Scalable training of neural network potentials for complex interfaces through data augmentation

Abstract

Artificial neural network (ANN) potentials enable accurate atomistic simulations of complex materials at unprecedented scales, but training them for potential energy surfaces (PES) of diverse chemical environments remains computationally intensive, especially when the PES gradients are trained on atomic force data. Here, we present an efficient methodology incorporating forces into ANN training by translating them to synthetic energy data using Gaussian process regression (GPR), leading to accurate PES models with fewer additional first-principles calculations and a reduced computational effort for training. We evaluated the method on hybrid density-functional theory data for ethylene carbonate (EC) molecules and their interfaces with Li metal, which are relevant for Li-metal batteries. The GPR-ANN potentials achieved an accuracy comparable to fully force-trained ANN potentials with a significantly reduced computational and memory overhead, establishing the method as a powerful and scalable framework for constructing high-fidelity ANN potentials for complex materials systems.