Complas 2025

COMPLAS 2025

AtomProNet: Data flow to and from machine learning interatomic potentials in materials science

Galib, Musanna (The University of British Columbia)
Isiet, Mewael (The University of British Columbia)
Ponga, Mauricio (The University of British Columbia)

In session: 701A - Machine Learning Methods for Inelastic/Multiscale/Multiphysics Material Modeling I

Please login to view abstract download link

Machine learning interatomic potentials (MLIPs) based on equivariant networks have been gaining potential in the materials science community. While MLIP development relies on curated data and flexible datasets from ab-initio simulations, transitioning seamlessly between ab-initio workflows and MLIP frameworks remains challenging. Here, we present AtomProNet, an open-source Python package that automates obtaining atomic structures, prepares and submits ab-initio jobs, and efficiently collects batch-processed data for streamlined neural network (NN) training. We show the ability of AtomProNet to facilitate the development of new potentials by developing an MLIP for alumina. We show that the new MLIP performs better than existing potentials. We further compare the results for high-strain rate deformation of alumina under hydrostatic tensile stress with experiments that are characteristic of spall failure in materials. Our work shows that AtomProNet is a useful tool to accelerate materials discovery using MLIP.