Prediction of Interatomic Potentials Combining Empirical Potential and Graph Neural Networks

Molecular dynamics (MD) simulation is constrained by the time and spatial scales as well as computational efficiency of traditional methods. While machine learning interatomic potentials (MLIPs) based on graph neural networks (GNNs) improve modeling accuracy, they suffer from insufficient short-range interaction modeling and lack of physical constraints. This paper proposes the DimezblNet model, which explicitly embeds the ZBL empirical potential function into the DimeNet framework. Through physically guided hybrid potential modeling, the description of short-range repulsive interactions is strengthened. On the MD17 dataset, the model reduces the mean absolute error (MAE) of energy and atomic force predictions for molecules like aspirin by 3.7%-5.3% compared to DimeNet. In the QM9 molecular property prediction task, the MAE of polarizability (α) is 0.0452 ų, outperforming DimeNet (0.0469 ų, a 3.6% improvement), and the MAE of dipole moment (μ) is 0.0273 D. Experiments show that combining physical priors with data-driven strategies significantly enhances the model's generalization ability and interpretability in complex molecular systems.