Automated training of neural-network interatomic potentials
Abstract
Neural-network interatomic potentials (NNIPs) have transformed atomistic simulations, by enabling molecular dynamics simulations with near ab initio accuracy at reduced computational costs and improved scalability. Despite these advances, crafting NNIPs remains complex, demanding specialized expertise in both machine learning and electronic-structure calculations. Here, we introduce an automated, open-source, and user-friendly workflow that streamlines the creation of accurate NNIPs. Our approach integrates density-functional theory, data augmentation strategies and classical molecular dynamics to systematically explore the potential energy landscape. Our active-learning strategy leverages on-the-fly calibration of committee disagreement against true errors to ensure reliable uncertainty estimates. We use electronic-structure descriptors and dimensionality reduction to analyze the efficiency of our active learning strategy, which is shown to minimize both false positives and false negatives when deciding what to relabel with ab initio calculations. The method is validated on the fully automated training of a NNIP for a diverse set of carbon allotropes, reaching state-of-the-art accuracy and data efficiency. This platform democratizes NNIP development, empowering users to achieve high-precision simulations with minimal human intervention.