Deep learning-based interatomic potentials for predicting thermodynamic properties of two-dimensional Ge/MoS2 systems

Abstract

This study presents the development and application of machine learning interatomic potentials based on neural networks for simulating two-dimensional materials, specifically monolayer Germanene, monolayer MoS₂, and the heterostructure Ge/MoS₂. Using first-principles molecular dynamics data, the ML models were trained via the DeePMD-kit framework to predict formation energies and atomic forces. The results demonstrate that machine learning can efficiently capture energy landscapes with high accuracy, even with limited training data. In contrast, predictions of moderately accurate atomic forces exhibit higher absolute errors, particularly affecting phonon calculations. Optimal neural network configurations were identified for each material system, enabling enhanced learning performance. This work confirms the potential of machine learning, particularly deep learning-based potentials, in accelerating thermodynamic modeling of 2D heterostructures with significantly reduced computational cost compared to traditional ab initio methods. Additionally, it underscores the need for more diverse datasets to enhance the reliability of force prediction in dynamical simulations.