Machine learning interatomic potentials (MLIPs) have revolutionized molecular dynamics by combining deep learning efficiency with quantum mechanical accuracy. However, the “black-box” nature of neural networks often obscures the physical mechanisms by which these models fit complex potential energy surfaces (PES). This study investigates the correlation between local atomic environments and energy prediction behaviors using the pre-trained DPA2 model applied to NaCl-KCl and Na 2 WO 4 molten salt systems. Utilizing ab initio molecular dynamics (AIMD) data for fine-tuning, we demonstrate that the model spontaneously internalizes electrostatic laws despite the absence of explicit charge or oxidation state inputs. Statistical analysis reveals that for cations, higher energy contributions correlate with cation-dense regions (repulsion), while anion-dense regions lead to stabilization (attraction); the inverse is observed for anions. Furthermore, a strong correlation was found between model-predicted atomic energies and local formal charges calculated via Gaussian expansion. Our findings substantiate the physical interpretability of DPA2, showing that it transcends mathematical interpolation to extract the physical essence of interatomic interactions, providing critical validation for the development of advanced MLIPs. Can machine learning potentials capture electrostatic laws without explicit supervision? This study demonstrates that the pre-trained DPA2 model spontaneously internalizes Coulomb's Law from geometric topology in molten salts. By quantifying correlations between atomic energy contributions and ionic neighborhoods, we prove that deep learning transcends mathematical interpolation to extract fundamental physical essence with high fidelity.
Guo et al. (Tue,) studied this question.