Improving Molecular Modeling with Geometric GNNs: an Empirical Study

TL;DR

This paper empirically evaluates various design choices in Geometric Graph Neural Networks for 3D molecular systems, providing insights to optimize performance, scalability, and symmetry enforcement in molecular modeling.

Contribution

It systematically investigates the effects of canonicalization, graph creation, and auxiliary tasks on GNN performance for molecular modeling, guiding future research.

Findings

Canonicalization methods significantly affect model accuracy.

Graph creation strategies impact scalability and symmetry enforcement.

Auxiliary tasks can improve model robustness and generalization.

Abstract

Rapid advancements in machine learning (ML) are transforming materials science by significantly speeding up material property calculations. However, the proliferation of ML approaches has made it challenging for scientists to keep up with the most promising techniques. This paper presents an empirical study on Geometric Graph Neural Networks for 3D atomic systems, focusing on the impact of different (1) canonicalization methods, (2) graph creation strategies, and (3) auxiliary tasks, on performance, scalability and symmetry enforcement. Our findings and insights aim to guide researchers in selecting optimal modeling components for molecular modeling tasks.