bond made between adjacent
SBUs, to account for bond orientations which cannot be rep-
resented by a single edge originating from the SBUs centre of
mass. An example of this can be seen in Fig. 4a for
Cd4(O2C)8. The geometry of the SBU and its connection sites
can best be described by a vertex of degree 8 located at the
SBU centre of mass, and 8 vertices of degree 2 which describe
the bonding orientation of the carboxylate moieties.
It should be noted that vertices of degree 2 are typically
not found in crystallographic nets, as they can be substituted
for a single edge between vertices of a higher degree. How-
Fig. 4 a) An octatopic inorganic SBU, Cd4(O2CR)8, along with a b)
linear organic SBU and their underlying vertex representations. Notice
that due to the bonding of the carboxylate groups to the Cd2+ ions,
the underlying representation in a) is best described by 9 points and 16
lines; a single (black) node at the centre of mass of the SBU, and
8 (orange) points located at its connection sites.
This journal is © The Royal Society of Chemistry 2016
View Article Online
Paper
ever, in the case of structure generation, we find it necessary
to include these vertices to ensure a good geometric match
with the SBUs. For example, all of the zeolite topologies pos-
sess quotient graphs with vertices of degree 4, which repre-
sent the tetrahedral Si4+ ions within the cell. However in or-
der to accurately describe the ‘bent’ 145° angle of the
connecting oxygen atoms, these bridging vertices must be
inserted between vertices of degree 4. By inserting these verti-
ces of degree 2, we do not alter the topology of the underlying
graph. This is ensured by assigning a label of ‘000’ to any
new edges such that the sum of edge labels in the cycle basis
yields the same value as the cycle in the un-augmented form.
A specific example is shown in Fig. 5. The labelled quotient
graph is augmented with several new vertices and edges to
account for the linear organic SBU and the bond between the
inorganic and organic SBUs.
We are now faced with the task of assigning the arcs of
the labelled quotient graph to specific SBU connecting sites.
We note that there are many possible ways to assign arcs to
SBU connection sites, particularly when the symmetry of the
SBU is low. The work of Schmid and coworkers have
performed genetic algorithm optimizations to find the lowest
energy (and therefore probable) SBU orientations within a
net representation, this was used in the context of explaining
why a particular nano-porous material will assemble to form
one net over another.31,32 In this work arc assignment is
based on a geometric comparison of the inner products of
the SBU vectors with the barycentric placement of the net
Fig. 5 Augmenting the labelled quotient graph to support the SBU
connectivities. a) The original quotient graph of bcu and b) it's
augmented form. The cycle spanning e1 in the original graph (loop) is
now a cycle spanning e1 + e5 + e6 + e7 + e8 + e9 in the augmented
graph, both cycles edge labels sum to ( ). The orange vertices
represent connection points between adjacent SBUs and the black
vertices represent the centre of mass of the SBUs. The additional
edges included in the augmented graph are all given a label of 000.
CrystEngComm, 2016, 18, 3777–3792 | 3783
 View Article Online

Paper CrystEngComm

vertices. A decision is made by pairing the lines of the
barycentric net embedding to a normalized set of vectors
representing the SBU connection sites that yields the lowest
root mean squared deviation.
Once the connection sites of each SBU have been assigned
to specific arcs from the labelled quotient graph, we can com-
pare their geometries to that of the inner products from a net
embedding generated using eqn (10). We define elements gij
of the net embedding's inner product matrix g as belonging
to a particular SBU if the arcs ei and ej in the augmented la-
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.
NSBU = {garas|(1 ≤ r ≤ n), (s ≤ r)} (16)
There are 16 arcs assigned to Cd4(O2CR)8 in the aug-
mented quotient graph of bcu. Thus n = 16 and the set NCd2+
runs over 136 elements of the inner product matrix.
It should be noted that when constructing the geometry of
the reduced SBU representation in eqn (9), the orientations
of the vectors must match the positive orientation of the as-
sociated arc in the labelled quotient graph. For example, the
arc +e9 in Fig. 5b points toward the vertex A. Thus the vector

belled quotient graph were assigned to that SBU. In the case
of the bcu topology, The Cd4(O2CR)8 SBU is assigned all of
the arcs connected to the central vertex A in Fig. 5b as well as
the adjacent arcs representing the SBU's connection sites.
Suppose the n edge indices assigned to a particular SBU are
found in the set a = (a1,⋯, ak,⋯, an), the elements of g asso-
ciated with an SBU would be,
associated with +e9 in the underlying SBU representation of
Cd4(O2C)8 must be oriented into the SBUs centre of mass
(which was assigned to the vertex A).
Now armed with the necessary elements required to com-
pare metric tensor parameters which optimizes the objective
function in eqn (11). Fig. 6 shows a graphical progression
from the barycentric embedding of the net bcu as taken from

Fig. 6 The graphical representation of progressing from the a) barycentric representation of the net bcu to b) an augmented form of the net with
orange vertices representing SBU connection sites, black vertices representing the centre of mass of the SBUs. c) The bcu net is adjusted into a
non-barycentric form to match the geometry of the SBUs in Fig. 4. d) The SBUs are super-imposed on the net vertices to form the final hypotheti-
cal MOF structure.

3784 | CrystEngComm, 2016, 18, 3777–3792 This journal is © The Royal Society of Chemistry 2016

CrystEngComm
the RCSR to the embedding which fits the SBUs from Fig. 4.
The barycentric representation of bcu is shown in Fig. 6a, the
augmented form with lines and points added to accommo-
date SBU connection sites is seen in Fig. 6b. This is followed
by an optimization of the X and P parameters of the net in
Fig. 6c, such that its geometry best accommodates the SBUs.
Finally, the SBUs are oriented onto their associated vertices
to construct the final hypothetical MOF in Fig. 6d.
Implementation
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.
The MOF building program was written primarily in Python,
with parts in C++ to interface with the NLOPT33 optimization
library. The outline of the program is as follows:
•The program reads in labelled quotient graphs obtained
from the RCSR. The format of which is general, and used in
other collections such as the EPINET database.34 Further de-
tails are given in the ESI.†
•SBUs are chosen if they agree with the degree of the net,
based on the number of connection sites they possess.
•The labelled quotient graph is augmented to introduce
more vertices and edges so they can be matched to each SBUs
centre of mass and connection sites. If linear two-connected
SBUs are used, they will be accommodated by inserting 3 di-
valent vertices into a single edge from the labelled quotient
graph; a central vertex for the SBU's centre of mass, and two
vertices for its connection sites. The edges split in this man-
ner are typically between vertices assigned to opposing SBU
type. For example an edge joining two inorganic SBUs will be
split to support a linear organic SBU.
•SBUs and their connection sites are assigned to specific
vertices and arcs in the labelled quotient graph based on a
lowest RMS score with respect to the barycentric embedding
of the net. The authors recognize that one can leverage sym-
metry elements found in the labelled quotient graph (auto-
morphisms) to assign arcs and vertices to SBUs (assuming
the point group symmetry of the SBU is known). However,
this is currently not implemented.
•The net embedding is optimized to the geometries of the
SBUs. This is accomplished by constructing the objective
function for the net and minimizing it using the L-BFGS algo-
rithm discussed above.
•Following optimization, the SBUs are oriented onto their
assigned points to complete the hypothetical MOF construc-
tion. This orientation consists of a translation of the SBU's
centre of mass to the central node, and a rotation to least
squares fit the connection sites to the appropriate points.35
At this stage, an atom collision check is performed to ensure
that no atoms are closer than the sum of their van der Waals
radii multiplied by a user-defined tolerance factor. In this
work, the tolerance factor was set to 0.4.
Results and discussion
Fitting different 4-c SBUs to nets
As an initial demonstration of the methodology, we will con-
struct periodic structures of different topologies using a four
This journal is © The Royal Society of Chemistry 2016
View Article Online
Paper
coordinate (4-c) inorganic SBU shown in Fig. 7a and a 4-c or-
ganic SBU shown in Fig. 7b. The copper paddlewheel can be
considered a square planar SBU with angles of 90° between
adjacent edges. The organic SBU can be considered tetrahe-
dral-like, which has four angles of 116.5° and two angles of
96.1°. The 3 topologies taken from the RCSR were chosen for
instructive purposes; the two nets dia and qtz whose labelled
quotient graphs are shown in Fig. 2, and the net pts. The first
two nets have only tetrahedral-like vertices in their
barycentric form, thus it represents a challenging case to fit
the square planar Cu paddlewheel to these nets, as it will re-
quire significant distortion of the net. The net pts contains
both square planar and tetrahedral like vertices in its
barycentric representation, thus it will be less challenging to
fit the net to these SBUs. Moreover this net is known to sup-
port geometries of the first two SBUs described in Fig. 7 and
is thus likely to give the best fit to the SBUs when optimizing
the net geometry.36,37 The labelled quotient graphs of dia
and pts are classified as bipartite, which means the vertices
in the graph can be separated into separate sets based on
their connectivity. This makes it easy for SBU placement, as
the inorganic SBUs can be placed in one set of vertices and
the organic SBUs can be placed in the other set, ensuring
that organic and inorganic SBUs connect only to each other.
However the qtz net produces a challenge for the choice of
SBU placements. It is not bipartite and so we must relax the
rule that organic and inorganic SBUs only join to each other
Fig. 7 SBUs used as input into the structure generation program and
their underlying vertex representations. a) (4-c) square planar copper
paddlewheel SBU, b) (4-c) tetrahedral organic SBU and c) (3-c) trigonal
planar organic SBU. Black vertices in the underlying SBU representa-
tions correspond to the SBU's centre of mass, and orange vertices cor-
respond to their connection sites.
CrystEngComm, 2016, 18, 3777–3792 | 3785

Paper
and not themselves. In this case only organic SBUs were
placed on adjacent vertices. This effectively creates a larger,
more complicated organic SBU in the hypothetical MOF.
The fitting of these nets to SBUs is not always perfect for a
number of reasons. The first being the program attempts to
match a net to a set of rigid SBU geometries, which, due to
the constraints of their angles, may not support a particular
periodic connectivity. A discussion of these types of geome-
tries can be found in ref. 38. Another reason is due to the
constraints set on the optimization parameters. For example,
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.
during the optimization, the cell angles are not permitted to
deviate lower than 60° and greater than 120°. For these rea-
sons, a ‘fitness’ of the embedded net is calculated once the
optimization has converged. This is computed as the average
deviation of the net geometry, in terms of its line lengths
(line length deviation shown in the first row of Table 1) and
angles (angle deviation shown in the second row of Table 1)
compared with underlying geometry of the rigid SBUs the
program fits the net to. Upon evaluating the net's fitness, one
can use discretion on whether to keep or discard the struc-
ture. In some cases the fit can be poor, but produce a reason-
able structure after relaxing the atoms and bonds by optimi-
zation with a molecular mechanics force field or ab initio
calculation. Table 1 reports how well each net embedding fits
to the rigid SBUs. While the averages of line length and angle
deviations were relatively small across all topologies, the
standard deviations appear to be the lowest for the pts net.
The stresses on the unit cell were computed from a single
point DFT calculation in VASP.39–42 Comparison of the stress
tensors show the pts net provides the lowest amount of strain
in the structure. It is clear that the significant distortions re-
quired to fit the dia and qtz nets to the square planar Cu
paddlewheel SBU has resulted in some unphysical stresses
on the system.
While not investigated here, it is possible that expanding
the labelled quotient graph (or reducing the translational
symmetry) so that the ‘unit cell’ of the net contains more ver-
tices and edges to support more SBUs may reduce the stress
values shown in Table 1. For example, 3 of the 4 arcs in the
labelled quotient graph of dia correspond to the three lattice
vectors 100, 010, and 001. Since these three arcs are assigned
to a square planar paddlewheel, there is considerable pres-
sure to make these lattice vectors coplanar. This can be seen
by the non-zero stress values in the XY YZ and ZX values for
Table 1 Measurements of the fit of three nets to the organic and inor-
ganic SBUs shown in Fig. 7
Topology dia qtz pts
Line length deviation (Å) 0.24 ± 0.84 −0.31 ± 1.66 −0.79 ± 0.65
Angle deviation (degrees) 1.13 ± 11.8 3.46 ± 11.46 1.18 ± 4.04
Stress XX (eV) −35.7 −95.4 0.9
Stress YY (eV) −66.1 −102.3 0.9
Stress ZZ (eV) −56.7 −119.9 1.8
Stress XY (eV) 6.8 −15.9 0.0
Stress YZ (eV) 12.8 50.0 0.0
Stress ZX (eV) −12.7 −27.4 0.0
3786 | CrystEngComm, 2016, 18, 3777–3792
View Article Online
CrystEngComm
dia. By increasing the number of SBUs in the unit cell, we
can effectively distribute this stress over a larger number of
bonds and degrees of freedom. Similar arguments can be
made for the qtz net whose lattice vectors are determined by
a sum of only two or three arcs. It is notable that the stress
values in the XX, YY, and ZZ directions are negative for both
the dia and qtz MOFs, indicating a drive towards volume
compression of the unit cells. The connectivity of these struc-
tures indicates that at least two connection sites of the tetra-
hedral SBU in Fig. 7b are connected to two sites of the same
Cu paddlewheel. While the tetrahedral orientation of the or-
ganic SBU is provided by a single sp3 hybridized carbon, the
square planar angles of the Cu paddlewheel are much more
rigid, being supported by two Cu2+ ions coordinated to conju-
gated carboxylate moieties. Therefore the compression likely
arises from the forced planarity of the tetrahedral SBUs.
Structures built from the RCSR
To demonstrate the breadth of structures the program is able
to generate, the (4-c) copper paddlewheel and a (3-c) SBU
shown in Fig. 7a and c, respectively were used to generate 46
MOFs unique in their topology. Note, at the time of this writ-
ing, the number of non-augmented, non-catenated nets with
mixed degree (3,4) in the RCSR is 49, however 3 structures
contained atomic collisions which could not be eliminated.
Many of the topologies shown along the x-axis of Fig. 8 are
not bipartite. Thus in some of the MOFs Cu paddlewheels
are bonded to Cu paddlewheels, and in others the 3-c SBU is
bonded to other 3-c SBUs. All of these structures were opti-
mized using a molecular mechanics force field prior to com-
puting their surface areas.
It is interesting to see such a diversity of surface areas in
Fig. 8 achieved from building MOFs with only two types of
SBUs. As a rough comparison, the database of over 130 000
MOFs constructed by Wilmer et al. possesses a similar
Fig. 8 Specific surface area of 46 MOFs assembled from topologies
found in the RCSR using the 4-c Cu paddlewheel and 3-c SBU shown
in Fig. 7a and c, respectively.
This journal is © The Royal Society of Chemistry 2016
 View Article Online

CrystEngComm Paper

distribution of gravimetric surface areas (from approximately

0–8000 m2 g−1).43 However the structures in that database
were assembled with 102 building blocks,1 and consisted of
only 6 topologies.11 This method therefore has the potential
to construct a much more diverse database of structures
holding the number of building blocks constant.
By its definition, the program forms MOFs with specific
topologies by performing the appropriate net isomorphism to
best match the SBU geometries. Thus it can be said that the
topologies of the hypothetical MOFs generated with this pro-
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.

gram will be the same as the underlying net initially used as

structure-directing. To test this, the topologies of the struc-
tures generated were determined by the TOPOS program,19
which abstracts chemical systems to their underlying nets
and classifies them against the TOPOS Topological Database
(TTD). As expected the TOPOS program provided the same to-
pology for each hypothetical MOF structure as was used to
build it. The structures and details of the TOPOS calculations
are provided in the ESI.†

Diverse MOF structures with the tbo topology

Delgado-Friedrichs and O'Keeffe demonstrated that the
highest possible symmetry of a net can be achieved when the
vertices are in their barycentric placement.8 The space groups
reported for each net in the RCSR database are indeed those
of the maximal achievable symmetry group of the net. Martin
and Harancyk demonstrated that MOFs could be constructed
using these nets represented by ideal symmetry operations.12
However, there are many cases where the barycentric repre-
sentation is not possible. For example, some nets contain
intersecting vertices in their barycentric representation, al-
though crystal structures have been observed with these to-
pologies.44 As another example, consider the bipartite net
tbo, which is associated with the MOF HKUST-1,45 one of the
most studied MOFs. The ideal space group of the net tbo is
Fm3̄m, which supports 3-c vertices with site symmetry 3m
(C3v in Shoenflies notation), and 4-c vertices of mmm symme-
try. Now suppose one wanted to construct a hypothetical ma-
terial using SBUs of significantly lower point group symmetry
than those supported by the idealized space group. Consider
the tritopic 1,3,5-benzenetricarboxylate (BTC) SBU used to
build HKUST-1 shown in Fig. 9a). The point group symme-
try of this molecule is D3h as there are three dihedral mirror
planes present in the molecule as well as a 3-fold rotation
axis. The SBU is supported by the ideal symmetry opera-
tions of the net as there is a 3-fold rotation axis at the 3-c
vertex. If, however, we were to instead use the 3-c SBU
biphenyl-3,4′,5-tricarboxylate (BPTC), we have effectively re-
duced the maximum possible point group symmetry of
the 3-c vertices in tbo to C2v. Any symmetry operation sup-
porting a three-fold rotation must therefore be scrapped
in the unit cell. It is important to stress here that the sym-
metry is not only of the atomic structure, but also of the
underlying connectivity. If one were to abstract the biphenyl
SBU seen in Fig. 9c to vertices and edges, there would be
Fig. 9 The generation of three different hypothetical structures based
on the same underlying topology, tbo. a) The highly symmetric
benzene-1,3,5-tricarboxylate (BTC) SBU and b) the primitive cell of the
hypothetical MOF generated by the program, viewing down the c axis.
c) The lower symmetry biphenyl-3,4′,5-tricarboxylate and d) its
hypothetical structure. e) The low symmetry biphenyl-3,3′,5-
tricarboxylate and f) its hypothetical structure. The point groups of the
SBUs are displayed with the mirror planes as purple circles and
rotation axes as yellow rods.
one long edge along the principle axis of rotation, and two
small edges pointing in the direction of the carboxylates
located on the same phenyl ring.
We also build a structure using a tritopic SBU with no ro-
tational symmetry in the molecule or the underlying connec-
tivity. This was accomplished by shifting the lone carboxylate
group of BPTC over one hydrogen to form biphenyl-3,3′,5-
tricarboxylate seen in Fig. 9e. Computational generation of
materials with such reductions in symmetry have not yet
been demonstrated by previously reported structure genera-
tion routines, which either snap SBUs together sequentially,1
place the SBUs in their barycentric positions, or a combina-
tion of both.12,46
We computed the resulting symmetry of the hypothetical
MOFs using a symmetry detection algorithm47 which finds
the appropriate space group of a crystal system by finding
equivalent atoms using a set of symmetry operations from a
lookup table. Equivalent atoms were found with a tolerance
of 0.4 Å.
Fig. 9b shows the generation a hypothetical version of
HKUST-1 MOF. The structure was generated with a negligible
line length deviation and an average angle deviation of 2.0 ±
3.3 degrees between the SBUs and the final embedded net. In

This journal is © The Royal Society of Chemistry 2016 CrystEngComm, 2016, 18, 3777–3792 | 3787

Paper
addition, the symmetry detected in the generated structure
was the ideal symmetry space group found in the barycentric
representation of the net, Fm3̄m. For the structure built with
the SBU with C2v symmetry in Fig. 9d the resulting net em-
bedding found from our algorithm, when compared with the
SBU geometries, was reported to possess an average line
length deviation of 0.3 ± 0.4 Å and angle deviation of 6.6 ± 10
degrees. The reported crystal system was found to be mono-
clinic C2 – which possesses two-fold screw axis and a two-
fold rotation axis. Both symmetry operations support the
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.
symmetry of the C2v SBU. The structure built from the low
symmetry SBU in Fig. 9f was found to have P1 symmetry, as
expected. The line length and angle deviations reported from
generating this material was 0.1 ± 0.4 Å and 7 ± 10 degrees,
respectively. Crystalline structures with this level of strain
and low symmetry will likely never be experimentally realized,
however generation of this material demonstrates the robust-
ness of the structure generation code. The displayed struc-
tures have not been geometry optimized in any form, they are
a direct output of the structure generation program.
Building MOFs with unstable nets
The advantage this program possesses over other methods is
that it can be extended beyond the nets found in the RCSR
in their barycentric representation. As an example, we will
build a structure based on an unstable net 3IJ32,4)2. Unstable
nets, so-called because they possess intersecting vertices in
their barycentric form, have been shown as important as de-
scriptions of crystallographic structures. Recently Delgado-
Friedrichs et al. discussed several crystalline materials which
could be classified as having unstable nets with collisions in
their barycentric placement.44 Here we assemble a MOF with
the copper paddlewheel (Fig. 7a) and a tritopic organic SBU
(Fig. 7b), using only the labelled quotient graph as input to
define the topology. The final hypothetical MOF built from
Fig. 10 Hypothetical MOF built with the unstable net 3IJ32,4)2. Left: The labelled quotient graph, middle: the SBUs used to build the structure,
right: the hypothetical MOF.
3788 | CrystEngComm, 2016, 18, 3777–3792
View Article Online
CrystEngComm
this net is shown in Fig. 10 and was confirmed to posses the
3IJ32,4)2 topology using the TOPOS program. It should be
noted that the embedding of this net produced somewhat un-
natural bending within the structure, which can be expected
when fitting rigid SBUs to an underlying graph. Potential en-
ergy based geometry optimization of the material with an ap-
propriate force field or similar would improve the structure,
however, the structure given in Fig. 10 is the ‘raw’ structure
obtained with our approach.
Challenging topologies and high surface area MOFs
Recently, Yaghi and coworkers reported a material, MOF-210
with, at the time, a record breaking Brunauer–Emmet–Teller
(BET) surface area of 6240 m2 g−1.48 Building this MOF repre-
sents a challenge for current structure generation algorithms
for several reasons. The MOF consists of 3 unique SBUs of
different geometry; a hexatopic Zn4O inorganic SBU found in
other MOFs such as MOF-5,18 a 3-c organic SBU; 4,4′,4″-
(benzene-1,3,5-triyl-trisIJethyne-2,1-diyl))tribenzoate (BTE), and
biphenyl-4,4′-dicraboxylate (BPDC), which is a linear 2-c SBU.
The rhombohedral primitive cell of this material is very large,
possessing cell dimensions of 71 Å and containing a total of
18 Zn4O inorganic SBU clusters, 18 BPDC SBUs, and 12 BTE
SBUs. These factors alone would pose a particular challenge
for the recursion method,1 which initiates a structure by plac-
ing a single SBU in space, and samples every possible combi-
nation of SBU bonds until a structure is made. To assemble
the hypothetical version of MOF-210, which possesses the toz
topology, with the tinker-toy-based construction method, one
would need to sample a very large number of possibilities be-
fore achieving the correct structure. Assuming that a
6-coordinate Zn4O SBU is initially placed, then each of the 6
connection sites would sample the 3 bonds from the BTE or-
ganic SBU and the 2 bonds from BPDC, yielding 65 = 7776
possible bonding combinations for the first SBU alone. There
This journal is © The Royal Society of Chemistry 2016
 View Article Online

CrystEngComm Paper

are a total of 48 SBUs in the primitive cell of this MOF, so
one would have to sample on the order of 1090 possible bond-
ing combinations, one of which would yield MOF-210. In
practice, an upper bound must be placed on the number of
bond combinations to sample, thus even if the tinker toy al-
gorithm can perform trial bonding moves in seconds, it is
unlikely that MOF-210 could be built. In addition, the coordi-
nation environments of the Zn4O SBUs are different. Each is
coordinated to 4 tritopic BTE SBUs and 2 linear BPDC SBUs,
however some Zn4O SBUs coordinate two BPDC in a cis fash-
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.
generated structure following geometry optimization of both
the atomic positions and the cell at the molecular mechanics
level. Table 2 also reveals that the calculated surface area and
pore volume of the generated structure are within 1.1% of
the experimental structure. We believe the agreement in the
cell parameters and other properties is excellent considering
how that only the labelled quotient graph of the net was used
as input and that the initial SBU geometries were taken from
sources not related to the crystal structure of MOF-210. To
compare the generated and experimental structures visually

ion, others in a trans fashion. Thus to build this structure
with the tinker-toy algorithm would require not only an in-
conceivable amount of compute resources, but also a large
amount of information from the experimental crystal struc-
ture to ensure the bonding motifs of the Zn SBUs are pre-
served. For these reasons, predictive capacity of the tinker-toy
algorithm is somewhat limited to relatively small, topologi-
cally and chemically simple structures.
These problems can be alleviated with our algorithm. By
using the labelled quotient graph of the toz net as structure
directing, and defining the 2-c, 4-c and 6-c SBUs, one can as-
semble a hypothetical version of MOF-210 with relative ease.
In this case, all 6-c vertices are assigned to the Zn4O SBU and
3-c vertices are assigned to the BTE SBU. With the placement
of Zn4O SBUs on the 6-c vertices, the program subsequently
recognizes these vertices as ‘inorganic’, thus in cases where
6-c vertices in the graph are incident upon other 6-c vertices,
it will automatically augment the edge between these vertices
with an additional 2-c vertex to support a bridging linear
ditopic BPDC SBU. To construct the hypothetical version of
MOF-210, the organic SBUs were built using a conventional
molecule building software. Unlike the SBUs found in the ex-
perimental crystal structure of MOF-210, these SBUs were
built entirely planar, with typical C–C and C–H bond lengths
and angles for aromatic and alkyne systems. The inorganic
SBU was extracted from MOF-5 (ref. 18) which contains the
same Zn4O cluster, only connected in a different topology to
different organic SBUs.
Compared in the first two columns of Table 2 are the ex-
perimental unit cell parameters of MOF-210, and those of our
and to depict the complexity of the structure, Fig. 11b, shows
an overlay of the experimental atomic positions in the unit
cell with that of the generated structure. It is worth noting
that the ‘raw’ unit cell vectors produced by the algorithm will
generally not be the same as the experimental vectors. How-
ever, one can define a linear combination of the original vec-
tors to yield cell vectors and dimensions to agree with the ex-
perimental counterparts. Further discussion of adjusting the
cell vectors in this way is provided in the ESI.† We note that
altering the unit cell dimensions does not change the inher-
ent physical and chemical characteristics of the periodic
structure. The presented values for surface area and pore vol-
ume in Table 2 were identical both before and after the cell
parameters were adjusted.
Farha et al.50 recently demonstrated a method for improv-
ing the surface area of an existing MOF, NU-100, based on
the ntt topology. By increasing the length of the organic SBU
arms with extra benzene and acetylene moieties, they were
able to generate NU-110, which surpassed MOF-210 as the
MOF with record breaking gravimetric surface area.50 The or-
ganic SBU of NU-110 is similar to the 3-c BTE SBU in that it
contains a central aromatic ring with three arms extending
outward with D3h symmetry. The differences being that the
length of these arms are approximately four times larger than
those of BTE, and at the terminus of these arms two carboxyl-
ate moieties are found, yielding a 6-c SBU (the authors called
LH6 (ref. 50)) instead of the 3-c SBU of BTE. It was shown50
that the computed physical properties of the hypothetical
structure generated with the recursion method were in good
agreement with the experimentally realized version of this

Table 2 Cell parameters and selected physical properties of two experimental MOFs, NU-110 and MOF-210, along with purely hypothetical MOFs gen-
erated with the ntt, toz and ith-d topologies

Experimental Hypothetical Experimental Hypothetical Hypothetical Hypothetical

MOF MOF-210 MOF-210 NU-110 NU-110 MOF-toz MOF-ith-d
a (Å) 50.7 51.4 48.6 49.6 103.7 69.2
b (Å) 50.7 51.4 48.6 49.7 98 69.3
c (Å) 194.3 191.6 48.6 49.6 461.7 69.3
α (deg) 90 90 60 60 54 89.6
β (deg) 90 90 60 60 105 93
γ (deg) 120 120 60 60 118 90
Calculated surface areaa (m2 g−1) 5760 5720 6258 6624 8737 8462
Calculated pore volumeb (cm3 g−1) 2.66 2.69 3.04 3.24 15.71 13.94
Experimental BET surface area (m2 g−1) 6580c — 7140b — — —
Experimental pore volume (cm3 g−1) 3.60c — 4.40b — — —
a
Calculated values from zeo++49 using a probe radius of 1.82 Å. b From ref. 50. c From ref. 48.

This journal is © The Royal Society of Chemistry 2016 CrystEngComm, 2016, 18, 3777–3792 | 3789
 View Article Online

Paper CrystEngComm
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.

Fig. 11 The hypothetical generation of two high surface area experimental MOFs. a) The SBUs used to assemble the hypothetical version of MOF-
210, including the inorganic Zn4O cluster, 4,4′,4″-(benzene-1,3,5-triyl-trisIJethyne-2,1-diyl))tribenzoate (BTE) and biphenyl-4,4′-dicarboxylate
(BPDC). b) Direct comparison of the hypothetical (blue) and experimental (red) MOF-210 structures. c) The organic SBUs used to assemble the hy-
pothetical version of NU-110. The original 6-c LH6 ligand (carboxylate moieties replaced with dotted lines) was fragmented into a 3-c benzene
SBU, and 3-c LH3 containing the remaining arms of the LH6 ligand.50 d) Superposition of the hypothetical (blue) and experimental (red) structures
of NU-110.

MOF, and so as a curiosity, in this work we assembled the hy-
pothetical version of NU-110, as well as new MOFs based on
the tritopic version of the LH6 SBU.
We constructed the hypothetical version of NU-110 using
the labelled quotient graph of ntt, which possesses vertices
of degree 3 and 4. This required an alteration of the SBUs
used to assemble the MOF; the 4-c Cu paddlewheel, and a
fragmentation of the 6-c LH6 organic ligand into large (LH3)
and small (benzene) 3-c SBUs to support the separate 3-c ver-
tices in the net. One challenge in constructing the hypotheti-
cal NU-110 structure was to ensure the 3-c organic SBUs were
placed at specific nodes in the ntt net such that they com-
bine to form the large (6-c) SBU in the final structure. Know-
ing that only the small 3-c benzene SBU form bonds with
copper paddlewheels in the crystal structure, we imposed a
constraint such that only these SBUs were assigned to 3-c ver-
tices adjacent to 4-c vertices in the labelled quotient graph.
The remaining 3-c vertices were then assigned LH3 SBUs. The
authors note that relaxing this constraint will still yield valid
(and entirely different) hypothetical MOF materials, however
for the purposes of constructing hypothetical NU-110, it was
necessary to judiciously assign 3-c SBUS to appropriate verti-
ces. Comparing the computed physical properties for both ex-
perimental and hypothetical NU-110 in columns 3 and 4 of
Table 2 shows a slight over-prediction with the hypothetical
MOF. We note that there is a slight distortion in the SBUs
when the hypothetical and experimental MOFs of NU-110
were superposed as seen in Fig. 11d, which is likely the cause
of the larger values in the reported properties for hypotheti-
cal NU-110.
The final two columns of Table 2 demonstrate how the al-
gorithm can predict new MOFs with incredibly high surface
areas and void volumes. Both the toz and ith-d hypothetical
MOFs were constructed using the exact same SBUs, one of
which is a 3-c adaptation of the LH6 SBU found in NU-110.
The other SBUs were the Zn4O inorganic SBU seen in Fig. 11a
and a 2-c organic SBU large enough to support the extended
3-c SBUs in both the toz and ith-d topologies. Hypothetical
MOF-toz was found to have a 1.4-fold increase in calculated
surface area over that of the NU-110 MOF. In addition the
pore volume is larger by a factor of approximately 5. Hypo-
thetical MOF-ith-d also has potential to break the gravimetric
surface area record, however the computed properties are
slightly lower than that of the SBUs assembled in the toz to-
pology. These structures are provided in the ESI† in the crys-
tallographic information file format.

3790 | CrystEngComm, 2016, 18, 3777–3792 This journal is © The Royal Society of Chemistry 2016

CrystEngComm
Conclusions
Through a series of theorems and proofs Eon13 formally
showed that one can obtain a representation of a
n-dimensional net with deviations from its barycentric place-
ment using colattice distortion vectors. Building on Eon's
work, we have developed a general method to generate hypo-
thetical materials beginning with a desired underlying net to-
pology. We demonstrate that a net embedding supporting
the geometries of SBUs can be realized prior to the actual
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.
placement of any SBUs in space, which poses an advantage
over the exhaustive recursion method presented by Wilmer
et al.1 In their method, the underlying topology is encoded in
the geometries of the parameters assigned to the connection
sites of each SBU, effectively limiting the topological space
sampled by the program. In our method, similar to the
method of Martin and Haranczyk,12 requires only the SBUs,
an indication of where they will form bonds with other SBUs,
and a blueprint describing the interconnectivity of the SBUs
within a periodic lattice. The major difference between the
method presented here and that of Martin and Haranczyk is
that we propose using the abstract definition of the net over
the barycentric realization. This allows us to manipulate the
net to match SBU geometries prior to assembly. Additionally
this method allows for assembly of MOFs based on topolo-
gies not described in the RCSR database, allowing for a much
wider range of hypothetical structures to sample.
In this work we demonstrate the method on MOFs with
3-dimensional periodicity, but it is applicable to the construc-
tion of other classes of materials, such as porous polymer
networks51 or covalent organic frameworks with
2-dimensional periodicity.52,53 This program can be used for
several applications, including the development of new struc-
tures based on the modification of SBUs in existing mate-
rials, or to generate a large database of topologically diverse
structures for high throughput screening purposes. Although
we have used the methodology to generate structures of
MOF-210 and NU-110 that were in remarkable agreement
with the experimental crystal structures, the topologies of
these MOFs were known and used as input information.
Thus, the method is not a general crystal structure prediction
algorithm. However, TOBASCCO could be used to generate
structures of all possible topologies as part of a crystal struc-
ture prediction system that would then also compare free en-
ergies of all possible topologies and geometric isomers, as
well factor in the kinetics of assembly in a given thermody-
namic and experimental environment.
Acknowledgements
The authors would like to thank the Natural Sciences and En-
gineering Research Council of Canada, the Canada Research
Chairs program, and Carbon Management Canada for
funding. We are also grateful for computing resources pro-
vided by Canada Foundation for Innovation and Compute
Canada. PGB would also like to thank NSERC and Ministry of
This journal is © The Royal Society of Chemistry 2016
View Article Online
Paper
Training, Colleges and Universities of Ontario for
scholarships.
Notes and references
1 C. E. Wilmer, M. Leaf, C. Y. Lee, O. K. Farha, B. G. Hauser,
J. T. Hupp and R. Q. Snurr, Nat. Chem., 2012, 4, 83–89.
2 R. L. Martin, C. M. Simon, B. Smit and M. Haranczyk, J. Am.
Chem. Soc., 2014, 136, 5006–5022.
3 L.-C. Lin, A. H. Berger, R. L. Martin, J. Kim, J. A. Swisher, K.
Jariwala, C. H. Rycroft, A. S. Bhown, M. W. Deem, M.
Haranczyk and B. Smit, Nat. Mater., 2012, 11, 633–641.
4 A. F. Wells, Three Dimensional Nets and Polyhedra, John
Wiley & Sons Ltd., New York, 1977.
5 M. O'Keeffe and B. G. Hyde, Philos. Trans. R. Soc., A,
1980, 295, 553–618.
6 E. V. Alexandrov, V. A. Blatov, A. V. Kochetkov and D. M.
Proserpio, CrystEngComm, 2011, 13, 3947.
7 M. O'Keeffe, M. A. Peskov, S. J. Ramsden and O. M. Yaghi,
Acc. Chem. Res., 2008, 41, 1782–1789.
8 O. Delgado-Friedrichs and M. O'Keeffe, Acta Crystallogr.,
Sect. A: Found. Crystallogr., 2003, 59, 351–360.
9 M. Fernandez, P. G. Boyd, T. D. Daff, M. Z. Aghaji and T. K.
Woo, J. Phys. Chem. Lett., 2014, 5, 3056–3060.
10 E. S. Kadantsev, P. G. Boyd, T. D. Daff and T. K. Woo,
J. Phys. Chem. Lett., 2013, 4, 3056–3061.
11 B. J. Sikora, R. Winnegar, D. M. Proserpio and R. Q. Snurr,
Microporous Mesoporous Mater., 2014, 186, 207–213.
12 R. L. Martin and M. Haranczyk, Cryst. Growth Des., 2014, 14,
2431–2440.
13 J.-G. Eon, Acta Crystallogr., Sect. A: Found. Crystallogr.,
2011, 67, 68–86.
14 O. Delgado-Friedrichs and M. O'Keeffe, J. Solid State Chem.,
2005, 178, 2480–2485.
15 S. J. Chung, T. Hahn and W. E. Klee, Acta Crystallogr., Sect.
A: Found. Crystallogr., 1984, 40, 42–50.
16 J.-G. Eon, J. Solid State Chem., 1998, 138, 55–65.
17 M. O'Keeffe, M. Eddaoudi, H. Li, T. Reineke and O. M.
Yaghi, J. Solid State Chem., 2000, 152, 3–20.
18 M. Eddaoudi, J. Kim, N. L. Rosi, D. T. Vodak, J. Wachter, M.
O'Keeffe and O. M. Yaghi, Science, 2002, 295, 469–472.
19 V. A. Blatov, A. P. Shevchenko and D. M. Proserpio, Cryst.
Growth Des., 2014, 14, 3576–3586.
20 T. H. Cormen, C. E. Leiserson, R. L. Rivest and C. Stein,
Introduction to Algorithms, MIT Press, 2nd edn, 2001.
21 M. T. Goodrich and R. Tamassia, Data Structures and
Algorithms in Java, John Wiley & Sons Ltd., 2nd edn, 2001.
22 D. Joyner, M. Van Nguyen and N. Cohen, Algorithmic Graph
Theory, 2010.
23 S. Sahni, Data Structures, Algorithms, and Applications in
Java, McGraw-Hill, 2000.
24 W. A. Stein, sage Dev. team, http://www.sagemath.org, 2013.
25 M. O'Keeffe and O. M. Yaghi, Chem. Rev., 2012, 112,
675–702.
26 C. Godsil and G. Royle, Algebraic Graph Theory, Springer,
New York, 1st edn, 2001.
CrystEngComm, 2016, 18, 3777–3792 | 3791

Paper
27 J. Nocedal, Math. Comp., 1980, 35, 773.
28 D. C. Liu and J. Nocedal, Mathematical Programming,
1989, 45, 503–528.
29 P. Kaelo and M. M. Ali, J. Optim. Theory Appl., 2006, 130,
253–264.
30 H. Chun, D. Kim, D. N. Dybtsev and K. Kim, Angew. Chem.,
Int. Ed., 2004, 43, 971–974.
31 S. Bureekaew and R. Schmid, CrystEngComm, 2013, 15, 1551.
32 S. Bureekaew, V. Balwani, S. Amirjalayer and R. Schmid,
CrystEngComm, 2015, 17, 344–352.
Published on 21 March 2016. Downloaded by University of Ottawa on 1/26/2022 1:47:31 AM.
33 S. G. Johnson, The NLopt nonlinear-optimization package,
http://ab-initio.mit.edu/nlopt.
34 S. J. Ramsden, V. Robins and S. T. Hyde, Acta Crystallogr.,
Sect. A: Found. Crystallogr., 2009, 65, 81–108.
35 K. S. Arun, T. S. Huang and S. D. Blostein, IEEE Trans.
Pattern Anal. Mach. Intell., 1987, PAMI-9, 698–700.
36 B. Chen, M. Eddaoudi, T. M. Reineke, J. W. Kampf, M. O'Keeffe
and O. M. Yaghi, J. Am. Chem. Soc., 2000, 122, 11559–11560.
37 J. Kim, B. Chen, T. M. Reineke, H. Li, M. Eddaoudi, D. B.
Moler, M. O'Keeffe and O. M. Yaghi, J. Am. Chem. Soc.,
2001, 123, 8239–8247.
38 T. R. Cook, Y.-R. Zheng and P. J. Stang, Chem. Rev.,
2013, 113, 734–777.
39 G. Kresse and J. Hafner, Phys. Rev. B: Condens. Matter Mater.
Phys., 1994, 49, 14251–14269.
40 G. Kresse and J. Hafner, Phys. Rev. B: Condens. Matter Mater.
Phys., 1993, 47, 558–561.
41 G. Kresse and J. Furthmüller, Comput. Mater. Sci., 1996, 6, 15–50.
3792 | CrystEngComm, 2016, 18, 3777–3792
View Article Online
CrystEngComm
42 G. Kresse and J. Furthmüller, Phys. Rev. B: Condens. Matter
Mater. Phys., 1996, 54, 11169–11186.
43 C. E. Wilmer, O. K. Farha, Y.-S. Bae, J. T. Hupp and R. Q.
Snurr, Energy Environ. Sci., 2012, 5, 9849.
44 O. Delgado-Friedrichs, S. T. Hyde, S. W. Mun, M. O'Keeffe
and D. M. Proserpio, Acta Crystallogr., Sect. A: Found.
Crystallogr., 2013, 69, 535–542.
45 S. S.-Y. Chui, S. M.-F. Lo, J. P. H. Charmant, A. Guy Orpen
and I. D. Williams, Science, 1999, 283, 1148–1150.
46 M. A. Addicoat, D. E. Coupry and T. Heine, J. Phys. Chem. A,
2014, 118, 9607–9614.
47 A. Togo, http://spglib.sourceforge.net/, 2012.
48 H. Furukawa, N. Ko, Y. B. Go, N. Aratani, S. B. Choi, E. Choi,
A. O. Yazaydin, R. Q. Snurr, M. O'Keeffe, J. Kim and O. M.
Yaghi, Science, 2010, 329, 424–428.
49 T. F. Willems, C. H. Rycroft, M. Kazi, J. C. Meza and M.
Haranczyk, Microporous Mesoporous Mater., 2012, 149,
134–141.
50 O. K. Farha, I. Eryazici, N. C. Jeong, B. G. Hauser, C. E.
Wilmer, A. A. Sarjeant, R. Q. Snurr, S. T. Nguyen, A. Ö.
Yazaydın and J. T. Hupp, J. Am. Chem. Soc., 2012, 134,
15016–15021.
51 W. Lu, D. Yuan, D. Zhao, C. I. Schilling, O. Plietzsch, T.
Muller, S. Bräse, J. Guenther, J. Blümel, R. Krishna, Z. Li and
H.-C. Zhou, Chem. Mater., 2010, 22, 5964–5972.
52 A. P. Côté, A. I. Benin, N. W. Ockwig, M. O'Keeffe, A. J.
Matzger and O. M. Yaghi, Science, 2005, 310, 1166–1170.
53 H. Xu, J. Gao and D. Jiang, Nat. Chem., 2015, 7, 905–912.
This journal is © The Royal Society of Chemistry 2016