Efficient, Interpretable Graph Neural Network Representation for Angle-dependent Properties and its Application to Optical Spectroscopy

Abstract

Graph neural networks (GNNs) are attractive for learning properties of atomic structures thanks to their intuitive, physically informed graph encoding of atoms and bonds. However, conventional GNN encodings do not account for angular information, which is critical for describing complex atomic arrangements in disordered materials, interfaces, and molecular distortions. In this work, we extend the recently proposed ALIGNN encoding, which incorporates bond angles, to also include dihedral angles (ALIGNN-d), and we apply the model to capture the structures of aqua copper complexes for spectroscopy prediction. This simple extension is shown to lead to a memory-efficient graph representation capable of capturing the full geometric information of atomic structures. Specifically, the ALIGNN-d encoding is a sparse yet equally expressive representation compared to the dense, maximally-connected graph, in which all bonds are encoded. We also explore model interpretability based on ALIGNN-d by elucidating the relative contributions of individual structural components to the optical response of the copper complexes. Lastly, we briefly discuss future developments to validate the computational efficiency and to extend the interpretability of ALIGNN-d.

Connectivity Optimized Nested Graph Networks for Crystal Structures

“…This generalized definition would also allow for incorporating dihedral angles by adding edges e ij e kl to the line graph for every path (e ij , e jk , e kl ) of length three in the graph G [Hsu et al, 2021, Klicpera et al, 2020c. Another way to incorporate information of dihedral edges into the model would be to construct the second-order line graph L(L(G)) and add another layer of nesting to the model architecture.…”

Molecular graph transformer: stepping beyond ALIGNN into long-range interactions

“…There has been rapid progress in the development of GNN architectures for predicting material properties, such as such as SchNet, 10 Crystal Graph Convolutional Neural Network (CGCNN), 11 MatErials Graph Network (MEGNet), 12 Atomistic Line Graph Neural Network (ALIGNN) 13 and similar variants. [14][15][16][17][18][20][21][22][23][24][25] These models consider only the pairwise interactions between bonded atoms or between atoms within a cut-off radius of typically 6 Å to 8 Å. Some have also incorporated many-body relationships, such as bond-angles, into the molecular representation.…”

“…Common applications of GDL include shape analysis and pose recognition in computer vision, 1 link and community detection on social media networks, [2][3][4] representation learning on textual graphs, 5,6 medical image analysis for disease detection [7][8][9] and property prediction for molecular and crystalline materials. [10][11][12][13][14][15][16][17][18] In the eld of quantum chemistry, the development of Graph Neural Networks (GNN) has provided a means of computing the properties of molecules and solids, without the need to approximate the solution to the Schrödinger equation. Furthermore, compared to other Machine Learning (ML) techniques, they have shown immense potential in the eld of chemistry, since they do not require manual feature engineering and have signicantly better performance compared to other ML models.…”

Materials fatigue prediction using graph neural networks on microstructure representations

“…Another extension to graph neural network techniques attempted to encode structural properties such as molecular bond angles in an additional line graph 13,14 . Similarly, the crystallographic unit cell structure was encoded in graph representations to predict energy band gaps from multi-fidelity ab-initio simulations was proposed previously 15 .…”