When Three-Body Interactions Become Redundant: Scaling Graph Neural Network Potentials

Read full text Summarise with VIRAT
Authors: Linh La, Yugam Aggarwal, Truyen Tran, Svetha Venkatesh, Sherif Abdulkader Tawfik
Cited by: 0
References: 25

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Graph neural networks (GNN) are currently the vehicle for large scale molecular dynamics at quantum accuracy, but their practical scalability remains limited by the rapidly increasing size of graph representations required for large atomistic systems. In particular, the construction of higherorder interactions such as triplets leads to substantial memory and computational overhead during both training and inference. In this work, we investigate strategies to improve the scalability of GNNbased potentials while preserving predictive accuracy. We show that pruning triplet interactions during training and inference has minimal impact on regression accuracy for formation energies, band gaps, and elastic moduli. For molecular dynamics simulations, we find that pruning triplets during inference preserves structural fidelity while significantly reducing computational cost. This allows for upscaling of GNNs to large structures (beyond 30k atoms). To further extend scalability, we introduce a spatial chunking scheme that decomposes large atomistic systems into padded subgraphs, enabling efficient evaluation of forces for systems containing hundreds of thousands of atoms. Using this approach, we perform molecular dynamics simulations of a 400,000-atom LGPS supercell on a single GPU using the M3GNet and DIEP models. Our results demonstrate that fully trained manybody GNN potentials contain significant redundancy that can be exploited to improve scalability, providing a practical pathway toward large-scale atomistic simulations with universal machinelearning potentials.

Fields of Study & Subjects

Machine Learning in Materials Science Advanced Chemical Physics Studies Model Reduction and Neural Networks

Publication details

Journal / source
ChemRxiv
Publication date
25 March 2026
Language
EN
Document type
Preprint