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

As the atomistic simulations of materials science move from traditional potentials to machine learning interatomic potential (MLIP), the field is entering the second phase, focused on discovering and explaining new material phenomena. 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. A global survey of 291 participants was conducted to understand the current standing (progress and bottleneck) of the machine learning-guided materials science research. The survey responses have been implemented to design an open-source software to reduce the access barriers of MLIP models for the global scientific community. 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 training. Finally, we compared 2 empirical and 2 state-of-the-art machine learning potentials on a dataset of 16000 structures (78–80 atoms each), showing the practicality of using MLIPs in molecular dynamics by analyzing computational performance with the number of unit cell repetitions (up to 2.5 million atoms) and CPUs (up to 256 processors).