Spherical Channels for Modeling Atomic Interactions

Abstract

Modeling the energy and forces of atomic systems is a fundamental problem in computational chemistry with the potential to help address many of the world's most pressing problems, including those related to energy scarcity and climate change. These calculations are traditionally performed using Density Functional Theory, which is computationally very expensive. Machine learning has the potential to dramatically improve the efficiency of these calculations from days or hours to seconds. We propose the Spherical Channel Network (SCN) to model atomic energies and forces. The SCN is a graph neural network where nodes represent atoms and edges their neighboring atoms. The atom embeddings are a set of spherical functions, called spherical channels, represented using spherical harmonics. We demonstrate, that by rotating the embeddings based on the 3D edge orientation, more information may be utilized while maintaining the rotational equivariance of the messages. While equivariance is a desirable property, we find that by relaxing this constraint in both message passing and aggregation, improved accuracy may be achieved. We demonstrate state-of-the-art results on the large-scale Open Catalyst 2020 dataset in both energy and force prediction for numerous tasks and metrics.Preprint. Under review.

DR-Label: Improving GNN Models for Catalysis Systems by Label Deconstruction and Reconstruction

“…The iterative relaxation method, such as (Zitnick et al, 2022;Klicpera et al, 2021;Sriram et al, 2022;Gasteiger et al, 2022), trains models as an imitation of DFT calculation. It predicts the transient energy and force of the atomic system, and iteratively updates atomic positions to the relaxed state.…”

“…(Jing et al, 2020) additionally added the vector channel besides the scalar representation for each node. More recently, (Zitnick et al, 2022) proposed to represent each node as a (L + 1) 2 × C matrix, which represents C channels of functions on the unit sphere, while each sphere function is parameterized by L degree of spherical harmonic basis.…”

Catalyst Energy Prediction with CatBERTa: Unveiling Feature Exploration Strategies through Large Language Models

“…Over successive development of GNNs for molecular property prediction has exhibited enhanced predictive accuracy for adsorption energies in the OC20 data set. Illustrated by examples such as GemNet-OC 16 and SCN, 17 leading-edge GNNs have achieved remarkable results in energy prediction. Notably, they have almost approached the DFT level accuracy by achieving mean absolute error (MAE) values as low as 0.28 eV.…”

Beyond Independent Error Assumptions in Large GNN Atomistic Models

“…For adsorption energy prediction, the mean absolute error (MAE) is used as the primary error metric to evaluate the performance of GNN models 22,23 . The MAE has decreased as newer GNNs have been developed, and their accuracy has approached that of DFT calculations [24][25][26] , which typically have an error range of 0.1 to 0.2 eV for adsorption energy prediction 27 .…”