# equivariant_networks_for_crystal_structures__0abae71c.pdf

Equivariant Networks for Crystal Structures

Sékou-Oumar Kaba, Siamak Ravanbakhsh

School of Computer Science,

Mc Gill University Mila - Quebec Artiﬁcial Intelligence Institute {kabaseko@mila.quebec, siamak@cs.mcgill.ca}

Supervised learning with deep models has tremendous potential for applications in materials science. Recently, graph neural networks have been used in this context, drawing direct inspiration from models for molecules. However, materials are typically much more structured than molecules, which is a feature that these models do not leverage. In this work, we introduce a class of models that are equivariant with respect to crystalline symmetry groups. We do this by deﬁning a generalization of the message passing operations that can be used with more general permutation groups, or that can alternatively be seen as deﬁning an expressive convolution operation on the crystal graph. Empirically, these models achieve competitive results with state-of-the-art on property prediction tasks.

1 Introduction

Deep learning has seen remarkable applications in computational chemistry, both on the side of molecular property prediction and molecule generation. For small molecules, these methods are close to a level of precision that would make them suitable for practical applications [46]. However, less attention has been put to the problem of designing neural network architectures for the larger assemblies of atoms that constitute materials. This is a crucial problem, as deep learning could advantageously complement computationally demanding ab initio simulation methods like DFT [38, 28].

In this paper, we address the problem of designing equivariant layers and use them for supervised learning on materials. We focus on crystals, materials characterized by the ordered arrangement of their atoms in lattices. Crystalline materials are largely present around us and essential in technological applications, as they include a large number of metals, ceramics, and salts, among others [3]. They can be described as the regular repetition of a set of atoms, called the unit cell, in all directions of space. These patterns are analogous to wallpaper patterns, for which the repeated structure is an image. Crystals are characterized by a high degree of symmetry which is fundamental in understanding their physical properties. Consequently, we suggest that equivariant deep learning provides an appropriate framework to design function approximators for crystals.

Graph Neural Networks (GNNs) have recently been proposed for supervised prediction tasks on crystals [54, 70, 12]. We hypothesize here that these models may fall short in exploiting crystal symmetry and that structural information is lost when mapping a crystal structure to a graph. In particular, GNNs are equivariant to the permutation of atoms given as input to the model. This condition is overly restrictive and amounts to forgetting about the ordered nature of a crystal. We propose to use models equivariant to a product of groups G SC, where G acts at the level of the Bravais lattice, the underlying periodic grid of the crystal, and SC on the unit cells. We show that our proposed equivariant architecture is more expressive than GNNs on this data structure.

36th Conference on Neural Information Processing Systems (Neur IPS 2022).

By using the crystal structure, our approach amounts to deﬁning a group-equivariant convolution kernel on the crystal in a way that is completely analogous to convolutional neural networks (CNNs). This convolution is deﬁned on a graph associated with a crystal structure. Our contributions are the following: 1) We derive different equivariant models based on the group-theoretical properties of crystals. 2) We show some of the limitations of GNNs for crystal data and propose an alternative data structure to be used with our architecture. 3) We perform a rigorous analysis, cleaning, and processing of the Materials Project database [33] and share the resulting processed dataset to serve as a benchmark for materials applications. 4) We perform experimental tests of our models on the Materials Project database and report results comparable to or better than baselines.

2 Related works

Equivariant neural networks It is well known that neural networks have to incorporate inductive biases to be useful in practice [67, 7]. Using models that are invariant or equivariant to the symmetry of the data has proven to be a particularly important inductive bias to promote generalization [8]. The ﬁrst notable application of this idea was the CNN, for which each convolution layer is equivariant to translation [42]. An alternative parameter-sharing view also appears in early works [57]; while more recent equivariant networks have used both convolutional [13, 40, 15, 19], and parameter-sharing view [51, 24]. A few notable symmetries considered in recent years are rotation in image and volumetric data [13, 68, 65], permutation symmetry in sets [72, 50] and graphs [39, 44], as well as Euclidean [60, 10], and rotational symmetry [14, 2, 21, 55, 23]. In this work, we build on foundational work on hierarchical symmetries [45, 64]

Deep learning for materials The increasing availability of large materials datasets from highthroughput calculations [18, 33, 36, 47, 31, 11], makes deep learning more and more relevant for materials science. Following the successes of GNNs on molecular data, similar models have been proposed for materials as alternatives to methods based on feature engineering. Note that in what follows, we refer to GNNs in a general sense that includes message passing neural networks (MPNNs) [26].Many variants of GNNs exist [35, 62, 71, 44], the underlying idea being for each node to aggregate features of neighboring nodes in a permutation invariant, or in the case of [39] in an equivariant way. The CGCNN [70] and MEGNet models [12] rely on mapping crystal structures to graphs and applying GNNs to obtain a prediction. For the Sch Net model, the correspondence with graphs is less explicit [54], but still present. Other approaches combine permutation equivariance with E(3)-equivariance to design models for molecules and materials [5, 53].

3 Background on crystal symmetry

We ﬁrst start by introducing some principles of crystallography that will be used to derive our main results. A more comprehensive treatment can also be found in the references [20, 59], whereas basics of group theory are covered in [52] for example.

Lattices A crystal can be described as the periodic and inﬁnite repetition of a pattern in all directions of space. Crystals are conveniently described using lattices as their underlying structures.

An n-dimensional lattice can be deﬁned as the set of integral combinations of the linearly independent lattice basis vectors ai 2 Rn:

miai | mi 2 Z

The lattice is entirely speciﬁed by its basis vectors ai. A lattice is associated with a group of translations T for which the multiplication rule is addition. This captures the translational symmetry of a crystal. A lattice also deﬁnes subsets of Rn called unit cells. These subsets have the property of tilling the space when translated by lattice vectors. Of particular importance is the primitive cell U for the basis ai:

xiai | 0 xi < 1

In a material, the unit cell comprises a set of atomic positions S = {(Zi, xi) | xi 2 U}, where the integer Zi is the atomic number and xi the position of the atom. S can contain an arbitrary number of atoms and must not possess a particular structure. Together with the lattice , the atomic positions provides a complete description of the crystal structure.

It is often useful to deﬁne the concept of a sublattice. A sublattice P of is a lattice with basis vectors bi such that P . Correspondingly, the supercell CP is the unit cell associated with the sublattice P , for which CP U. The full lattice is generated by translations of the sublattice by a set of centring vectors {0 . . . vs}. More formally, the sublattice is associated with a normal subgroup T P of lattice translations, which speciﬁes a coset decomposition of the original lattice T = (0 + T P ) [ [ (vs + T P ). The centring vectors are coset representatives with respect to that decomposition. It is clear that the centering vectors are also the set of lattice points contained in the supercell CP .

Example 3.1 (Graphene) In graphene, carbon atoms are arranged in a two-dimensional crystal with honeycomb structure. The underlying lattice has basis vectors a1 = a 2

and a2 = a 2

, where a is a lattice constant. The set of atoms in the unit cell is S = &'

Figure 1: Graphene crystal structure

In Fig.1, blue dots represent carbon atoms. Black and white dots identify points of lattice with unit cell U. Black points identify a sublattice P deﬁned by basis vectors b1 and b2 and with supercell UP . The centring vectors {0, a1} correspond to lattice points contained in the supercell.

Space groups Symmetry plays a major role in the description of crystals. This symmetry is often directly visible through facets in naturally occurring crystals. Mathematically, it is described by a space group G, the set of isometries that maps a crystal structure to itself. As isometries, space groups are subgroups of the Euclidean group E (n). A space group element can be described as a tuple (W, t), where W is the linear part of the transformation and t a translation. An element maps a vector x 2 Rn to Wx + t. The multiplication of space group elements is therefore given by (W1, t1) (W2, t2) = (W1W2, W2t1 + t2). The point group P of a space group is the group obtained from linear part operations in G, which will in general be rotations and reﬂections. Considering only elements in G for which the linear part is identity (I, t), we obtain the translation subgroup of the space group T. It is a normal subgroup, which allows deﬁning the factor group G/T isomorphic to the point group P.

The crystallographic restriction theorem guarantees that only certain ﬁnite groups are valid point groups of space groups [17]. In particular, in 2 and 3 dimensions, only n-fold rotations with n 2 {2, 3, 4, 6} are allowed. Two space groups G and G0 are said to belong to be of same type if they can be related by a change of coordinate system. There are 17 space group types in 2 dimensions and 230 in 3 dimensions [59]. For a lattice , we call the Bravais group of the lattice P the set of linear isometries that map to itself. Bravais groups provide a way to classify lattices according to their

symmetry. We say that two lattices and 0 belong to the same Bravais type if their Bravais groups are the same matrix groups when written for suitable basis vectors of each lattice. The 5 Bravais types in 2 dimensions and the 14 in 3 dimensions are enumerated in the Appendix. The full symmetry group G of lattice a is given by the semidirect product of its translation group with its Bravais group : G = T o P .

Consider the space group G a crystal structure with lattice and unit cell U. From the deﬁnition of a crystal structure, it is clear that the translation subgroup of G will be T . However, the point group of G will, in general, not be P = P . This is because the unit cell may have less symmetry than the underlying lattice, which will result in P P (see example 3.2). This is the reason why the number of space groups is much larger than the number of Bravais lattices.

Example 3.2 (Wallpaper pattern) The wallpaper in Figure 2 is described by a square Bravais lattice and unit cell U with 4-fold rotational symmetry. The square lattice has a symmetry group P4m, with D4 Bravais group. However, the unit cell reduces the symmetry of the overall pattern since it does not have reﬂection axes. The symmetry group of the wallpaper is, therefore, P4, with point group C4.

Figure 2: Egyptian wallpaper [34]

4 Limitations of GNNs

In this section, we examine some of the limitations of GNNs on crystalline data, motivating our model by addressing these limitations. Typical models for crystalline data [70, 12] will build a graph from atoms in the unit cell and assign a feature vector hi 2 Rn to each. Edge features encoding distance might also be used. For convenience, this graph can be represented as a sparse tensor H 2 RN 2 n, where N is the number of atoms considered, and in which node and edge features are encoded on diagonal and off-diagonal entries, respectively. A GNN is a function approximator f : RN 2 n ! Y that produce a prediction, by successive application of layers φ: RN 2 n ! RN 2 n +1, where is the layer index.

Expressivity GNNs are usually stack of layers that are permutation equivariant

φ ((P(g) P(g))H) = (P(g) P(g))φ (H) , 8g 2 SN, H 2 RN 2 n, (3) where P(g) is the permutation matrix associated with the group element g. While P(g) permutes nodes, P(g) P(g) permutes the vectorized adjacency matrix. Our use of Kronecker product is due to the equality vec(PAP>) = (P P)vec(A) for A 2 RN N.

This captures the fact that these models are designed to be insensitive to node ordering in H. The material is, in a sense, treated as if it was a molecule. Much of the crystal structure is forgotten because it cannot be captured by ordering the nodes in a speciﬁc way and is only encoded in positions or distances. We argue that permutation equivariance is an overly strong requirement and was mainly used for practical reasons. When possible, it should be more beneﬁcial to use a model equivariant to the actual symmetry of the data; in this case, the crystal space groups, which will, in general, be much smaller than the symmetric group. Being equivariant to a smaller group results in less restrictive parameter sharing and a more expressive model.

Note that this requirement is different from that of E(3)-equivariance characteristic of some architectures [60, 53, 5]. In these models, the E(n) group only acts on the position xi 2 R3 part of each

atom s feature vector. By contrast, the symmetry group of a crystal structure maps the crystal to itself, and its action is a permutation. It therefore offers an alternative to building more powerful architectures for these systems that does not suppose that coordinate information is available. In particular, this approach is more suitable for abstract condensed matter systems, like spin and freefermion lattice models, in which coordinates are not relevant. Moreover, space groups are subgroups of the Euclidean group; equivariance to space groups is thus less restrictive. In this work, we choose to concentrate on space group equivariance, but we still provide an E(n)-equivariant version of our architecture, which is a straightforward extension of [53] in Appendix A.5. Note that equivariance (and not only invariance) is important, as expressive invariant functions can be built by composing equivariant layers with an output pooling layer. Equivariant functions can also be used to predict local properties like magnetization and charge distribution for example.

Invariance and symmetry breaking Even if the arrangement of atoms in a crystal is symmetric, this does not have to carry over to all local properties of the material. Spontaneous symmetry breaking is common in materials and crucial to describing phenomena such as magnetism and superconductivity [41, 6]. Building a graph only at the unit cell level does not allow to capture local properties that differ across unit cells. Using a supercell of multiple unit cells does not sufﬁce to solve this problem since permutation equivariant models have the property that equal input elements will be mapped to equal outputs elements [58, 73].

5 Input representation

Supercell Following the arguments of Section 4 we consider the set of atoms in a supercell instead of only in the unit cell, and add explicity symmetry breaking to increase the representational power of the model. Since this increases the computational complexity of the method, we choose to keep the supercells small and deﬁne them with sublattice vectors bi = 2ai. The supercell is therefore 8 times larger than the unit cell, with centring vectors

i miai | mi 2 {0, 1}

. For each atom, we build a feature vector with a one-hot encoding of the atomic number and a one-hot encoding identiﬁcation of the unit cell it belongs to using the corresponding centering vector. We keep track of the index of each atom within the unit cell, and the index of the atoms unit cell h(a,i) = [onehot (Zi) , onehot (a)], where a 2 {1, . . . , 8} and i 2 C, where C is the number of atoms in the unit cell. The encoding of the unit cell allows to break the symmetry between atoms mapped into each other by lattice translations.

Graph We construct a graph from atoms in the supercell, encoding relative distances between atoms as edge features. Inspired by [54], an edge feature vector is built from the distance between atoms dij = kxi xjk as eij = exp

. The vector of Gaussian centers µ and γ are hyperparameters. We use this approach to facilitate comparaison to previous works [54, 70], although Bessel encondings [37] could also be considered. It can be seen as soft" binning of interatomic distances. Using this approach, we use only distance features in contrast to methods that use position vectors. This has the beneﬁt of simplicity while still allowing a complete description of the input structure [66, 4, 63].

Sparsity has proven to be a useful inductive bias in graph representation learning [25], and it is also beneﬁcial in reducing computational complexity. However, atomic bonds are not unambiguously deﬁned in crystals [16, 1]. We choose to follow a similar approach to [32]: an edge is drawn between atoms if they share a Voronoi face and if the distance between atoms is smaller than the sum of atomic Cordero radii plus a cutoff = 0.5Å. This approach has the advantage of being physically sound and producing graphs that are relatively sparse. We provide more details on the graph-building strategy and compute metrics on the resulting graphs in Appendix A.1.

To preserve translational invariance for atoms at the boundary of the supercell, edges are initially also drawn to atoms outside the supercell. Then, if an edge points outside the supercell, its head is mapped to the corresponding representative node inside the supercell. This is analogous to circular padding in image processing.

(a) Identiﬁcation of the supercell. The sublattice vectors are bi = 2ai

(b) Voronoï tessellation.

(c) Drawing edges between atoms that share Voronoï faces. The additional condition related to the Cordero radius is not relevant here.

(d) Periodization of the graph. Edges that point to nodes outside the supercell are mapped back to the corresponding atom inside the supercell. Markings show identical atoms.

Figure 3: Building a graph for the graphene crystal structure from Example 3.1.

6 Equivariant crystal networks

Product groups Consider a crystal structure with space group G and in which each atom has a feature vector hi. Then the space group acts as a permutation of the input atoms

gh(a,i) = h( S g (a), U g (i))8g 2 G, (4)

g are the permutations associated with group element g on the supercells and within the unit cells respectively.

If a dataset contains only samples that share the same crystal structure, then a model equivariant with respect to G can readily be used. However, the case of a dataset with multiple different crystal structures that would be used in a typical supervised learning setting is more challenging for two reasons. First, samples may have different space groups, which would require different models, each be trained with a fraction of the data. Second, the group action may be different even for the same space group. This is because the unit cells may have different structures and numbers of atoms.

A solution to address these issues is to consider equivariance to a direct product of groups G SC, where the symmetry group of the lattice G acts across unit cells and SC, the symmetric group, acts within unit cells:

(g, h)h(a,i) = h( S g (a), U

h (i))8g 2 G , h 2 SC. (5)

In this way, differences in unit cell structures are dealt with by the symmetric group, where parametersharing can handle variable-sized inputs [72]. We still have to accommodate 14 different group actions G corresponding to the different Bravais lattices. To avoid using a different model for each lattice, we propose two groups to deal with all the lattices. The ﬁrst option is to consider the least symmetric Bravais lattice of primitive triclinic type and use its symmetry group P 1 = T o C2. This group is a subgroup of all the other lattice symmetry groups. The second option we consider is to simply use the symmetric group S that is the symmetric group across unit cells, which is an overgroup of all the lattice symmetry groups. This leaves us with a hierarchy of groups, with SN, the symmetric group over all atoms of a supercell, being the largest :

P 1 SC | {z } P 1-model

G SC S SC | {z } S -model

In our experiments, we use the two groups in this hierarchy for different levels of expressivity.

Equivariant message passing Having deﬁned the group action, we can now build the Equivariant Crystal Network (ECN). We will seek to use the message passing framework, which has demonstrated good performance on molecular data, and generalize it to obtain equivariance to other groups than the symmetric group. The update equations for message passing framework are

The idea is to deﬁne parameter-sharing patterns for functions φe and φh, such that there can be multiple versions while still retaining equivariance. Following [51], we ﬁrst deﬁne the parametersharing pattern of the set of input nodes N, with respect to group G as the colored bipartite graph (N, , β), with the edge-color function : N N ! {1, . . . , Ce} and node-color function β : N ! {1, . . . , Ch}. We also consider the action of the group G on edges as g (i, j) .= ( g (i) , g (j)) 8g 2 G. We deﬁne the orbit G (i, j) of edge (i, j) as the set of edges in which it

can be moved to by the group action : G (i, j) .= {g (i, j) | g 2 G}. A similar deﬁnition applies to the orbit of a node, G i .= { g (i) | g 2 G}.

We then make the following claim:

Claim 6.1 The layer deﬁned by the Ce functions φ (i,j)

e and the Cn functions φβ(i)

mij = φ (i,j)

is G-equivariant if the parameter-sharing pattern respects the equivariance condition:

(i, j) = (k, l) () (k, l) 2 G (i, j) , (11) β (i) = β (j) () j 2 G i. (12)

The proof of this claim follows in Appendix A.2. In words, the group action on the graph creates node and edge orbits, and we use a different copy of φe and φh for each edge and node orbit, respectively. The computational process for producing the pattern is to ﬁnd the orbit of G-action on the edges (nodes) [51], and the computational cost of this orbit-ﬁnding process grows linearly with the number of edges (nodes) [30].

This layer generalizes both MPNNs and equivariant multilayer perceptrons, such as CNNs. The MPNN is recovered with G = Sn and a standard CNN with circular convolution with G = T , φ (i,j)

= w (i,j) ht

i and φβ(i)

i, mi) = Re LU

We now consider a product group G H, acting according to Eq. (5). From Claim 1 of [64], the equivariant linear map for this group is the Kronecker product of equivariant maps for individual groups; see also [45]. The reformulation for parameter-sharing patterns is the following. If parameterssharing patterns 1 and 2 satisfy the equivariance condition for G and H respectively, then the parameter-sharing pattern = (N M, , β) satisﬁes the equivariant condition if

: N M N M ! {1, . . . , Ce,1} {1, . . . , Ce,2} , (a, i, b, j) = ( 1 (a, b) , 2 (i, j)) ,

β : N M ! {1, . . . , Ch,1} {1, . . . , Ch,2} , (13) β (a, i) = (β1 (a) , β2 (i)) . (14)

This simply means that a new color is deﬁned in the product pattern for each possible combination of colors in the original patterns. Example 6.1 demonstrates this idea with a simple example.

We use MLPs to build functions φ (i,j)

e and φβ(i)

h . The functions used in the experiments are detailed in Appendix A.4. In addition, we add a weighting factor to the edge aggregation 9, as this as been shown to be beneﬁcial by [70, 53]:

eijmij, where eij = φa (mij) , (15)

and φa is simply a linear layer. This change does not affect the equivariance of the model.

Example 6.1 In our running example, the parameter-sharing pattern for the message passing is produced by the Kronecker product of the pattern for the P2 group (see Appendix A.3 for a similar example with the P6m group) and the pattern for the symmetric group S2 as shown in the ﬁgure below (Top left). However, we only need to keep the colors for which there is a corresponding edge (Top right). Message passing is used on the resulting edge/node colored graph where similar colored nodes and edges use the same functions in message passing.

Figure 4: Equivariant message passing for graphene

7 Model implementation

Network architecture We aimed to keep the architecture of our ECN model as simple as possible. The model receives crystal structure graphs as input, with one-hot encoded feature vectors for each atom. We use 128-dimensional embeddings and keep the same dimension for hidden layers throughout the network. The ECN consists of 6 layers of the equivariant message passing operation 6.1. This is followed by a mean-pooling operation over node embeddings of each graph and a two-layer MLP that outputs the ﬁnal prediction. Mean pooling is preferred to other options because we only predict intensive physical quantities. These properties are "per-atom" and do not depend on the choice of supercell. Therefore, selecting a pooling operation that respects this invariance makes sense. If we were to predict extensive quantities, sum pooling would be the preferred option.

Input features Many choices to encode atomic features are possible and have been suggested in the literature. We perform experiments over these variants. For each atom, we encode its type in a feature vector. We consider two strategies, using only information available from the periodic table. The ﬁrst one is to use only the atomic number. The second one is to use the one-hot encoded group and period numbers for each atom. The second strategy could beneﬁt from promoting better generalization since embeddings of atoms belonging to the same group or period will show a certain similarity. Around 24% of the structures retained in the Materials Project dataset have been computed using the Hubbard-U extension of DFT. Since this information can signiﬁcantly inﬂuence the resulting properties [61] it is added to the atomic feature vector as a binary feature.

Other details We implemented our model using Pytorch [49]. We use a sparse implementation of the equivariant message-passing based on the Pytorch Scatter package [22]. We use the Adam W

optimizer [43] with weight decay regularization. The full hyperparameter setup is provided in Appendix A.7.

8 Experiments

Materials Project We perform experiments using the Materials Project dataset [33] 1. This standard dataset of materials informatics comprises more than 120K materials with a complete speciﬁcation of their crystal structure and some important physical properties obtained with high-throughput DFT calculations.

The Materials Project dataset has not been initially built to serve as a machine learning benchmark; we perform some preprocessing to make it suitable. First, since multiple DFT calculations were sometimes performed on close initial conﬁgurations, some samples only show marginal structure differences and resulting properties. This is exempliﬁed by the compound Li9Mn2Co5O16, which appears 322 times in the database. Such duplicates can result in training-test leakage. To prevent this, we consider two structures redundant if they have the same unit cell chemical formula and the same space group. Amongst a set of duplicate structures, the one with the lowest formation energy is chosen. Second, we ﬁlter out one-dimensional and two-dimensional materials from the dataset to only keep three-dimensional materials. Finally, we remove materials for which the unit cell contains more than 50 atoms. These are often associated with molecular and inorganic crystals with very different properties than the other materials. Training, validation, and test splits are 80%, 10%, and 10% of the dataset. We provide statistics on the processed dataset in Appendix A.6.

Following previous work, we predict a few relevant energetic properties: the formation energy E, the Fermi energy EF , and the band gap Eg for the subset of insulating materials. We also predict the binary insulator or conductor character material. Finally, we also predict the magnetic moment per atom M. This can be seen as a graph regression or classiﬁcation task. For regression, training is performed using the mean-squared error (MSE) loss function, but we report the mean-absolute error (MAE). For classiﬁcation, we use the cross-entropy loss function.

Table 1: Results on the Materials Project dataset.

Method Property

E(e V/atom) EF (e V) M (µB/atom) Eg (e V) Metal precision Nonmetal precision

CGCNN 0.039 0.363 - 0.388 80% 95% MEGNET 0.028 - - 0.33 78.9% 90.6% SCHNET 0.041 - - - - -

CGCNN 0.048 0.0002 0.307 0.001 0.111 0.001 0.399 0.006 81.2% 3.0 86.3% 3.0 MEGNET 0.056 0.0002 0.365 0.007 0.110 0.001 0.434 0.006 72.1% 3.0 81.6% 4.0

ECN-P 1 0.052 0.001 0.303 0.004 0.108 0.002 0.44 0.02 80% 4.0 84% 4.0 ECN-S 0.046 0.002 0.281 0.007 0.106 0.002 0.390 0.02 79.8% 2.0 83.2% 1.0

We compare the results obtained by our models to baselines [54, 70, 12]. Note that because these papers used different training and test splits and preprocessing schemes (even between themselves), our results cannot be directly compared. To alleviate that, we trained our own versions of two of the baselines using our splits and a similar training procedure. We obtain slightly better or comparable results to the baselines on all targets when evaluated on the same splits. The S version offers better performance than the P 1 model overall, showing that it is more beneﬁcial to lean on the side of having slightly more symmetry than necessary at the cost of some expressivity. We provide additional results for the model variants in Appendix A.8. The beneﬁts of the increased expressivity on this task is in the not crucial, which we think can be explained in part by the relatively small size of the Materials Project dataset. In a larger data regime, we expect that the beneﬁt of increased expressivity will outweight the cost in generalization capability.

Perov-5 Finally, we perform experiments using the Perov-5 dataset [9] as provided by [69]. In this dataset, all the materials share the same Perovskite crystal structure. The task considered is the regression of the heat of formation computed through DFT. Results are shown on Table 2.The improvement on this dataset is signiﬁcantly more important for the proposed model compared to the

1We use version 2021.05 of the dataset

baselines than on the Materials Project dataset. We hypothesize that the fact that all the structures are shared in this dataset allows the model to specialize more efﬁciently leading to better generalization.

Method Property

CGCNN 0.047 0.000 MEGNET 0.059 0.006 ECN-S 0.038 0.004

Table 2: Perov-5 results

We have shown how to leverage crystal symmetry to build more expressive and physically motivated neural networks for materials data. This allows us to obtain a close equivalent of group equivariant convolution on this data structure. These models show excellent accuracy in supervised property prediction, which supports the idea that symmetry is a useful inductive bias. Such models could be used for other tasks on materials such as dynamics prediction, if the dynamics approximately preserves the crystal structure. We also think that these models have signiﬁcant potential on more abstract condensed matter systems such as spin models and free-fermion models. We have also deﬁned equivariant message passing, a generalization of the MPNN framework that can potentially be used on any data structures for which a group can capture the symmetry in sparse interactions between the basic elements. One limitation of this approach is that it is not clear how to handle structures with different groups without using a larger group like S . A potential solution is drawing inspiration from the Natural Graph Neural Networks framework introduced in [27]. Another area of future improvement is on the computational efﬁciency of the equivariant message passing, which does not beneﬁt from optimized algorithms available for convolutions.

Acknowledgments and Disclosure of Funding

We thank Mehran Shakerinava, Christopher Morris, Joey Bose, Simon Verret and the anonymous reviewers for their valuable comments. This project is in part supported by the CIFAR AI chairs program and NSERC Discovery. S.-O. K. s research is also supported by IVADO and the Deep Mind Scholarship. Computational resources were provided by Mila and Compute Canada.

[1] Alvarez Santiago. A cartography of the van der Waals territories // Dalton Trans. 2013. 42.

[2] Anderson Brandon, Hy Truong Son, Kondor Risi. Cormorant: Covariant molecular neural

networks // Advances in neural information processing systems. 2019. 32.

[3] Ashcroft Neil W, Mermin N David, others . Solid state physics. 1976.

[4] Bartók Albert P, Kondor Risi, Csányi Gábor. On representing chemical environments // Physical

Review B. 2013. 87, 18. 184115.

[5] Batzner Simon, Musaelian Albert, Sun Lixin, Geiger Mario, Mailoa Jonathan P., Kornbluth

Mordechai, Molinari Nicola, Smidt Tess E., Kozinsky Boris. E(3)-Equivariant Graph Neural Networks for Data-Efﬁcient and Accurate Interatomic Potentials. 2021.

[6] Beekman Aron, Rademaker Louk, Wezel Jasper van. An introduction to spontaneous symmetry

breaking // Sci Post Physics Lecture Notes. 2019. 011.

[7] Bengio Y., Courville A., Vincent P. Representation Learning: A Review and New Perspectives //

IEEE Transactions on Pattern Analysis and Machine Intelligence. Aug 2013. 35, 8. 1798 1828.

[8] Bronstein Michael M, Bruna Joan, Cohen Taco, Velickovi c Petar. Geometric deep learning:

Grids, groups, graphs, geodesics, and gauges // ar Xiv preprint ar Xiv:2104.13478. 2021.

[9] Castelli Ivano E, Landis David D, Thygesen Kristian S, Dahl Søren, Chorkendorff Ib, Jaramillo

Thomas F, Jacobsen Karsten W. New cubic perovskites for one-and two-photon water splitting

using the computational materials repository // Energy & Environmental Science. 2012. 5, 10. 9034 9043.

[10] Cesa Gabriele, Lang Leon, Weiler Maurice. A Program to Build E (N)-Equivariant Steerable

CNNs // International Conference on Learning Representations. 2021.

[11] Chanussot* Lowik, Das* Abhishek, Goyal* Siddharth, Lavril* Thibaut, Shuaibi* Muhammed,

Riviere Morgane, Tran Kevin, Heras-Domingo Javier, Ho Caleb, Hu Weihua, Palizhati Aini, Sriram Anuroop, Wood Brandon, Yoon Junwoong, Parikh Devi, Zitnick C. Lawrence, Ulissi Zachary. Open Catalyst 2020 (OC20) Dataset and Community Challenges // ACS Catalysis. 2021.

[12] Chen Chi, Ye Weike, Zuo Yunxing, Zheng Chen, Ong Shyue Ping. Graph Networks as a Universal

Machine Learning Framework for Molecules and Crystals // Chemistry of Materials. 2019. 31, 9. 3564 3572.

[13] Cohen Taco, Welling Max. Group equivariant convolutional networks // International conference

on machine learning. 2016. 2990 2999.

[14] Cohen Taco S., Geiger Mario, Köhler Jonas, Welling Max. Spherical CNNs // International

Conference on Learning Representations. 2018.

[15] Cohen Taco S, Geiger Mario, Weiler Maurice. A general theory of equivariant cnns on homogeneous spaces // Advances in neural information processing systems. 2019. 32.

[16] Cordero Beatriz, Gómez Verónica, Platero-Prats Ana E., Revés Marc, Echeverría Jorge, Cre-

mades Eduard, Barragán Flavia, Alvarez Santiago. Covalent radii revisited // Dalton Trans. 2008. 2832 2838.

[17] Coxeter H. S. M. Introduction to geometry. New York: John Wiley & Sons Inc., 1961. xvii+443.

[18] Curtarolo Stefano, Setyawan Wahyu, Hart Gus L.W., Jahnatek Michal, Chepulskii Roman V.,

Taylor Richard H., Wang Shidong, Xue Junkai, Yang Kesong, Levy Ohad, Mehl Michael J., Stokes Harold T., Demchenko Denis O., Morgan Dane. AFLOW: An automatic framework for high-throughput materials discovery // Computational Materials Science. 2012. 58. 218 226.

[19] Dehmamy Nima, Walters Robin, Liu Yanchen, Wang Dashun, Yu Rose. Automatic Symme-

try Discovery with Lie Algebra Convolutional Network // Advances in Neural Information Processing Systems. 2021. 34.

[20] Dresselhaus Mildred S, Dresselhaus Gene, Jorio Ado. Group theory: application to the physics

of condensed matter. 2007.

[21] Esteves Carlos, Makadia Ameesh, Daniilidis Kostas. Spin-weighted spherical cnns // Advances

in Neural Information Processing Systems. 2020. 33. 8614 8625.

[22] Fey Matthias. Py Torch Scatter. 2022.

[23] Finkelshtein Ben, Baskin Chaim, Maron Haggai, Dym Nadav. A simple and universal rotation

equivariant point-cloud network // ar Xiv preprint ar Xiv:2203.01216. 2022.

[24] Finzi Marc, Welling Max, Wilson Andrew Gordon. A Practical Method for Constructing Equiv-

ariant Multilayer Perceptrons for Arbitrary Matrix Groups // ar Xiv preprint ar Xiv:2104.09459. 2021.

[25] Garg Vikas, Jegelka Stefanie, Jaakkola Tommi. Generalization and representational limits of

graph neural networks // International Conference on Machine Learning. 2020. 3419 3430.

[26] Gilmer Justin, Schoenholz Samuel S., Riley Patrick F., Vinyals Oriol, Dahl George E. Neural

Message Passing for Quantum Chemistry // Proceedings of the 34th International Conference on Machine Learning. 70. International Convention Centre, Sydney, Australia: PMLR, 06 11 Aug 2017. 1263 1272. (Proceedings of Machine Learning Research).

[27] Haan Pim de, Cohen Taco S, Welling Max. Natural Graph Networks // Advances in Neural

Information Processing Systems. 33. 2020. 3636 3646.

[28] Hasnip Philip J., Refson Keith, Probert Matt I. J., Yates Jonathan R., Clark Stewart J., Pickard

Chris J. Density functional theory in the solid state // Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences. 2014. 372, 2011. 20130270.

[29] Hiß Gerhard, Holt Derek F, Newman Michael F. Computational Group Theory // Oberwolfach

Reports. 2007. 3, 3. 1795 1878.

[30] Holt Derek F, Eick Bettina, O Brien Eamonn A. Handbook of computational group theory.

[31] Horton Matthew Kristofer, Montoya Joseph Harold, Liu Miao, Persson Kristin Aslaug. High-

throughput prediction of the ground-state collinear magnetic order of inorganic materials using Density Functional Theory // npj Computational Materials. jun 2019. 5, 1.

[32] Isayev Olexandr, Oses Corey, Toher Cormac, Gossett Eric, Curtarolo Stefano, Tropsha Alexan-

der. Universal fragment descriptors for predicting properties of inorganic crystals // Nature Communications. Jun 2017. 8, 1.

[33] Jain Anubhav, Ong Shyue Ping, Hautier Geoffroy, Chen Wei, Richards William Davidson, Dacek

Stephen, Cholia Shreyas, Gunter Dan, Skinner David, Ceder Gerbrand, al. et. Commentary: The Materials Project: A materials genome approach to accelerating materials innovation // APL Materials. Jul 2013. 1, 1. 011002.

[34] Jones Owen, Bedford Francis, Waring J. B., Westwood J. O., Wyatt M. Digby. The Grammar

of Ornament. London: Published by Day and Son, Lithographers to the Queen, Gate Street, Lincoln s Inn Fields, 1856.

[35] Kipf Thomas N., Welling Max. Semi-Supervised Classiﬁcation with Graph Convolutional

Networks // 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. 2017.

[36] Kirklin Scott, Saal James E, Meredig Bryce, Thompson Alex, Doak Jeff W, Aykol Muratahan,

Rühl Stephan, Wolverton Chris. The Open Quantum Materials Database (OQMD): assessing the accuracy of DFT formation energies // npj Computational Materials. 2015. 1. 15010.

[37] Klicpera Johannes, Groß Janek, Günnemann Stephan. Directional Message Passing for Molec-

ular Graphs // International Conference on Learning Representations. 2020.

[38] Kohn W., Sham L. J. Self-Consistent Equations Including Exchange and Correlation Effects //

Phys. Rev. Nov 1965. 140. A1133 A1138.

[39] Kondor Risi, Son Hy Truong, Pan Horace, Anderson Brandon, Trivedi Shubhendu. Covariant

compositional networks for learning graphs // ar Xiv preprint ar Xiv:1801.02144. 2018.

[40] Kondor Risi, Trivedi Shubhendu. On the Generalization of Equivariance and Convolution in

Neural Networks to the Action of Compact Groups // Proceedings of the 35th International Conference on Machine Learning. 80. 10 15 Jul 2018. 2747 2755. (Proceedings of Machine Learning Research).

[41] Landau Lev Davidovich. On the theory of phase transitions. I. // Phys. Z. Sowjet. 1937. 11. 26.

[42] Le Cun Yann, Bengio Yoshua, others . Convolutional networks for images, speech, and time

series // The handbook of brain theory and neural networks. 1995. 3361, 10. 1995.

[43] Loshchilov Ilya, Hutter Frank. Decoupled Weight Decay Regularization // International

Conference on Learning Representations. 2019.

[44] Maron Haggai, Ben-Hamu Heli, Shamir Nadav, Lipman Yaron. Invariant and Equivariant Graph

Networks // International Conference on Learning Representations. 2019.

[45] Maron Haggai, Litany Or, Chechik Gal, Fetaya Ethan. On learning sets of symmetric elements

// International Conference on Machine Learning. 2020. 6734 6744.

[46] Noé Frank, Tkatchenko Alexandre, Müller Klaus-Robert, Clementi Cecilia. Machine Learning

for Molecular Simulation // Annual Review of Physical Chemistry. 2020. 71, 1. 361 390. PMID: 32092281.

[47] O Mara Jordan, Meredig Bryce, Michel Kyle. Materials Data Infrastructure: A Case Study of

the Citrination Platform to Examine Data Import, Storage, and Access // JOM. Jun 2016. 68, 8. 2031 2034.

[48] Ong Shyue Ping, Richards William Davidson, Jain Anubhav, Hautier Geoffroy, Kocher Michael,

Cholia Shreyas, Gunter Dan, Chevrier Vincent L., Persson Kristin A., Ceder Gerbrand. Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis // Computational Materials Science. 2013. 68. 314 319.

[49] Paszke Adam, Gross Sam, Massa Francisco, Lerer Adam, Bradbury James, Chanan Gregory,

Killeen Trevor, Lin Zeming, Gimelshein Natalia, Antiga Luca, others . Pytorch: An imperative style, high-performance deep learning library // Advances in neural information processing systems. 2019. 32. 8026 8037.

[50] Qi Charles R, Su Hao, Mo Kaichun, Guibas Leonidas J. Pointnet: Deep learning on point

sets for 3d classiﬁcation and segmentation // Proceedings of the IEEE conference on computer vision and pattern recognition. 2017. 652 660.

[51] Ravanbakhsh Siamak, Schneider Jeff G., Póczos Barnabás. Equivariance Through Parameter-

Sharing // ICML. 2017. 2892 2901.

[52] Rotman Joseph J. An introduction to the theory of groups. 148. 2012.

[53] Satorras Victor Garcia, Hoogeboom Emiel, Welling Max. E (n) equivariant graph neural

networks // ar Xiv preprint ar Xiv:2102.09844. 2021.

[54] Schütt Kristof, Kindermans Pieter-Jan, Felix Huziel Enoc Sauceda, Chmiela Stefan, Tkatchenko

Alexandre, Müller Klaus-Robert. Schnet: A continuous-ﬁlter convolutional neural network for

modeling quantum interactions // Advances in Neural Information Processing Systems. 2017. 991 1001.

[55] Shakerinava Mehran, Ravanbakhsh Siamak. Equivariant Networks for Pixelized Spheres //

Proceedings of the 38th International Conference on Machine Learning. 139. 18 24 Jul 2021. 9477 9488. (Proceedings of Machine Learning Research).

[56] Shakerinava Mehran, Ravanbakhsh Siamak. Equivariant networks for pixelized spheres //

International Conference on Machine Learning. 2021. 9477 9488.

[57] Shawe-Taylor J. Building symmetries into feedforward networks // 1989 First IEE International

Conference on Artiﬁcial Neural Networks, (Conf. Publ. No. 313). 1989. 158 162.

[58] Smidt Tess E., Geiger Mario, Miller Benjamin Kurt. Finding symmetry breaking order parame-

ters with Euclidean neural networks // Phys. Rev. Research. Jan 2021. 3. L012002.

[59] Souvignier B. A general introduction to space groups // International Tables for Crystallography.

2016. 1.3, 22 41.

[60] Thomas Nathaniel, Smidt Tess, Kearnes Steven, Yang Lusann, Li Li, Kohlhoff Kai, Riley Patrick.

Tensor ﬁeld networks: Rotation-and translation-equivariant neural networks for 3d point clouds // ar Xiv preprint ar Xiv:1802.08219. 2018.

[61] Tolba Sarah A., Gameel Kareem M., Ali Basant A., Almossalami Hossam A., Allam Nageh K.

The DFT+U: Approaches, Accuracy, and Applications // Density Functional Calculations. Rijeka: Intech Open, 2018. 1.

[62] Velickovi c Petar, Cucurull Guillem, Casanova Arantxa, Romero Adriana, Liò Pietro, Bengio

Yoshua. Graph Attention Networks // International Conference on Learning Representations.

[63] Villar Soledad, Hogg David W, Storey-Fisher Kate, Yao Weichi, Blum-Smith Ben. Scalars are

universal: Equivariant machine learning, structured like classical physics // Advances in Neural Information Processing Systems. 2021.

[64] Wang Renhao, Albooyeh Marjan, Ravanbakhsh Siamak. Equivariant Maps for Hierarchical

Structures. 2020.

[65] Weiler Maurice, Geiger Mario, Welling Max, Boomsma Wouter, Cohen Taco. 3d steerable cnns:

Learning rotationally equivariant features in volumetric data // Advances in Neural Information Processing Systems. 2018. 10381 10392.

[66] Weyl Hermann. The Classical Groups. 1939.

[67] Wolpert D.H., Macready W.G. No free lunch theorems for optimization // IEEE Transactions

on Evolutionary Computation. 1997. 1, 1. 67 82.

[68] Worrall Daniel E, Garbin Stephan J, Turmukhambetov Daniyar, Brostow Gabriel J. Harmonic

networks: Deep translation and rotation equivariance // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017. 5028 5037.

[69] Xie Tian, Fu Xiang, Ganea Octavian-Eugen, Barzilay Regina, Jaakkola Tommi. Crystal Diffusion Variational Autoencoder for Periodic Material Generation // ar Xiv preprint ar Xiv:2110.06197. 2021.

[70] Xie Tian, Grossman Jeffrey C. Crystal Graph Convolutional Neural Networks for an Accurate

and Interpretable Prediction of Material Properties // Physical Review Letters. Apr 2018. 120, 14.

[71] Xu Keyulu, Hu Weihua, Leskovec Jure, Jegelka Stefanie. How Powerful are Graph Neural

Networks? // International Conference on Learning Representations. 2019.

[72] Zaheer Manzil, Kottur Satwik, Ravanbakhsh Siamak, Poczos Barnabas, Salakhutdinov Russ R,

Smola Alexander J. Deep Sets // Advances in Neural Information Processing Systems 30. 2017. 3391 3401.

[73] Zhang Yan, Zhang David W, Lacoste-Julien Simon, Burghouts Gertjan J., Snoek Cees G. M.

Multiset-Equivariant Set Prediction with Approximate Implicit Differentiation // International Conference on Learning Representations. 2022.

1. For all authors...

(a) Do the main claims made in the abstract and introduction accurately reﬂect the paper s

contributions and scope? [Yes] On both the theoretical and experimental side, the paper s contributions are aligned with the abstract and introduction. (b) Did you describe the limitations of your work? [Yes] See Conclusion

(c) Did you discuss any potential negative societal impacts of your work? [N/A] We do

not think that this is relevant to this work (d) Have you read the ethics review guidelines and ensured that your paper conforms to

them? [Yes] 2. If you are including theoretical results...

(a) Did you state the full set of assumptions of all theoretical results? [Yes] See Theorem 6.1

and its proof in the Appendix A.2 (b) Did you include complete proofs of all theoretical results? [Yes] The only proof is

included in Appendix A.2 3. If you ran experiments...

(a) Did you include the code, data, and instructions needed to reproduce the main exper-

imental results (either in the supplemental material or as a URL)? [Yes] The details needed to reproduce the experiments are included in the supplementary material Appendix A.7. A link to the relevant code will also be accessible upon publication of the paper. (b) Did you specify all the training details (e.g., data splits, hyperparameters, how they

were chosen)? [Yes] The details needed to reproduce the experiments are included in the supplementary material Appendix A.7. (c) Did you report error bars (e.g., with respect to the random seed after running experi-

ments multiple times)? [Yes] (d) Did you include the total amount of compute and the type of resources used (e.g.,

type of GPUs, internal cluster, or cloud provider)? [Yes] This is provided in the supplementary material Appendix A.7 4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...

(a) If your work uses existing assets, did you cite the creators? [Yes] See Introduction and

Model implementation sections (b) Did you mention the license of the assets? [Yes] This is mentioned in the supplementary

material Appendix A.6 (c) Did you include any new assets either in the supplemental material or as a URL? [No]

The used dataset will be provided upon publication of the paper. (d) Did you discuss whether and how consent was obtained from people whose data you re

using/curating? [N/A] This is not relevant according to the license of the used assets. (e) Did you discuss whether the data you are using/curating contains personally identiﬁable

information or offensive content? [N/A] This is not relevant for this paper. 5. If you used crowdsourcing or conducted research with human subjects...

(a) Did you include the full text of instructions given to participants and screenshots, if

applicable? [N/A] (b) Did you describe any potential participant risks, with links to Institutional Review

Board (IRB) approvals, if applicable? [N/A] (c) Did you include the estimated hourly wage paid to participants and the total amount

spent on participant compensation? [N/A]