betaFeedback

WorksPhysics and Astronomy

StrainTensorNet: Predicting crystal structure elastic properties using SE(3)-equivariant graph neural networks

Abstract

Accurately predicting the elastic properties of crystalline solids is vital for computational materials science. However, traditional atomistic scale ab initio approaches are computationally intensive, especially for studying complex materials with a large number of atoms in a unit cell. We introduce a novel data-driven approach to efficiently predict the elastic properties of crystal structures using SE(3)-equivariant graph neural networks (GNNs). This approach yields important scalar elastic moduli with the accuracy comparable to recent data-driven studies. Importantly, our symmetry-aware GNNs model also enables the prediction of the strain energy density (SED) and the associated elastic constants, the fundamental tensorial quantities that are significantly influenced by a material's crystallographic group. The model consistently distinguishes independent elements of SED tensors, in accordance with the symmetry of the crystal structures. Finally, our deep learning model possesses meaningful latent features, offering an interpretable prediction of the elastic properties.

Open data: The dataset used in this work has been made available at https://github.com/trachote/predict_elastic_tensor

Introduction

The elastic properties of crystalline solids such as elastic constants, bulk modulus, and shear modulus are important macroscopic quantities that determine materials’ mechanical characteristics. Computational study of which can provide theoretical guidelines to understand various phenomena in solid materials, e.g., mechanical stability of crystal structures [1, 2], pressure-driven deformation and phase transition of materials [3, 4], the response of materials to sound wave propagation [5, 6], and the hardness of materials [7, 8, 9], to name a few. From an atomistic scale description, an ab initio approach based on density functional theory has been employed to investigate the macroscopic elastic properties of materials, yielding, for example, elastic constants and elastic moduli that are in agreement with experimental results [10, 11, 12, 13].

Three atomistic scale computational methods are usually adopted to calculate the elastic constants of crystal structures: an energy-based method [14], a stress-strain method [10, 15], and a calculation from group velocity [16, 17, 18]. The energy-based and the stress-strain methods are more standard, and derive the elastic constants using respectively an energy and a stress tensor acquired from an ab initio calculation. Although both methods yield a comparable prediction of elastic constants, the energy-based method is computationally less efficient, involving a larger number of plane-wave basis set and kk-point meshes so that the predicted elastic constants converge to reliable values [19]. Indeed, the more efficient stress-strain method is commonly employed to determine elastic constants by established softwares such as VASP, Quantum Espresso and CASTEP [20, 21, 22]. The stress-strain method is also utilized to create a database for elastic properties of inorganic compounds [23, 24].

However, atomistic scale simulation can be computationally prohibitive, especially when the number of atoms in a unit cell grows large. Such constraint limits an ab initio method’s capability to investigate more complex crystalline solids. On the other hand, advances in machine learning bring about alternative approaches to computational materials science. This data-driven paradigm can reasonably predict the elastic properties of crystal structures, provided sufficient training data from an ab initio method [25, 26]. Even when the number of atoms in a unit cell is large, such approach can efficiently predict the bulk and shear moduli of complex materials such as alloys [27, 28]. Applying machine learning together with ab initio calculations is also potentially useful in searching for novel superhard materials [29, 30].

Machine learning models based on graph neural networks (GNNs) have received increasing attention in studying solid materials and molecules. With GNNs, it is natural to encode atomistic scaled descriptions of solids into a computational framework; atomic locations and the associated atomic attributes can be directly embedded into node attributes of a graph, whereas pairwise interactions among atoms can be encoded into the edge attributes. Efficient GNNs training procedures have also been proposed [31, 32, 33], enabling GNNs to learn the representation of a complex relationship between the input and its associated prediction [34]. These neural networks can also be endowed with translation-rotation equivariant property, so that input atomic locations of molecules (or point clouds) in R3\mathbb{R}^{3} that differ only in their orientation or their centroid can be identified. Enforcing an SE(3) equivariant property helps the networks to extract a more compact (translation and rotation independent) relationship between the inputs and their predictions. Due to these appealing features, variations of SE(3) equivariant GNNs have been developed to study materials science [35].

In this work, we adopt a data-driven approach using graph neural networks (GNNs) to predict the elastic properties of crystal structures. Accounting for the symmetry of crystalline solids, our SE(3) equivariant GNNs take as an input atomistic descriptions of strained materials, such as atomic locations and their associated atomic attributes, and predict the strain energy density (SED) of the materials. The prediction of the SED, which is the energy stored in a crystal structure when distorted by either tensile or compressive strains in particular directions, can be obtained efficiently and relatively accurately, given sufficient training data. The model thus provides an alternative approach to the standard ab initio energy-based prediction method.

After the SED is computed, we can then calculate the elastic constants and construct the elastic tensor with a simple analytical expression, see § 1.1 A. Other macroscopic (scalar) elastic properties including bulk modulus, shear modulus, Young’s modulus, and Poisson’s ratio immediately follow from the elastic constants. Sec. IV reports our model prediction results of these elastic properties. The prediction performances of the scalar elastic properties are comparable to those of the recent data-driven work [25, 26]. Importantly, however, the data-driven prediction of the SED and the associated elastic constants, which are fundamental tensorial quantities that depend on the crystallographic group, are first reported here. The trained model consistently reveal the independent components of strain energy tensors dictated by the symmetry of the crystal structures, as evident by the degeneracy structure of the distance metrics computed in Sec. IV B. Our symmetry-aware GNNs architecture as well as the dataset used to train the model are provided in Sec. III. Also, as opposed to a black-box deep learning prediction, we show in Sec. IV C that the latent features (data representations) that the model learned are physically meaningful, rendering our GNNs explainable. We conclude this work in § 2, and provide supplementary results and necessary derivations in the Appendices.

Linear Elasticity Background

The strain energy tensor and the elastic tensor

Assuming linear elastic deformation and no external stress, the strain energy density (SED) of isotropic materials can be expressed as [36]

U(ϵ1,ϵ2,…,ϵ6)=12∑i,j=16Cijϵiϵj,(1)U(\epsilon_1,\epsilon_2,\ldots,\epsilon_6)=\frac{1}{2}\sum_{i,j=1}^{6}C_{ij}\epsilon_i\epsilon_j, \tag*{(1)}

where CijC_{ij} is an elastic constant and ϵi\epsilon_i is a strain component i∈{1,2,…,6}i \in\{1,2,\ldots,6\} in Voigt notation. With the goal of obtaining the elastic constant, it suffices to focus on the SED tensor of rank 2, which we now show. Denote Uij≡U(ϵi,ϵj)U_{ij} \equiv U(\epsilon_i,\epsilon_j) for i≠ji \ne j and Uii≡U(ϵi)U_{ii} \equiv U(\epsilon_i), which are the energy stored in a distorted crystal structure when the lattice is strained by at most two strain components (two Voigt indices). In this study, we consider UijU_{ij} in units of eV per atom. Since the elastic constant tensor is symmetric, the SED tensor is also symmetric and can be expressed as

Uij=Cijϵiϵj+12(Ciiϵi2+Cjjϵj2)for i≠j and,U_{ij}=C_{ij}\epsilon_i\epsilon_j+\frac{1}{2}\left(C_{ii}\epsilon_i^2+C_{jj}\epsilon_j^2\right)\quad\text{for }i\ne j\text{ and,}
Uii=12Ciiϵi2.(2)U_{ii}=\frac{1}{2}C_{ii}\epsilon_i^2. \tag*{(2)}

Thus, the elastic constants can then be analytically obtained as a function of the SED:

Cij=1ϵiϵj[(1+δij)Uij−(1−δij)(Uii+Ujj)],(3)C_{ij}=\frac{1}{\epsilon_i\epsilon_j}\left[(1+\delta_{ij})U_{ij}-(1-\delta_{ij})(U_{ii}+U_{jj})\right], \tag*{(3)}

where δij\delta_{ij} is the Kronecker delta. Because the SED tensor of interest and the elastic tensor are symmetric tensors of rank 2 with 6×66\times6 components, so there are (6+1)6/2=21(6+1)6/2=21 independent components. In the matrix form, we denote the SED tensor as

U=[U11U12U13U14U15U16U22U23U24U25U26U33U34U35U36U44U45U46U55U56U66],(4)\mathbf{U}= \begin{bmatrix} U_{11} & U_{12} & U_{13} & U_{14} & U_{15} & U_{16}\\ & U_{22} & U_{23} & U_{24} & U_{25} & U_{26}\\ & & U_{33} & U_{34} & U_{35} & U_{36}\\ & & & U_{44} & U_{45} & U_{46}\\ & & & & U_{55} & U_{56}\\ & & & & & U_{66} \end{bmatrix}, \tag*{(4)}

where only the upper triangular elements are shown (the lower diagonal elements are left blank for brevity). For crystalline solids, however, the upper triangular elements are not completely independent, depending on the symmetry of the crystal structure. For instance, the SED tensor of a cubic lattice is

Ucubic=[U11U12U12U14U14U14U11U12U14U14U14U11U14U14U14U44U45U45U44U45U44].(5)\mathbf{U}_{\mathrm{cubic}}= \begin{bmatrix} U_{11} & U_{12} & U_{12} & U_{14} & U_{14} & U_{14}\\ & U_{11} & U_{12} & U_{14} & U_{14} & U_{14}\\ & & U_{11} & U_{14} & U_{14} & U_{14}\\ & & & U_{44} & U_{45} & U_{45}\\ & & & & U_{44} & U_{45}\\ & & & & & U_{44} \end{bmatrix}. \tag*{(5)}

Note that some components of the elastic constants CijC_{ij} can be zero in many crystal structures, e.g. a cubic lattice has 9 zeros out of 21 independent components. However, due to the property of SED from Eq. (2), UijU_{ij} is never 0. For the purpose of machine learning regression, working with SED helps avoid zero inflation problems in the training dataset. Figure E1 shows the distributions of CijC_{ij} and UijU_{ij}, which cover different ranges depending on the indices. Importantly, the distribution of elastic constants can concentrate around zero, but this is not the case for the SED. As an illustrative example of how the SED can avoid a zero inflation problem, consider the elastic constants of a diamond cubic crystal structure. For diamond, C11C_{11}, C12C_{12}, C14C_{14}, C44C_{44}, and C45C_{45} are 1054, 126, 0, 562, and 0 GPa, respectively. Then, Eq. (2) gives U11U_{11}, U12U_{12}, U14U_{14}, U44U_{44}, and U45U_{45} subjected to 2% strain to be 7.506, 16.807, 11.509, 4.002, and 8.005 meV/atom, respectively. Note that, in this study, we consider UijU_{ij} in the unit of eV per atom instead of per volume. The magnitude of U12U_{12} is the largest as it is the sum of C12C_{12} and C11C_{11}, while U14U_{14} and U45U_{45} are smaller (but non-zero) because C14C_{14} and C45C_{45} are zero.

Histograms of elastic constants and strain energy densities

Figure E1. The histograms of (a) CijC_{ij} and of (b) UijU_{ij} in the training set and of (c) CijC_{ij} and of (d) UijU_{ij} in the test set. For visualization purposes, only some ijij components, specifically 11, 12, 14, 33, 44, and 45, are presented. For each ijij component, UijU_{ij} (CijC_{ij}) are distributed over different range of energy (modulus). For i≠ji \ne j, UijU_{ij} are distributed over a broader range than UiiU_{ii} are. For ij=45ij = 45, the elastic constant is concentrated around 0 GPa. Importantly, this concentration around zero issue can be ameliorated when the elastic constants are converted to the strain energy density. This conversion to SED reduces the number of zero values in the dataset, which significantly improves model training and model predictive performances.

Strain operations

The unit cell of a crystal can be compactly parametrized by a lattice matrix L=[a b c]\mathbf{L} = [\mathbf{a}\ \mathbf{b}\ \mathbf{c}], where a\mathbf{a}, b\mathbf{b}, and c\mathbf{c} are vectors of lattice parameters in the Cartesian coordinate. If the system is applied by strain, the lattice matrix will be deformed by

L′(ϵ1,ϵ2,…,ϵ6)=εIL,(6)\mathbf{L}'(\epsilon_1,\epsilon_2,\ldots,\epsilon_6)=\boldsymbol{\varepsilon}_{I}\mathbf{L}, \tag*{(6)}

where εI\boldsymbol{\varepsilon}_{I} is the strain matrix

εI=[ϵ1ϵ62ϵ52ϵ62ϵ2ϵ42ϵ52ϵ42ϵ3]+I.(7)\boldsymbol{\varepsilon}_{I} = \begin{bmatrix} \epsilon_1 & \frac{\epsilon_6}{2} & \frac{\epsilon_5}{2} \\ \frac{\epsilon_6}{2} & \epsilon_2 & \frac{\epsilon_4}{2} \\ \frac{\epsilon_5}{2} & \frac{\epsilon_4}{2} & \epsilon_3 \end{bmatrix}+\mathbf{I}. \tag*{(7)}

In this work, we assume the crystal structures will be deformed by at most two strain components, so that L′=L′(ϵi,ϵj)\mathbf{L}'=\mathbf{L}'(\epsilon_i,\epsilon_j).

Due to an applied strain, the atomic coordinates r\mathbf{r} must also be transformed accordingly as

r′=εILrf,(8)\mathbf{r}'=\boldsymbol{\varepsilon}_{I}\mathbf{L}\mathbf{r}_{f}, \tag*{(8)}

where rf\mathbf{r}_{f} is a fractional coordinate of an atom in an unstrained lattice. Noting that rf\mathbf{r}_{f} typically shifts from its equilibrium value when the strain is applied (say, in density function theory or in experiments); however, the atomic relaxation is neglected for simplicity.

Machine Learning with SE(3)-Equivariant Graph Neural Networks

Crystal graphs

A crystal structure can be represented as a multigraph whose nodes and edges, respectively, encode atoms and their pairwise connections. A pair of nodes describing an atom of type mm and an atom of type nn can be connected by multiple edges, encapsulating the interactions between the atom of type mm in a unit cell and atoms of type nn in the unit cell as well as in other periodic cells of consideration, see Figure 1. An atom of type mm can interact with an atom of the same type in periodic cells if the interaction range of consideration is large enough, represented by self-loops in the node mm of the crystal graph. Each atom’s location and its atomic features (e.g., atomic mass, electronegativity, polarizability, atomic radius, and etc.) in a unit cell is accounted for by the attributes of a single node.

A 2-dimensional lattice and its strained crystal graph embedding

Figure 1. A 2-dimensional lattice and its strained crystal graph embedding. Grey dashed lines are periodic boundaries. The whole structure can be spanned by the 3 atoms in 3 different bases in the unit cell, represented by 3 different nodes. Black lines are (undirected) edges connecting two neighbor nodes. The strained lattice is represented as a multigraph on the right, which is the crystal graph input into our SE(3)\mathrm{SE}(3)-equivariant GNN building block. Note that to describe the crystal graph for the unit cell and the atoms on its boundary, only three nodes are required. The multi-edges of each pair of nodes are distinguished by different crystallographic directions.

More formally, a crystal graph G=(V,E)\mathcal{G}=(\mathcal{V},\mathcal{E}) consists of a set of nodes (vertices) V\mathcal{V} and a set of edges E\mathcal{E} defined as

V={(fn,rn)∣fn∈RM, rn=Lrfn∈R3},\mathcal{V}=\{(\mathbf{f}_n,\mathbf{r}_n)\mid\mathbf{f}_n\in\mathbb{R}^{M},\ \mathbf{r}_n=\mathbf{L}\mathbf{r}_{f_n}\in\mathbb{R}^{3}\},
E={Δrmn(T)∣Δrmn(T)=rm−rn+T; rm,rn∈R3},\mathcal{E}=\{\Delta\mathbf{r}_{mn}^{(\mathbf{T})}\mid\Delta\mathbf{r}_{mn}^{(\mathbf{T})}=\mathbf{r}_m-\mathbf{r}_n+\mathbf{T};\ \mathbf{r}_m,\mathbf{r}_n\in\mathbb{R}^{3}\},

where mm and nn are indices of the two atoms in the unit cell, fn\mathbf{f}_n is a vectorized atomic feature of an atom of type nn, MM is the number of atomic features, rfn\mathbf{r}_{f_n} is the fractional coordinate of an atom of type nn, and T\mathbf{T} is a translation vector within the space of interest. Each node embeds two kinds of information: atomic attributes f\mathbf{f} and atomic positions r\mathbf{r} in the unit cell. From this definition, the unit cell and its periodic images is compactly represented as a multigraph; only a finite number of nodes describing the atoms in the unit cell is required, and the number of edges joining the two nodes grows linearly with the number of periodic images of the atoms in the structure of interest (see Figure 1).

A crystal graph describing the whole crystal structure will require infinitely many edges information. In practice, one typically embeds only partial information of the structure using the description of the unit cell and a finite number of edges encoding interactions with neighbor atoms [37]. One method to construct the edges is to connect a target atom to other atoms within a finite radius; however, still, this method generates an excessive number of edges, which is computationally infeasible for GNNs machine learning. Alternatively, some algorithms determine the connection between atoms using, e.g., the Voronoi tessellation, generating a moderate number of edges [38, 39]. However, for layered materials, using Voronoi tessellation causes the edges connecting atoms between layers to be absent. These inter-layer connections are crucial in differentiating the structure strained in out-of-plane direction from the structures strained in other directions. In this work, we define an edge between a pair of atoms Δr\Delta\mathbf{r} (dropping the translation vector superscript and the atom indices subscript for brevity) to be non-zero only if each component Δrβ\Delta r_{\beta} where β∈{x,y,z}\beta\in\{x,y,z\} satisfies the following spatial cutoff criterion: ∣Δrβ∣≤min⁡(∣aβ∣+∣bβ∣+∣cβ∣,s)+δ|\Delta r_{\beta}|\leq\min(|a_{\beta}|+|b_{\beta}|+|c_{\beta}|,s)+\delta where aβa_{\beta}, bβb_{\beta}, and cβc_{\beta} are the β\beta components of the corresponding lattice parameters constituting the lattice matrix L\mathbf{L}, ss is a cutoff distance, and δ\delta is a small cutoff extension of our choice. This can reduce the number of edges from the hard radial cutoff criterion and ensures the connections of atoms between layers for layered materials by an appropriate choice of ss and δ\delta.

Crystal systems# of training indicesTraining indices
cubic611, 12, 14, 15, 44, 45
hexagonal1311, 12, 13, 14, 15, 16, 33, 34, 36, 44, 45, 46, 66
tetragonal1411, 12, 13, 14, 15, 16, 26, 33, 34, 36, 44, 45, 46, 66
trigonal1611, 12, 13, 14, 15, 16, 24, 25, 33, 34, 36, 44, 45, 46, 66
orthorhombic21all
monoclinic21all
triclinic21all

Table 1. The minimal set of ijij strained component of each crystal system used to generate an initial crystal graph input

Rotational and translational equivariance

If two crystal graph inputs constructed from an identical crystal graph strained by two different sets of strained components (ϵi,ϵj)(\epsilon_i, \epsilon_j) and (ϵk,ϵl)(\epsilon_k, \epsilon_l) yielding the new vertices V′={(fn,rn′)}\mathcal{V}^{\prime} = \{(f_n, \mathbf{r}^{\prime}_n)\} and V′′={(fn,rn′′)}\mathcal{V}^{\prime\prime} = \{(f_n, \mathbf{r}^{\prime\prime}_n)\}, respectively, are equivalent up to a 3D rotation and a translation, then we expect our machine learning model to predict the same SED. This symmetry-aware property can be implemented with geometric deep learning models [40]. We will equip a latent representation of our model with this symmetry-aware (equivariant) property, which is defined as follows.

A latent representation ϕ:V→Y\phi: \mathcal{V} \to\mathcal{Y} is equivariant if for each transformation Tg:V→VT_g : \mathcal{V} \to\mathcal{V} with gg being an element of an abstract group GG, there exists a transformation Sg:Y→YS_g : \mathcal{Y} \to\mathcal{Y} such that the following condition holds:

Sg[ϕ(v)]=ϕ(Tg[v]),S_g[\phi(v)] = \phi\left(T_g[v]\right),

for all g∈Gg \in G, v∈Vv \in\mathcal{V}. In our case, the group GG of interest is SE(3)\mathrm{SE}(3), providing the latent representation with a 3D rotation and translation equivariant property. This latent feature ϕ\phi can be achieved with SE(3)\mathrm{SE}(3)-equivariant GNNs, known as Tensor Field Network (TFN) [41], and with a TFN with an appropriate attention mechanism, known as SE(3)\mathrm{SE}(3)-transformers [42], see more detailed recipes of these two models in Appendix A.

The key idea that enables SE(3)\mathrm{SE}(3)-equivariant property of these two GNNs is that its message passing kernel is constructed from translational invariant spherical harmonic bases, with the structure-preserving transformation SgS_g given by the Wigner-D matrices (see Appendix A). Under a 3D rotation, each spherical harmonic JJ basis function of the kernel transforms according to

YJ(Rg−1Δr∥Δr∥)=DJ∗(g)YJ(Δr∥Δr∥),Y_J\left(\mathbf{R}_g^{-1}\frac{\Delta\mathbf{r}}{\lVert\Delta\mathbf{r}\rVert}\right) = D_J^*(g)Y_J\left(\frac{\Delta\mathbf{r}}{\lVert\Delta\mathbf{r}\rVert}\right),

where DJ(g)\mathbf{D}_J(g) is the JthJ^{\mathrm{th}} Wigner-D matrix, and Rg\mathbf{R}_g is a rotation matrix associated with g∈SO(3)g \in\mathrm{SO}(3), making

the learned latent representation ϕ\phi equivariant [41]. The multi-head attention mechanism in SE(3)\mathrm{SE}(3)-transformers also induces an equivariant property on the product rule between each key and value pairs, rendering the whole message passing algorithm equivariant under SE(3)\mathrm{SE}(3) [42]. In this work, we use SE(3)\mathrm{SE}(3)-transformers to first build a compressed representation of the orientations of a strained crystal structure, then such learned latent representation will be passed to other networks with a high expressive power to perform the prediction (regression) task of the SED. The following section contains the entire framework for the SED prediction.

Model architecture and training procedure

We now describe our GNNs-based model that predicts the SED tensor. The architecture of our model, which we term StrainTensorNet, is illustrated in Figure 2. The model is designed to receive two types of input: a crystal graph representing a strained crystal structure, and a one-hot vector of dimension 6(6+1)/2=216(6 + 1)/2 = 21 [43] indicating the strained components, with a value 1 in the ijij component that the strain operation is applied on and with a value 0 otherwise.

StrainTensorNet architecture

Figure 2. With a strained crystal graph and the strained component one-hot vector as the input, StrainTensorNet employs a self-supervised approach combining both classification and regression networks to predict both the degeneracy class vector and the UijU_{ij}. The fin,n0≡fnf_{\mathrm{in},n}^{0} \equiv f_n denotes the input feature of the nthn^{\mathrm{th}} node, where a superscript, 0, indicates the input into the l=0l=0 channel of the spherical harmonics in SE(3)-Transformer, which is a channel whose feature is invariant under rotation (transforms like a scalar quantity). Δrmn′\Delta\mathbf{r}_{mn}^{\prime} is the edge connecting a target node mm and a neighbor node nn. The degeneracy class vector from the classification network will be concatenated with the latent representation of the strained crystal graph (freg,n0f_{\mathrm{reg},n}^{0}) for the final regression network to predict the SED.

The crystal graph of a strained crystal structure is an input feature of the SE(3)-equivariant GNNs building block, which generates two latent features. The first latent feature (fclass,n0f_{\mathrm{class},n}^{0} in Figure 2) is fed into a classification neural network that predicts a vector of dimension 21, identifying all the upper-triangular elements of the SED tensor whose values are exactly identical (by symmetry) to the component indicated by the input one-hot vector. This inherent classification informs the model to discern the degenerate structure of the SED tensor that depends on the symmetry of an input crystal graph (see Figure 3). For example, for a cubic lattice strained in the direction 11, the input one-hot vector (1,0,…,0)T(1,0,\ldots,0)^{T} together with the latent feature fclass,n0f_{\mathrm{class},n}^{0} shall give a prediction of a vector that has the value close to 1 in the indices ij=11,22,33ij=11,22,33, and the value close to 0 in the other 18 indices. Finally, the predicted degeneracy class vector (DCV) together with the second latent feature (freg,n0f_{\mathrm{reg},n}^{0} in Figure 2) will be fed into the final neural network that predicts (regresses on) the SED.

Distance metrics for cubic, hexagonal, and monoclinic materials

Figure 3. The distance metrics d(Uij,Ukl)d(U_{ij}, U_{kl}) computed from the test set of the ground-truth SED (top row) and the predicted SED (bottom row). The columns organize the crystal structure, appearing from left to right is for cubic, hexagonal, and monoclinic materials respectively. The model prediction shows an excellent agreement with the ground-truth for crystals with high symmetry (cubic), and a good agreement with the ground truths for crystals with lower symmetry (hexagonal and monoclinic).

To generate an SE(3)-equivariant latent representation of an input crystal graph, as alluded to in the previous section, we first used the SE(3)-Transformer that receives the strained crystal graph input. The material structure from the database is first transformed according to Eq. (6) by a strain of a fixed magnitude, which is then converted into the strained crystal graph input. Noting that in the ab initio calculation, the atomic coordinates in the strained lattice are optimized to be in the equilibrium positions under an applied strain, but in this work, the fractional coordinates are kept as in pre-strain condition. The strained crystal graph input is given by

V′={(fn,rn′)∣fn∈RM, rn′=εILrn∈R3},E′={Δrmn′(T′)∣Δrmn′(T′)=rm′−rn′+T′; rm′,rn′∈R3}.\mathcal{V}^{\prime}=\{(f_n,\mathbf{r}_n^{\prime})\mid f_n\in\mathbb{R}^{M},\ \mathbf{r}_n^{\prime}=\varepsilon_I L\mathbf{r}_n\in\mathbb{R}^{3}\}, \qquad \mathcal{E}^{\prime}=\{\Delta\mathbf{r}_{mn}^{\prime(T^{\prime})}\mid\Delta\mathbf{r}_{mn}^{\prime(T^{\prime})}=\mathbf{r}_m^{\prime}-\mathbf{r}_n^{\prime}+\mathbf{T}^{\prime};\ \mathbf{r}_m^{\prime},\mathbf{r}_n^{\prime}\in\mathbb{R}^{3}\}.

where the prime notation indicates that the atomic positions and the translation vector are of the strained lattice. The training data of the {Uij}\{U_{ij}\} is computed, via Eq. (2), from {Cij}\{C_{ij}\} extracted from Materials Project database [24] (see Appendix § 1.2.5). For molecules and materials, if the number of training data is large enough, only atomic numbers are sufficient for the node (atom) attributes [44]. Nevertheless, in this work, the features such as atomic mass, electronegativity, polarizability, atomic radius, atomic number, and ijij indices, are used as node attributes. Polarizability data is obtained from [45], whereas other attributes are obtained from Materials Project database. The atomic number is vectorized through an embedding layer of dimension 512, which is then concatenated with the other four node attributes (see Figure 2). Note that the elastic constants are influenced by the atomic mass since the phonon frequencies in sound waves, particularly the acoustic modes, are inversely proportional to the square root of atomic mass. Additionally, the resistance to changes in bond length correlates with both electronegativity and polarizability. These three atomic properties are thus incorporated into node attributes, and they significantly improve the model prediction accuracy.

Since the atomic features are scalar attributes, the input feature vector will be fed into the l=0l=0 channel of the transformer network since such channel transforms like a scalar (see Appendix A). We’ll denote each node’s input feature vector as fnin=fn\mathbf{f}^{\mathrm{in}}_n = \mathbf{f}_n. For computational feasibility, we will restrict the output feature vector of the SE(3)-transformers to consist of 4 channels fhid,nl\mathbf{f}^{l}_{\mathrm{hid},n} with l∈{0,1,2,3}l \in\{0,1,2,3\}. These latent features are fed into two different SE(3)-equivariant GNNs without attention (TFNs), outputing a node-wise classification feature vector fclass,n0\mathbf{f}^{0}_{\mathrm{class},n}, and the node-wise regression feature vector freg,n0\mathbf{f}^{0}_{\mathrm{reg},n}, see Figure 2. Note that the attention mechanism is not deployed in the last GNN layer as it yields a better prediction accuracy.

To train the model to discern different orientations of a strained crystal graph and classify its appropriate degeneracy class vector, in each epoch, we draw a new realization of a crystal graph in a different orientation and of a new one-hot vector in the same class of vector associated with the original strained component. Specifically, a new crystal graph input is sampled from a uniformly random orientation (uniformly random Euler angles) around the center of mass in the unit cell of the original strained crystal graph. A new one-hot vector of the strained components is chosen such that the element 1 is drawn uniformly from one of the non-zero components of the DCV. For example, if the original strained component is 11 for a cubic material, a new one-hot vector in each epoch is drawn from the situations where only one of the strained components 11, 22, or 33 is 1, while the other components are 0.

Our self-supervised approach combines both the regression errors and the classification errors in the global loss function:

L=1N∑n=1N∣Uij(n)−Uij(n),pred∣+λLclass,(9)\mathcal{L} = \frac{1}{N}\sum_{n=1}^{N}\left|U_{ij}^{(n)}-U_{ij}^{(n),\mathrm{pred}}\right|+\lambda\mathcal{L}_{\mathrm{class}}, \tag*{(9)}

where the SED prediction Uij(n),predU_{ij}^{(n),\mathrm{pred}} is the output from the regression network, Lclass\mathcal{L}_{\mathrm{class}} is the binary cross-entropy loss function for multi-label classification, λ\lambda specifies the relative significance between the classification and the regression errors, and NN is the total number of training samples. We stop the training when the gradient of the loss is relatively small over multiple epochs. Training was performed on NVIDIA A100 GPUs.0

Choosing a subset of strained crystal structures for training data

By the virtue of rotational equivariance, we expect StrainTensorNet to efficiently predict the 21 components of UU using training data consisting of only non-degenerate components. We have selected the number of non-degenerate components for training that depends on crystal systems, i.e. cubic, tetragonal, hexagonal, trigonal, monoclinic, and triclinic. For cubic materials, only 6 components are used for training (the number of non-degenerate components for other crystal systems can be found in Table I). Recall that, as discussed earlier, the cubic lattice has 5 distinct UijU_{ij}, as shown in Eq. (5); however, the strain tensor transform the cubic structures into 6 distinct structures (according to the symmetry in Laue class m3ˉmm\bar{3}m) as follows

(T1T2T2O2M2M2T1T2M2O2M2T1M2M2O2O1M1M1O1M1O1)\begin{pmatrix} \mathcal{T}_1 & \mathcal{T}_2 & \mathcal{T}_2 & \mathcal{O}_2 & \mathcal{M}_2 & \mathcal{M}_2 \\ & \mathcal{T}_1 & \mathcal{T}_2 & \mathcal{M}_2 & \mathcal{O}_2 & \mathcal{M}_2 \\ & & \mathcal{T}_1 & \mathcal{M}_2 & \mathcal{M}_2 & \mathcal{O}_2 \\ & & & \mathcal{O}_1 & \mathcal{M}_1 & \mathcal{M}_1 \\ & & & & \mathcal{O}_1 & \mathcal{M}_1 \\ & & & & & \mathcal{O}_1 \end{pmatrix}

where each element in the matrix stands for a crystal system of L′(ϵi,ϵj)\mathrm{L}'(\epsilon_i,\epsilon_j), T\mathcal{T}, O\mathcal{O}, and M\mathcal{M} are tetragonal, orthorhombic, and monoclinic lattices, respectively, and a different number labeling the subscript indicates a different structure. Despite the fact that U14U_{14} (C14C_{14}) is equal to U15U_{15} (C15C_{15}) by the cubic symmetry, L′(ϵ1,ϵ4)\mathbf{L}'(\epsilon_1,\epsilon_4) and L′(ϵ1,ϵ5)\mathbf{L}'(\epsilon_1,\epsilon_5) possess different lattice symmetries which are orthorhombic and monoclinic lattices, respectively. This is because ϵ1\epsilon_1 strains the structure in the xx direction, while ϵ4\epsilon_4 and ϵ5\epsilon_5 strain the structure in both the yy and the zz directions, and both the xx and the yy directions, respectively. Since L′(ϵ1,ϵ4)\mathbf{L}'(\epsilon_1,\epsilon_4) and L′(ϵ1,ϵ5)\mathbf{L}'(\epsilon_1,\epsilon_5) are not equivalent up to a rotation, the SE(3) kernel regards the two inputs as different inputs. By exploiting SE(3) kernel to identify inputs that are equivalent up to a rotation, for cubic materials, it suffices to use the training data consisting of the distinct input structures, strained only in 6 directions, i.e., 11, 12, 14, 15, 44, and 45.

The distance between SED components

To evaluate our model capability to predict the symmetry-dependent degeneracy pattern of the SED tensor, we use the Canberra distance between two components of the SED tensor as a metric:

d(Uij,Ukl)≡1N∑n=1N∣Uij(n)−Ukl(n)∣∣Uij(n)∣+∣Ukl(n)∣,(10)d(U_{ij},U_{kl}) \equiv\frac{1}{N}\sum_{n=1}^{N}\frac{\left|U_{ij}^{(n)}-U_{kl}^{(n)}\right|}{\left|U_{ij}^{(n)}\right|+\left|U_{kl}^{(n)}\right|}, \tag*{(10)}

where NN represents the number of samples, and ijij and klkl are Voigt indices. With this metric, the distance between UijU_{ij} and UklU_{kl} is zero if they belong to the same degeneracy class (their values are identical.) The top row of Figure 3 shows the ground-truth distance pattern for cubic, hexagonal, and monoclinic materials. Dark blue and bright green colors indicate the d(Uij,Ukl)d(U_{ij}, U_{kl}) values that are closest to zero and 0.6, respectively. It is important to note that the degeneracy pattern of UijU_{ij} depends on the crystal symmetry, which results in a unique Canberra distance pattern for each crystal system. For instance, in cubic materials, U11U_{11}, U22U_{22}, and U33U_{33} are identical (in the same degeneracy class), so the distance pattern of cubic materials displays a dark blue box corresponding to indices 11, 22, and 33. The distance patterns in Figure 3 and Figure 4 are averaged over the samples with the same crystal system in the dataset. Hence, the distance pattern of low-symmetry crystals, such as monoclinic and triclinic, will strongly depend on the dataset.

Distance metric heatmaps for tetragonal, trigonal, orthorhombic, and triclinic materials

Figure 4. The distance metric (computed from the test set) of tetragonal, trigonal, orthorhombic, and triclinic materials (top to bottom respectively) for (left) the ground-truth and (right) StrainTensorNet’s prediction.

Discussion and Results

Rationale for the StrainTensorNet model architecture

The underlying principle for using an equivariant network as a core component is to construct a compact representation of the SED tensor, such that the crystal structures strained by different sets of ϵi\epsilon_i and ϵj\epsilon_j that are equivalent up to rotation (and translation) possess the same SED. Such symmetry-dependent SED tensor of specific crystal structures is demonstrated in the ground-truth distance patterns d(Uij,Ukl)d(U_{ij}, U_{kl}) of Figure 3 (top), where the StrainTensorNet can well approximate such symmetry-dependent patterns, see Figure 3 (bottom).

Figure G1 shows our earlier model development with the goal to compactly represent the SED tensor, beginning from the simplest but less expressive to StrainTensorNet which can compactly represent the SED tensor rather accurately. All these trial GNN-based models were trained on crystal structures with fewer than 500 edges. The simplest model (Figure G1(a)) takes as an input only the crystal graph with node attributes, i.e., atomic mass, atomic radius, electronegativity, polarizability, and embedded atomic number. This minimal equivariant model does not yield a sufficiently accurate representation of UijU_{ij}; we found that the model is more biased towards predicting the SED tensor whose distance matrix pattern is more akin to that of the cubic crystals. This could be because the model is not sufficiently expressive, and thus attempts to fit the majority of the training dataset comprising more higher-symmetry structures (see Table IV for the crystal system statistics of our curated dataset).

Crystal systemTraining set CountTraining set %Test set CountTest set %
Cubic386546.1744147.83
Hexagonal136916.3514615.84
Tetragonal158018.8716317.68
Trigonal5306.33606.51
Orthorhombic7318.73798.57
Monoclinic2533.02272.93
Triclinic430.5160.65

Table IV. The number of data for training and test sets categorized by crystal systems.

Earlier trial models with SE(3)-equivariant GNNs

Figure G1. Earlier trial models with SE(3)SE(3)-equivariant GNNs as a fundamental building block. (Top) The building block consists of a strained crystal graph input that is fed into SE(3)SE(3)-equivariant GNNs (SE(3)-transformer followed by a TFN, similar to that of the building block in StrainTensorNet), whose output is the global max-pooled feature from the TFN. (a) A minimal equivariant GNN model for SED regression. (b) An SED regression network that take both the known degeneracy class vector and the minimal equivariant GNN’s representation as an input. (c) The same network as (a) but the input node attributes also incorporates the degeneracy class vector.

To improve the minimal model’s expressiveness, we introduce the degeneracy class vector (DCV) as an additional input, to inform the model about the symmetry-dependent class of the SED tensor. The DCV is first embedded by fully-connected layers and then concatenated with global max-pooled features or node attributes, as shown in Figure G1(b) and (c), respectively. These symmetry-informed models yield lower MAE and RMSE of the SED tensor compared to that of the minimal model, while the model in Figure G1(b) performs slightly better than the model in Figure G1(c) (see Table VI). Moreover, the predicted distance patterns are less biased towards those of the cubic lattice. Specifically, for crystal structures with lower symmetry such as hexagonal, tetragonal, trigonal, and orthorhombic, the predicted distance patterns are relatively similar to their ground-truth distance patterns. As expected, informing the model about the DCV of the input helps force the distance between UijU_{ij} and UklU_{kl} belonging to the same degeneracy class to be closer to zero.

PropertiesAverageMAE of Fig. G1(a)MAE of Fig. G1(b)MAE of Fig. G1(c)
Uˉij\bar{U}_{ij} (meV/atom)2.7491.6280.8420.866
CijC_{ij} (GPa)44.1078.9013.2918.66
BVB_{V} (GPa)108.6767.9515.3517.07
GVG_{V} (GPa)53.9279.9412.4614.63
YY136.2686.0529.0831.93
ν\nu0.3941.7580.0440.061

Table VI. The MAE of predicted elastic properties using the model in Figure G1(a), (b), and (c). The dataset for these earlier models are the materials whose number of edges is less than 500.

While these models with a DCV included as an input give good predictions of the SED tensor, they require the knowledge of the prestrained crystal symmetry a priori (to properly assign the value of the DCV). To overcome this limitation, we use a self-supervised method schematically shown Figure 2 to modify the model to predict, rather than to require, the DCV. The classification neural network that predicts the DCV takes as an input a one-hot vector of the strained ijij components, representing the ϵi\epsilon_i and ϵj\epsilon_j components in (7). This self-supervised technique enables the StrainTensorNet to predict UijU_{ij} without any prior knowledge of the prestrained crystal symmetry. Importantly, StrainTensorNet can express the degeneracy class of the SED tensor relatively well, see Figure 3 and Figure 4.

StrainTensorNet’s prediction results

Table II summarizes StrainTensorNet’s prediction errors on the test set. To the best of our knowledge, StrainTensorNet provides the first data-driven prediction of UijU_{ij} with a reasonable accuracy. Using these predicted SED tensors together with Eq. (3), elastic tensors of materials can also be calculated. Our model’s prediction accuracy significantly depends on the elemental composition of the compounds. Figure 5 depicts the MAE of UijU_{ij} and CijC_{ij}, averaged over 21 ijij components and categorized by elements. It is interesting to note that the compounds containing period 2 elements (Be, B, C, N, O, and F), some period 6 transition metals (Ta, W, Re, and Os), and some actinides (Th, Pa, and Pu), which have high bulk and shear moduli on average, exhibit a higher MAE for UijU_{ij} compared to the overall MAE of the entire test set. As a result, their MAE for CijC_{ij} are also significantly higher than the overall MAE of the test set. In contrast, the compounds containing lanthanides, which exhibit medium bulk and shear moduli on average, except for Gd, demonstrate a considerably lower MAE than the overall MAE of the test set, for both UijU_{ij} and CijC_{ij}.

Bar charts displaying the MAE of predicted $U_{ij}$ and $C_{ij}$ components

Figure 5. The bar charts display the MAE of the predicted UijU_{ij} and CijC_{ij} components, averaged over 21 ijij components. The charts are categorized based on the elements present in the compounds in the test set and are colored according to the average bulk modulus for (a) and (c), and average shear modulus for (b) and (d), as indicated in the legend below. The dashed lines represent the MAE of each property, averaged over the data in the test set. It can be seen that most elements, whose compounds have high bulk and shear moduli, exhibit errors of UijU_{ij} and CijC_{ij} that exceed the dataset’s MAE.

PropertiesThis work AverageThis work MAEThis work RMSEMazhnik et al. [25] AverageMazhnik et al. [25] MAEMazhnik et al. [25] RMSEZhao et al. [26] RMSE
U‾ij\overline{U}_{ij} (meV/atom)2.6520.6551.288
CijC_{ij} (GPa)42.9210.3717.69
BVB_V (GPa)107.0711.4819.73111.8311.1119.5416.530
GVG_V (GPa)50.989.6117.1054.818.2411.4315.780
YY129.6222.2938.63138.9519.1526.23
ν\nu0.4010.0370.1330.2860.0410.105

Table II. Statistics of the dataset and the model prediction

To further investigate the prediction results of the elastic constants, we plot the comparisons between the prediction (obtained from the predicted SED tensors together with Eq. (3)) and the ground truth of the elastic constants in Figure 6. Figure 6(a) shows the prediction results of every CijC_{ij} components of all crystal structures, whereas Figure 6(b)–(Figure 6f) show the prediction results of C22C_{22}, C23C_{23}, C26C_{26}, C55C_{55}, and C56C_{56} of cubic materials. For cubic materials, the model is trained only on a non-degenerate subset of 21 UijU_{ij} components, specifically components 11, 12, 14, 15, 44, and 45. Notably, the model is able to predict other unseen components reasonably well. However, the prediction of the elastic constant components that are exactly zero is challenging. For example, the ground-truth values of C26C_{26} and C56C_{56} of cubic materials are exactly 0 GPa, while the means and standard deviations of these components from our model predictions are 0.69 and 4.45 GPa, and -0.43 and 4.85 GPa respectively (see Figure G3 for the histograms of the prediction results). Although the standard deviations are relatively large, the means are less than 1 GPa. It is worth noting that the first-principles calculations also make some errors in these values, but these are usually not computed since they are known to be exactly 0 GPa.

Predicted values versus ground-truth values of the elastic constants, with panels (a)–(f)

Figure 6. The predicted values (vertical axes) versus the ground-truth values (horizontal axes) of the elastic constants. (a) All the components of the elastic constants of the whole test set of all crystal systems. The results for only cubic materials are shown in (b) C22C_{22}, (c) C23C_{23}, (d) C26C_{26}, (e) C55C_{55}, and (f) C56C_{56}, where we note that the 22, 23, 26, 55, and 56 indices did not appear in the input of the training set for cubic materials.

Statistics of the prediction results for the elastic constant components concentrated near 0 GPa

Figure G3. Statistics of the prediction results (vertical axes) vs of the ground truths (horizontal axes) of the elastic constant components that are concentrated near 0 GPa. (a) The predicted C26C_{26} of a cubic system has the mean and the standard deviation of 0.69 and 4.45 GPa. (b) The predicted C56C_{56} of a cubic system has the the mean and the standard deviation of -0.43 and 4.85 GPa. (c) The predicted C26C_{26} of all the crystal systems has the mean and the standard deviation of 0.99 and 6.08 GPa. (d) The predicted C56C_{56} of all the crystal systems has the mean and the standard deviation of -0.20 and 9.22 GPa, respectively. The ground-truth distributions of C26C_{26} and C56C_{56} are centered around 0 GPa with vanishing widths. Although some of the predicted C26C_{26} and C56C_{56} are not exactly 0 GPa (for cubic system), most of the predicted values are near 0 GPa.

To see how well StrainTensorNet discerns the geometric relationship between different strained crystal graph inputs, we plot the Canberra distance metric of the predicted and ground-truth SED tensors for comparison in Figure 3. The predicted distance metric of cubic materials, as shown in the left column of Figure 3, closely resembles the ground-truth distance metric, although there are minor discrepancies in the values that belong to the same degeneracy class. For instance, the distances between the predicted U11U_{11}, U22U_{22}, and U33U_{33} are 4.8−6.24.8 - 6.2 μ\mueV/atom, resulting in the percentage errors of 1.83−2.43%1.83 - 2.43\% (computed from 1N∑n=1N∣Uij(n)−Ukl(n)∣/Uavg(n)×100%\frac{1}{N}\sum_{n=1}^{N}|U_{ij}^{(n)}-U_{kl}^{(n)}|/U_{\mathrm{avg}}^{(n)} \times100\% where Uavg(n)=(∣Uij(n)∣+∣Ukl(n)∣)/2U_{\mathrm{avg}}^{(n)}=(|U_{ij}^{(n)}|+|U_{kl}^{(n)}|)/2). For hexagonal materials, U11U_{11} are identical to U22U_{22}, but not to U33U_{33}. The predicted distance metric of hexagonal materials accordingly forms a dark blue box in the 11 and 22 components. The percentage error between the predicted U11U_{11} and U22U_{22} is 2.22%. For monoclinic materials, the values of each UijU_{ij} are not necessarily identical to one another, and the distance metric is averaged over monoclinic materials data. The resulting distance pattern may differ significantly depending on the dataset. Notably, these agreements between the predicted and the ground-truth distance metric in various crystal structures reveal that StrainTensorNet gives an excellent prediction of the degenerate structure of SED tensors of strained crystalline materials.

Lastly, the model can also predict other elastic properties, including Voigt’s bulk modulus (BVB_V), Voigt’s shear modulus (GVG_V), Young’s modulus (YY), and Poisson’s ratio (ν\nu), using the predicted CijC_{ij} together with Eqs. (19)–(22). Table II summarizes the MAE and RMSE of our predicted elastic properties compared to previous works. The MAE and RMSE of BVB_V are 12.59 and 20.83 GPa, respectively, which are comparable to those of Mazhnik et al. [25]. The MAE and RMSE of GVG_V are 10.49 and 18.95 GPa, respectively, with the RMSE of GVG_V being comparable to that reported by Zhao et al. [26]. Our work also yields smaller errors for ν\nu than Mazhnik et al. [25]. Note that our dataset (see Sec. E) differs from the dataset in the relevant work of [25, 26]. Figure G4 presents the MAE of BVB_V, GVG_V, YY and ν\nu, categorized by elements. The model produces high MAE for compounds containing period 2 elements, mainly due to their larger MAE in UijU_{ij} and CijC_{ij} (see Figure 5). On the other hand, the model can produce smaller MAE for lanthanide compounds, as their MAE in UijU_{ij} and CijC_{ij} (see Figure 5) are smaller.

Four bar charts showing MAE of predicted elastic properties categorized by elements

Figure G4. The bar charts display the MAE of the predicted elastic properties, i.e., (a) BVB_V, (b) GVG_V, (c) YY, and (d) ν\nu. The charts are categorized based on the elements present in the compounds in the test set and colored according to the average bulk modulus, as indicated in the legend below. The dashed lines represent the MAE of each property, averaged over the data in the test set. It can be seen that most elements, whose compounds have high bulk modulus, exhibit errors of BVB_V, GVG_V, and YY that exceed the dataset’s MAE.

Interpretability of StrainTensorNet’s latent features

To investigate how latent features facilitate successful prediction of the SED tensor, we employed the diffusion map to faithfully visualize the three-dimensional data representation of the high-dimensional latent features of the entire test set [46, 47]. Figure 7 and Figure 8 reveal the dimensionality reduction of the latent features in the diffusion coordinates (φ1\varphi_1, φ2\varphi_2, φ3\varphi_3). Two different latent features were considered: the global max-pooling layer from the GNN representation of the input crystal graph (denoted as A in Figure 2), and the concatenation between A and the embedded DCV (denoted as B in Figure 2). These low-dimensional representations in Figure 7 and Figure 8 are colored by the energy scale (left column) and by the strained component ijij (middle and right columns).

Three-dimensional diffusion map representation of the concatenated latent features

Figure 7. Three-dimensional diffusion map representation of the concatenated latent features that are fed into the SED tensor prediction layer (the latent feature B\mathbf{B} in Figure 2). (a) is colored by the SED scale, and (b) is colored by the strained component ijij. (c) highlights the latent features of 3 example materials colored by the strained component ijij as in (b); the features of other materials are shown in transparent grey. The 3 materials are trigonal LuTiTe2_2 phase (+), tetragonal YMnSi phase (△), and cubic Ta3_3Ru phase (□), which have low, medium, and high elastic constants, respectively. These are all the latent features in the test dataset. More elaborated discussion can be found in Sec. IV C.

Three-dimensional diffusion map representation of the global max-pooled latent features of the graph neural networks

Figure 8. Three-dimensional diffusion map representation of the global max-pooled latent features of the graph neural networks (i.e., the latent feature A\mathbf{A} in Figure 2). (a) is colored by the SED scale, and (b) is colored by the strained component ijij. (c) highlights the latent features of 3 example materials colored by the strained component ijij as in (b); the features of other materials are shown in transparent grey. The 3 materials are trigonal LuTiTe2_2 phase (+), tetragonal YMnSi phase (△), and cubic Ta3_3Ru phase (□), which have low, medium, and high elastic constants, respectively. See the discussion and the interpretation of these plots in Sec. IV C.

What does the max-pooled feature of the crystal graph input A represent? Figure 8(c) shows that the crystal graph latent representations of the same material that are strained in different ijij components are almost identical, in fact sharing greater resemblance than the representations of different materials that belong to the same symmetry class. This is consistent with the result of Bronstein et al. which shows that the latent representation in SE(3)-equivariant GNNs of the original input is very similar to those of the mildly spatially distorted inputs, and, interestingly, is less similar to those of inputs spatially translated further away [40]. Hence, the latent feature A seems to uniquely encode the information of the input material, rather than the symmetry of the input. Thus A alone does not suffice to meaningfully represent the symmetry-dependent SED tensor (see Figure 8(a) and (Figure 8b)).

On the other hand, with the embedded DCV in the latent feature B, the network can differentiate different strained components within the same material class. When strained in 21 different components, the latent features A of three example materials (trigonal LuTiTe2_2 phase, tetragonal YMnSi phase, and cubic Ta3_3Ru phase) that were not differentiable within the same material class (Figure 8(c)) are now clearly segregated in the φ2\varphi_2 variable through the help of the embedded DCV, see Figure 7(c). In fact, the latent feature B which combines both the input crystal graph representation and the embedded DCV offers an interpretable latent representation of the SED (Figure 7), which we now discuss.

The latent feature B is organized such that the (φ1\varphi_1, φ2\varphi_2) coordinate encodes the strain energy density that varies from a smaller value in the (+, +) quadrant to a larger value in the (−, −) quadrant (see Figure 7(a)). Since UijU_{ij} stored by the tensile strain (i,j∈{1,2,3}i,j \in\{1,2,3\}) is typically higher than UijU_{ij} stored by the shear strain (i,j∈{4,5,6}i,j \in\{4,5,6\}), Figure 7(a) and (Figure 7b) consistently reveal that the latent features of the materials strained by the 11, 22, or 33 component have a negative φ2\varphi_2 value, whereas those of the materials strained by the 44, 55, or 66 components have a positive φ2\varphi_2 value. In addition, since UijU_{ij} for i≠ji \ne j is computed from the sum of CiiC_{ii}, CjjC_{jj}, and CijC_{ij}, it is typically larger than UiiU_{ii} and UjjU_{jj}. Figure 7(a) and (Figure 7b) also consistently reveal that the latent features of the materials strained by 44, 45, 11, and 12 components are respectively organized in the φ2\varphi_2 coordinate from a more positive to a more negative value. Additionally, Figure 7(a) and (Figure 7c) show that materials with higher (lower) average SED will be represented by a larger negative (larger positive) φ1\varphi_1 variable. In summary, the max-pooled feature from the graph neural networks A encodes the material information, and, together with the information of the DCV, the concatenated latent feature B effectively encapsulates both the material information and its strained component-dependent SED.

Conclusions

We have demonstrated a novel data-driven framework for predicting the elastic properties of crystal structures using SE(3)-equivariant graph neural networks (GNNs). By leveraging SE(3)-equivariant GNNs as building blocks, our self-supervised deep learning model accurately predicts Voigt’s bulk modulus, Voigt’s shear modulus, Young’s modulus, and Poisson’s ratio that are comparable to other non-equivariant deep learning studies [25, 26].

A key contribution is the prediction of the strain energy density (SED) and its associated elastic constants, which are tensorial quantities that depend on a material’s crystallographic group. The similarity between the distance metrics of the SED components between the ground truths and the predictions demonstrates our model’s capability to identify different symmetry groups of strained crystal structures. Requiring only a strained crystal graph and the strained component as the input, our approach offers an efficient alternative to the standard ab initio method for designing new materials with tailored elastic properties.

The interpretability of the model is also a notable feature. The learned latent representations taking into account the degeneracy class vector are organized in a physically meaningful structure for predicting the SED tensors. This interpretability aspect enhances the transparency of model prediction, enabling the justification of whether the prediction is physically relevant.

The combination of interpretability and the consideration of crystallographic groups sets our model apart from recent data-driven methods for predicting elastic properties of materials. We hope this work is a stepping stone toward an efficient data-driven approach for materials discovery and design, and opens up avenues for approaching more challenging tensor prediction tasks, e.g., predicting dielectric and piezoelectric tensors, which are second-rank and third-rank tensors, respectively.

Acknowledgments

This research is supported by the Program Management Unit for Human Resources and Institutional Development, Research and Innovation (Grant No. B05F650024) and by Thailand Science research and Innovation Fund Chulalongkorn University (CU FRB65 ind (5) 110 23 40). The authors acknowledge high performance computing resources including NVIDIA A100 GPUs from Chula Intelligent and Complex Systems Lab, Faculty of Science, and from the Center for AI in Medicine (CU-AIM), Faculty of Medicine, Chulalongkorn University, Thailand. The authors also acknowledge the National Science and Technology Development Agency, National e-Science Infrastructure Consortium, Chulalongkorn University and the Chulalongkorn Academic Advancement into Its 2nd Century Project (Thailand) for providing computing infrastructure.

Ingredients of SE(3)-Transformers

Here we summarize the computational building block of SE(3)-equivariant GNNs with a built-in attention mechanism (SE(3)-Transformers). More detailed discussion of this deep learning architecture can be found in the original proposal [42].

In an SE(3)-equivariant GNN without attention, also known as Tensor Field Network (TFN) [41], given a multi-channel input feature of a node jthj^{\mathrm{th}}, fl,n,jkf_{l,n,j}^{k}, of a crystal graph, the GNN will update each crystal graph input feature by the following message-passing convolution

fout,il=Wllfin,il+∑k≥0∑j≠iWlk(rj−ri)fin,jk,(11)f_{\mathrm{out},i}^{l}=\mathbf{W}^{ll}f_{\mathrm{in},i}^{l}+\sum_{k\geq0}\sum_{j\ne i}\mathbf{W}^{lk}(\mathbf{r}_{j}-\mathbf{r}_{i})f_{\mathrm{in},j}^{k}, \tag*{(11)}

where ii and jj are respectively the indices of a target node and its neighbouring nodes, ll and kk specify type-ll and type-kk feature vectors which respectively transform under SO(3) rotation according to the (2l+1)×(2l+1)(2l+1)\times(2l+1) and (2k+1)×(2k+1)(2k+1)\times(2k+1) Wigner-D matrices (type-0 vector is invariant under SO(3) rotation), Wll\mathbf{W}^{ll} is the diagonal matrix whose diagonal is identically parametrized by a trainable parameter wllw^{ll} describing a self-interaction message passing, and Wlk\mathbf{W}^{lk} is a convolution kernel that propagates the channel-kk feature from neighboring nodes to the channel-ll feature of the node of interest such that

Wlk(rj−ri)=∑J=∣k−l∣k+lφJlk(∥rj−ri∥)WJlk(rj−ri),(12)\mathbf{W}^{lk}(\mathbf{r}_{j}-\mathbf{r}_{i})=\sum_{J=|k-l|}^{k+l}\varphi_{J}^{lk}(\lVert\mathbf{r}_{j}-\mathbf{r}_{i}\rVert)\mathbf{W}_{J}^{lk}(\mathbf{r}_{j}-\mathbf{r}_{i}), \tag*{(12)}

with

WJlk(rj−ri)=∑m=−JJYJm(rj−ri∥rj−ri∥)QJmlk,(13)\mathbf{W}_{J}^{lk}(\mathbf{r}_{j}-\mathbf{r}_{i})=\sum_{m=-J}^{J}Y_{Jm}\left(\frac{\mathbf{r}_{j}-\mathbf{r}_{i}}{\lVert\mathbf{r}_{j}-\mathbf{r}_{i}\rVert}\right)\mathbf{Q}_{Jm}^{lk}, \tag*{(13)}

where φJlk(∥r∥)\varphi_{J}^{lk}(\lVert\mathbf{r}\rVert) is a learnable radial function, YJmY_{Jm} is the mthm^{\mathrm{th}} component (mm magnetic quantum number in quantum mechanics) of the spherical harmonic function YJ:R3→R2J+1Y_{J}:\mathbb{R}^{3}\to\mathbb{R}^{2J+1}, QJmlk\mathbf{Q}_{Jm}^{lk} is a Clebsch-Gordan matrix with dimension (2l+1)×(2k+1)(2l+1)\times(2k+1). The SO(3)-equivariant basis kernel WJlk\mathbf{W}_{J}^{lk} imposes a fixed functional constraint on the convolution kernel along the angular direction, whereas the radial direction is still trainable via φJlk\varphi_{J}^{lk}.

The SE(3)-Transformer introduces the attention αij\alpha_{ij} to each edge of a TFN, so that the message passing rule of Eq. (A1) is weighted by the attentions. Without isolating out the self-interaction term, for brevity, this weighted message passing rule can be compactly expressed as

fout,il=∑jαijvijl,(14)f_{\mathrm{out},i}^{l}=\sum_{j}\alpha_{ij}v_{ij}^{l}, \tag*{(14)}

where the channel-ll value message, vijlv_{ij}^{l}, is the same as that of the above TFN, defined as

vijl=⨁l≥0∑k≥0WVlk(rj−ri)fin,jk,(15)v_{ij}^{l} = \bigoplus_{l \ge0} \sum_{k \ge0} W_{V}^{lk}(\mathbf{r}_{j}-\mathbf{r}_{i})f_{\mathrm{in},j}^{k}, \tag*{(15)}

and the attention on the edge ijij (which is invariant under SO(3) [42]) is

αij=exp⁡(qiTkij)∑j′exp⁡(qiTkij′),(16)\alpha_{ij} = \frac{\exp(\mathbf{q}_{i}^{\mathsf{T}}\mathbf{k}_{ij})}{\sum_{j'}\exp(\mathbf{q}_{i}^{\mathsf{T}}\mathbf{k}_{ij'})}, \tag*{(16)}

with a query vector on a node ii, qi\mathbf{q}_{i}, and a key vector on the edge ijij, kij\mathbf{k}_{ij}, given by, respectively,

qi=⨁l≥0∑k≥0WQlkfin,ik,(17)\mathbf{q}_{i} = \bigoplus_{l \ge0} \sum_{k \ge0} W_{Q}^{lk}f_{\mathrm{in},i}^{k}, \tag*{(17)}
kij=⨁l≥0∑k≥0WKlk(rj−ri)fin,jk.(18)\mathbf{k}_{ij} = \bigoplus_{l \ge0} \sum_{k \ge0} W_{K}^{lk}(\mathbf{r}_{j}-\mathbf{r}_{i})f_{\mathrm{in},j}^{k}. \tag*{(18)}

⨁l≥0\bigoplus_{l \ge0} denotes the direct sum over the ll channels. Figure A1 depicts the message passing with the attention mechanism of the SE(3)-Transformer model. Detailed numerical implementation of these models can be found in https://github.com/trachote/predict_elastic_tensor.

A schematic of a GNN message passing rule

Figure A1. A schematic of a GNN message passing rule that aggregates features from neighboring nodes of a crystal graph with an attention mechanism. A white square on each edge depicts an attention head. A target node (pink circle) sends a query vector of the node ii, (qi\mathbf{q}_i, pink arrow), to match with key vectors (kij\mathbf{k}_{ij}, dotted purple arrow) and the value message (vij\mathbf{v}_{ij}, purple arrow) from neighboring nodes (blue and orange circles). The angular momentum channels are not illustrated in the visualization.

Macroscopic elastic moduli

Here are the formulae of elastic moduli computable from the elastic constants.

Voigt’s bulk modulus:

BV=19∑i,j=13Cij.(19)B_V = \frac{1}{9}\sum_{i,j=1}^{3} C_{ij}. \tag*{(19)}

Voigt’s shear modulus:

GV=130∑i,j=13(Cii−Cij)+15∑i=46Cii.(20)G_V = \frac{1}{30}\sum_{i,j=1}^{3}(C_{ii}-C_{ij})+\frac{1}{5}\sum_{i=4}^{6}C_{ii}. \tag*{(20)}

Young’s modulus:

Y=9BG3B+G.(21)Y = \frac{9BG}{3B+G}. \tag*{(21)}

The Possion’s ratio:

ν=3B−2G6B+2G.(22)\nu= \frac{3B-2G}{6B+2G}. \tag*{(22)}

Symmetry in the strain energy tensor

The strain energy density (SED) tensor (U\mathbf{U}) is derived from the elastic tensor (C\mathbf{C}). The symmetry in U\mathbf{U} follows the symmetry in C\mathbf{C} which depends on the crystal system.

Hexagonal and Tetragonal I (for Laue class 4/mmm)

Chex/tet=(C11C12C13000C11C13000C33000C4400C440C66).(23)\mathbf{C}_{\mathrm{hex/tet}} = \begin{pmatrix} C_{11} & C_{12} & C_{13} & 0 & 0 & 0 \\ & C_{11} & C_{13} & 0 & 0 & 0 \\ & & C_{33} & 0 & 0 & 0 \\ & & & C_{44} & 0 & 0 \\ & & & & C_{44} & 0 \\ & & & & & C_{66} \end{pmatrix}. \tag*{(23)}
Uhex/tet=(U11U12U13U14U14U16U11U13U14U14U16U33U34U34U36U44U45U46U44U46U66).(24)\mathbf{U}_{\mathrm{hex/tet}} = \begin{pmatrix} U_{11} & U_{12} & U_{13} & U_{14} & U_{14} & U_{16} \\ & U_{11} & U_{13} & U_{14} & U_{14} & U_{16} \\ & & U_{33} & U_{34} & U_{34} & U_{36} \\ & & & U_{44} & U_{45} & U_{46} \\ & & & & U_{44} & U_{46} \\ & & & & & U_{66} \end{pmatrix}. \tag*{(24)}

Tetragonal II (for Laue class 4/m4/m)

CtetII=[C11C12C1300C16C11C1300−C16C33000C4400C440C66].(25)\mathbf{C}_{\mathrm{tetII}} = \begin{bmatrix} C_{11} & C_{12} & C_{13} & 0 & 0 & C_{16} \\ & C_{11} & C_{13} & 0 & 0 & -C_{16} \\ & & C_{33} & 0 & 0 & 0 \\ & & & C_{44} & 0 & 0 \\ & & & & C_{44} & 0 \\ & & & & & C_{66} \end{bmatrix}. \tag*{(25)}
UtetII=[U11U12U13U14U14U16U11U13U14U14′U16′U33U34U34U36U44U45U46U44U46U66],(26)\mathbf{U}_{\mathrm{tetII}} = \begin{bmatrix} U_{11} & U_{12} & U_{13} & U_{14} & U_{14} & U_{16} \\ & U_{11} & U_{13} & U_{14} & U_{14}' & U_{16}' \\ & & U_{33} & U_{34} & U_{34} & U_{36} \\ & & & U_{44} & U_{45} & U_{46} \\ & & & & U_{44} & U_{46} \\ & & & & & U_{66} \end{bmatrix}, \tag*{(26)}

where U16′=−C16ϵ1ϵ6+12(C11ϵ12+C66ϵ62)U_{16}'=-C_{16}\epsilon_1\epsilon_6+\frac{1}{2}(C_{11}\epsilon_1^2+C_{66}\epsilon_6^2).

Rhombohedral I (for Laue class 3ˉm\bar{3}m)

CrhomI=[C11C12C13C1400C11C13−C1400C33000C4400C44C14C66].(27)\mathbf{C}_{\mathrm{rhomI}} = \begin{bmatrix} C_{11} & C_{12} & C_{13} & C_{14} & 0 & 0 \\ & C_{11} & C_{13} & -C_{14} & 0 & 0 \\ & & C_{33} & 0 & 0 & 0 \\ & & & C_{44} & 0 & 0 \\ & & & & C_{44} & C_{14} \\ & & & & & C_{66} \end{bmatrix}. \tag*{(27)}
UrhomI=[U11U12U13U14U14′′U16U11U13U14′U14′′U16U33U34U34U36U44U45U46U44U46′U66].(28)\mathbf{U}_{\mathrm{rhomI}} = \begin{bmatrix} U_{11} & U_{12} & U_{13} & U_{14} & U_{14}'' & U_{16} \\ & U_{11} & U_{13} & U_{14}' & U_{14}'' & U_{16} \\ & & U_{33} & U_{34} & U_{34} & U_{36} \\ & & & U_{44} & U_{45} & U_{46} \\ & & & & U_{44} & U_{46}' \\ & & & & & U_{66} \end{bmatrix}. \tag*{(28)}

where U14′=−C14ϵ1ϵ4+12(C11ϵ12+C44ϵ42)U'_{14}=-C_{14}\epsilon_{1}\epsilon_{4}+\frac{1}{2}(C_{11}\epsilon_{1}^{2}+C_{44}\epsilon_{4}^{2}), U14′′=12(C11ϵ12+C44ϵ42)U''_{14}=\frac{1}{2}(C_{11}\epsilon_{1}^{2}+C_{44}\epsilon_{4}^{2}), U46′=C14ϵ4ϵ6+12(C44ϵ42+C66ϵ62)U'_{46}=C_{14}\epsilon_{4}\epsilon_{6}+\frac{1}{2}(C_{44}\epsilon_{4}^{2}+C_{66}\epsilon_{6}^{2}).

Rhombohedral II (for Laue class 3)

CrhomII=[C11C12C13C14C150C11C13−C14−C150C33000C440−C15C44C14C66].(29)\mathbf{C}_{\mathrm{rhomII}}= \begin{bmatrix} C_{11} & C_{12} & C_{13} & C_{14} & C_{15} & 0 \\ & C_{11} & C_{13} & -C_{14} & -C_{15} & 0 \\ & & C_{33} & 0 & 0 & 0 \\ & & & C_{44} & 0 & -C_{15} \\ & & & & C_{44} & C_{14} \\ & & & & & C_{66} \end{bmatrix}. \tag*{(29)}
UrhomII=[U11U12U13U14U15U16U11U13U14′U15′U16U33U34U34U36U44U45U46′′U44U46′U66],(30)\mathbf{U}_{\mathrm{rhomII}}= \begin{bmatrix} U_{11} & U_{12} & U_{13} & U_{14} & U_{15} & U_{16} \\ & U_{11} & U_{13} & U'_{14} & U'_{15} & U_{16} \\ & & U_{33} & U_{34} & U_{34} & U_{36} \\ & & & U_{44} & U_{45} & U''_{46} \\ & & & & U_{44} & U'_{46} \\ & & & & & U_{66} \end{bmatrix}, \tag*{(30)}

where U46′′=−C15ϵ4ϵ6+12(C44ϵ42+C66ϵ62)U''_{46}=-C_{15}\epsilon_{4}\epsilon_{6}+\frac{1}{2}(C_{44}\epsilon_{4}^{2}+C_{66}\epsilon_{6}^{2}).

Orthorhombic, Monoclinic, and Triclinic

Strain energy tensors of orthorhombic, monoclinic, and triclinic lattices do not possess any degenerate structure, and are parametrized by the most general tensor of the form (4) in the main text.

The minimal set of strained components for each crystal systems in a training data set

The equivariant property of the SE(3)-transformer should be able to identify strained crystal graph inputs that are equivalent up to a rotation, thus to reduce the number of training data, we require one independent ijij component of the SED tensor for each degeneracy class vector. The other strained components can be incorporated during the randomized update of the one-hot strained component vector within the same degeneracy class in each training epoch, see Sec. III C. The minimal set of ijij indices of each crystal system can be chosen from the degenerate structure of the SED tensor for each crystal system in the previous Appendix C. Our chosen set is presented in Table I.

Dataset

Material data is pooled from the Materials Project database [24]. There are 12,105 materials in the database that have the elastic tensor. However, only some materials are used in this work due to memory restriction. The memory usage increases with the number of edges, so the computationally feasible data is restricted to the materials with the maximum number of edges of 2,000 constructed by a cell-radii approach. There are 9,296 out of 12,105 materials from the database with this property (about 76.79%). Figure E2 shows that the selected materials have bulk and shear moduli distributing in the range of 0 − 600 GPa, which can be a good representative of the whole database. In addition, we have found that the validation errors do not decrease significantly even when the number of edges increases to 3,000. The statistics of crystal systems curated for training and test sets of these 9,296 materials is shown is Table IV. The dataset is partitioned into a training set and a validation set using the stratified method (in scikit-learn) with a random seed number 0.

Bulk modulus plotted against the number of edges, with hue indicating the shear modulus

Figure E2. The bulk modulus of materials is plotted against the number of edges created by cell radii method, with the hue indicating the shear modulus. The data below a hard cutoff of 2000 edges can already cover the range of values of bulk and shear moduli of the whole dataset.

Cell radiiCrystalNN
mean2115.16143.47
std4665.66129.67
min66
25%36858
50%1024112
75%1776224
max758642272

Table III. Statistics of number of edges for 12105 materials created by cell radii and CrystalNN methods.

Model parameters

The hyperparameters that significantly affect the model performance are the number of convolution layers, the number of ll-channels in the SE(3) kernel, and the number of hidden units of the learnable radial function in the kernel ϕ\phi. The number of convolution layers is 4, where the model performance only slightly improves with more layers. The learnable radial function ϕ\phi is a feedforward network, consisting of an input layer that is fed into a fully-connected hidden layer with a ReLU activation, which is forwarded to another fully-connected layer that outputs a hidden feature vector. We use the number of hidden units in the hidden layer of ϕ\phi to be 128 to avoid over-squashing, and the performance only marginally improves expanding beyond 128 hidden units. The dimension of the hidden feature vectors fhid,nl\mathbf{f}^{l}_{\mathrm{hid},n} is 32, whereas the dimension of freg,nl\mathbf{f}^{l}_{\mathrm{reg},n} and fclass,nl\mathbf{f}^{l}_{\mathrm{class},n} are 128. The dimension of the input atomic number embedding is 512. The maximum degree of ll is 4 where, of course, the model performance also improves if the degree is higher especially for low-symmetry crystal structures; however, increasing the maximum degree of ll beyond 4 becomes infeasible on our computational infrastructure, with other hyperparameters fixed. The number of transformer-head is 2. We use ADAM optimizer and cosine annealing method for a scheduler with the initial learning rate of 10−410^{-4}.

Trial models and additional results

References

PropertiesAverageMAERMSE
UijU_{ij} (meV/atom)2.6520.6611.258
CijC_{ij} (GPa)42.9210.9817.86
BVB_{V} (GPa)107.0711.5819.10
GVG_{V} (GPa)50.989.9116.31
YY129.6222.8137.24
ν\nu0.4010.0350.081

Table V. The MAE and RMSE of predicted elastic properties using the model in Figure G1(b). The dataset for this model consists of materials whose number of edges is less than 2000.

Distance matrices for cubic, hexagonal, and monoclinic materials and prediction results from three models

Figure G2. From top to bottom rows are, respectively, the distance matric (of the materials whose numbers of edges are less than 500) of the ground truths, of the model in Figure G1(a), of the model Figure G1(c), and of the model model in Figure G1(b). The last row suggests that the concatenation of the strained crystal graph latent representation and the degeneracy class vector significantly increases the expressiveness of the SED regression network, motivating the self-supervised network architecture in the main text.

References

  1. [1]M. Born and K. Huang, Dynamical Theory of Crystal Lattices, International series of monographs on physics (Clarendon Press, 1988).
  2. [2]F. Mouhat and F. m. c.-X. Coudert, Necessary and sufficient elastic stability conditions in various crystal systems, Phys. Rev. B 90, 224104 (2014).DOI
  3. [3]J. R. MACDONALD, Review of some experimental and analytical equations of state, Rev. Mod. Phys. 41, 316 (1969).DOI
  4. [4]V. V. Brazhkin, High-pressure synthesized materials: treasures and hints, High Pressure Research 27, 333 (2007).DOI
  5. [5]M. Yamaguchi, T. Yagi, T. Azuhata, T. Sota, K. Suzuki, S. Chichibu, and S. Nakamura, Brillouin scattering study of gallium nitride: elastic stiffness constants, Journal of Physics: Condensed Matter 9, 241 (1997).DOI
  6. [6]H. J. Reichmann, S. D. Jacobsen, S. J. Mackwell, and C. A. McCammon, Sound wave velocities and elastic constants for magnesiowüstite using gigahertz interferometry, Geophysical Research Letters 27, 799 (2000).DOI
  7. [7]D. M. Teter, Computational alchemy: The search for new superhard materials, MRS Bulletin 23, 22 (1998).DOI
  8. [8]X.-Q. Chen, H. Niu, D. Li, and Y. Li, Modeling hardness of polycrystalline materials and bulk metallic glasses, Intermetallics 19, 1275 (2011).DOI
  9. [9]Y. Tian, B. Xu, and Z. Zhao, Microscopic theory of hardness and design of novel superhard crystals, International Journal of Refractory Metals and Hard Materials 33, 93 (2012).DOI
  10. [10]O. H. Nielsen and R. M. Martin, Stresses in semiconductors: Ab initio calculations on si, ge, and gaas, Phys. Rev. B 32, 3792 (1985).pubmed.ncbi.nlm.nih.gov/9937529
  11. [11]P. Labéguerie, M. Harb, I. Baraille, and M. Rérat, Structural, electronic, elastic, and piezoelectric properties of α-quartz and mco4 (m = Al, ga, fe; x = P, as) isomorph compounds: A dft study, Phys. Rev. B 81, 045107 (2010).
  12. [12]Y. Wang, J. J. Wang, H. Zhang, V. R. Manga, S. L. Shang, L.-Q. Chen, and Z.-K. Liu, A first-principles approach to finite temperature elastic constants, Journal of Physics: Condensed Matter 22, 225404 (2010).DOI
  13. [13]A. Ektarawong, S. I. Simak, and B. Alling, Carbon-rich icosahedral boron carbides beyond b4c and their thermodynamic stabilities at high temperature and pressure from first principles, Phys. Rev. B 94, 054104 (2016).
  14. [14]R. Stadler, W. Wolf, R. Podloucky, G. Kresse, J. Furthmüller, and J. Hafner, Ab initio calculations of the cohesive, elastic, and dynamical properties of cosi2 by pseudopotential and all-electron techniques, Phys. Rev. B 54, 1729 (1996).pubmed.ncbi.nlm.nih.gov/9986018
  15. [15]Y. Le Page and P. Saxe, Symmetry-general least-squares extraction of elastic data for strained materials from ab initio calculations of stress, Phys. Rev. B 65, 104104 (2002).
  16. [16]K. Sarasamak, S. Limpijumnong, and W. R. L. Lambrecht, Pressure-dependent elastic constants and sound velocities of wurtzite sic, gan, inn, zno, and cdse, and their relation to the high-pressure phase transition: A first-principles study, Phys. Rev. B 82, 035201 (2010).
  17. [17]L. A. Chernozatonskii, P. B. Sorokin, A. A. Kuzubov, B. P. Sorokin, A. G. Kvashnin, D. G. Kvashnin, P. V. Avramov, and B. I. Yakobson, Influence of size effect on the electronic and elastic properties of diamond films with nanometer thickness, The Journal of Physical Chemistry C 115, 132 (2011), https://doi.org/10.1021/jp1080687.
  18. [18]T. Pakornchote, A. Ektarawong, U. Pinsook, and T. Bovornratanaraks, Modifying electronic and elastic properties of 2-dimensional [110] diamond by nitrogen substitution, C 7, 10.3390/c7010008 (2021).DOI
  19. [19]M. A. Caro, S. Schulz, and E. P. O’Reilly, Comparison of stress and total energy methods for calculation of elastic properties of semiconductors, Journal of Physics: Condensed Matter 25, 025803 (2012).DOI
  20. [20]G. Kresse and J. Furthmüller, Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set, Computational Materials Science 6, 15 (1996).DOI
  21. [21]P. Giannozzi, S. Baroni, N. Bonini, M. Calandra, R. Car, C. Cavazzoni, D. Ceresoli, G. L. Chiarotti, M. Cococcioni, I. Dabo, A. D. Corso, S. de Gironcoli, S. Fabris, G. Fratesi, R. Gebauer, U. Gerstmann, C. Gougoussis, A. Kokalj, M. Lazzeri, L. Martin-Samos, N. Marzari, F. Mauri, R. Mazzarello, S. Paolini, A. Pasquarello, L. Paulatto, C. Sbraccia, S. Scandolo, G. Sclauzero, A. P. Seitsonen, A. Smogunov, P. Umari, and R. M. Wentzcovitch, Quantum espresso: a modular and open-source software project for quantum simulations of materials, Journal of Physics: Condensed Matter 21, 395502 (2009).DOI
  22. [22]S. J. Clark, M. D. Segall, C. J. Pickard, P. J. Hasnip, M. I. J. Probert, K. Refson, and M. C. Payne, First prin- ciples methods using castep, Zeitschrift für Kristallographie - Crystalline Materials 220, 567 (2005).
  23. [23]M. de Jong, W. Chen, T. Angsten, A. Jain, R. Notestine, A. Gamst, M. Sluiter, C. Krishna Ande, S. van der Zwaag, J. J. Plata, C. Toher, S. Curtarolo, G. Ceder, K. A. Persson, and M. Asta, Charting the complete elastic properties of inorganic crystalline compounds, Scientific Data 2, 150009 (2015).
  24. [24]A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K. a. Persson, The Materials Project: A materials genome approach to accelerating materials innovation, APL Materials 1, 011002 (2013).DOI
  25. [25]E. Mazhnik and A. R. Oganov, Application of machine learning methods for predicting new superhard materials, Journal of Applied Physics 128, 075102 (2020).DOI
  26. [26]Y. Zhao, K. Yuan, Y. Liu, S.-Y. Louis, M. Hu, and J. Hu, Predicting elastic properties of materials from electronic charge density using 3d deep convolutional neural networks, The Journal of Physical Chemistry C 124, 17262 (2020), https://doi.org/10.1021/acs.jpcc.0c02348.
  27. [27]H. Levämäki, F. Tasnádi, D. G. Sangiovanni, L. J. S. Johnson, R. Armiento, and I. A. Abrikosov, Predicting elastic properties of hard-coating alloys using ab-initio and machine learning methods, npj Computational Materials 8, 17 (2022).
  28. [28]N. Linton and D. S. Aidhy, A machine learning framework for elastic constants predictions in multi-principal element alloys, APL Machine Learning 1, 101063/5.0129928 (2023), 016109, https://pubs.aip.org/aip/aml/article-pdf/doi/10.1063/5.0129928/16772730/016109_1-online.pdf.DOI
  29. [29]P. Avery, X. Wang, C. Oses, E. Gossett, D. M. Pros- erpio, C. Toher, S. Curtarolo, and E. Zurek, Predicting superhard materials via a machine learning informed evolutionary structure search, npj Computational Materials 5, 89 (2019).arxiv.org/abs/1906.05886
  30. [30]W.-C. Chen, J. N. Schmidt, D. Yan, Y. K. Vohra, and C.-C. Chen, Machine learning and evolutionary prediction of superhard b-c-n compounds, npj Computational Materials 7, 114 (2021).
  31. [31]S. Kearnes, K. McCloskey, M. Berndl, V. Pande, and P. Riley, Molecular graph convolutions: moving beyond fingerprints, Journal of Computer-Aided Molecular Design 30, 595 (2016).DOI
  32. [32]J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, Neural message passing for quantum chemistry (2017).arxiv.org/abs/1704.01212
  33. [33]K. T. Schütt, H. E. Sauceda, P.-J. Kindermans, A. Tkatchenko, and K.-R. Müller, SchNet – a deep learning architecture for molecules and materials, The Journal of Chemical Physics 148, 241722 (2018), https://doi.org/10.1063/1.5019779.
  34. [34]N. Dym and H. Maron, On the universality of rotation equivariant point cloud networks, in International Conference on Learning Representations (2021).arxiv.org/abs/2010.02449
  35. [35]Tensor Field Network (TFN), L1Net, GemNet, and EGNN, have exploited this computational architecture to make better data-driven predictions on the properties of chemical compounds [41, 50–52]. The SE(3)-Transformers model based on TFN with an attention mechanism also demonstrates a good predictive performance on the standard QM9 dataset [42]. In addition, translation-rotation equivariance concept has also been incorporated into generative modeling, in an attempt to search for novel molecules and crystal structures with desirable properties [54, 55].
  36. [36]W. S. Slaughter, The Linearized Theory of Elasticity (Birkhäuser Boston, 2002).DOI
  37. [37]T. Xie and J. C. Grossman, Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties, Phys. Rev. Lett. 120, 145301 (2018).DOI
  38. [38]N. E. R. Zimmermann, M. K. Horton, A. Jain, and M. Haranczyk, Assessing local structure motifs using order parameters for motif recognition, interstitial identification, and diffusion path characterization, Frontiers in Materials 4, 10.3389/fmats.2017.00034 (2017).DOI
  39. [39]H. Pan, A. M. Ganose, M. Horton, M. Aykol, K. A. Persson, N. E. R. Zimmermann, and A. Jain, Benchmarking coordination number prediction algorithms on inorganic crystal structures, Inorganic Chemistry 60, 1590 (2021), pMID: 33417450.DOI
  40. [40]M. M. Bronstein, J. Bruna, T. Cohen, and P. Velickovic, Geometric deep learning: Grids, groups, graphs, geodesics, and gauges, CoRR abs/2104.13478 (2021), 2104.13478.arxiv.org/abs/2104.13478
  41. [41]N. Thomas, T. Smidt, S. Kearnes, L. Yang, L. Li, K. Kohlhoff, and P. Riley, Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds (2018).
  42. [42]F. Fuchs, D. Worrall, V. Fischer, and M. Welling, Se(3)-transformers: 3d roto-translation equivariant attention networks, in Advances in Neural Information Processing Systems, Vol. 33, edited by H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin (Curran Associates, Inc., 2020) pp. 1970–1981.arxiv.org/abs/2006.10503
  43. [43]Due to the symmetry of SED, we identify only an upper triangular element.
  44. [44]V. Fung, J. Zhang, E. Juarez, and B. G. Sumpter, Benchmarking graph neural networks for materials chemistry, npj Computational Materials 7, 84 (2021).DOI
  45. [45]P. Schwerdtfeger and J. K. Nagle, 2018 table of static dipole polarizabilities of the neutral elements in the periodic table, Molecular Physics 117, 1200 (2019).DOI
  46. [46]S. Lafon, Diffusion maps and geometric harmonics, Ph.D. thesis, Yale University (2004).
  47. [47]R. R. Coifman and S. Lafon, Diffusion maps, Appl. Comput. Harmon. Anal. 21, 5 (2006).DOI

Paper details

vlnc:W_y2q7xw8d
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents