Scaling a Universal Graph Neural Network Force Field to Billions of Atoms on Desk-Side GPUs

Machine-learning force fields are capable of simulating billions of atoms, but usually require high-performance supercomputing clusters. We show that a universal graph neural network force field, ALIGNN-FF, evaluates energies and forces of more than one billion atoms on desk-side computers such as the NVIDIA GB10, which has 128 GB of unified memory, by using a cube-and-halo decomposition. The halo contains every atom within the model's 20 Å receptive field, which makes the result exact. One GB10 holds 216,000 atoms with forces as a single graph. The cube-and-halo decomposition removes this limit, enabling an evaluation of 2 billion atoms on nine GB10s. We also demonstrate three case studies: a billion-atom polycrystal, a 102-million-atom microelectronic gate stack and a 1.07-million-atom satellite tobacco mosaic virus in water. This brings single-point evaluations at experimentally relevant sizes within reach of one research group. The code is available at: https://github.com/atomgptlab/scalingmat.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-11-16
DOI
https://doi.org/10.48550/arxiv.2611.20585
Primary Topic
Machine Learning in Materials Science
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Scaling a Universal Graph Neural Network Force Field to Billions of Atoms on Desk-Side GPUs

Kamal Choudhary, Jaehyung Lee
arXiv (Cornell University)
Machine Learning in Materials Science
preprint

Scaling a Universal Graph Neural Network Force Field to Billions of Atoms on Desk-Side GPUs

Kamal Choudhary, Jaehyung Lee
preprint en

Abstract

Machine-learning force fields are capable of simulating billions of atoms, but usually require high-performance supercomputing clusters. We show that a universal graph neural network force field, ALIGNN-FF, evaluates energies and forces of more than one billion atoms on desk-side computers such as the NVIDIA GB10, which has 128 GB of unified memory, by using a cube-and-halo decomposition. The halo contains every atom within the model's 20 Å receptive field, which makes the result exact. One GB10 holds 216,000 atoms with forces as a single graph. The cube-and-halo decomposition removes this limit, enabling an evaluation of 2 billion atoms on nine GB10s. We also demonstrate three case studies: a billion-atom polycrystal, a 102-million-atom microelectronic gate stack and a 1.07-million-atom satellite tobacco mosaic virus in water. This brings single-point evaluations at experimentally relevant sizes within reach of one research group. The code is available at: https://github.com/atomgptlab/scalingmat.

arXiv (Cornell University)
Machine Learning in Materials Science
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon