Machine-learned interatomic potential (MLIP) has become a powerful tool to combine the accuracy of quantum mechanics with the efficiency of molecular dynamics in the era of artificial intelligence. However, a key open question persists: what physical mechanism is behind the atomic model that generates the MLIP and what physical information determines the final outputs? To address this problem, we use molten Na2WO4 as a representative system and fine-tune a pretrained deep potential model (DPA2) with ab initio molecular dynamics data of Na2WO4. We find a strong correlation between the model's final output and the projected density of states (PDOS) in energy regions exhibiting high electron density and distinct local atomic environments. This result indicates that a well-constructed neural network inherently captures the quantum-mechanical information and its predictions represent meaningful physicochemical interactions rather than purely statistical patterns. Importantly, the mechanistic insights gained in this work-which links model's outputs to electronic structure descriptors- are general in nature. It provides an electronic-structure-informed metric for feature learning and a general strategy for building interpretable, transferable MLIPs across diverse material systems.
Guo et al. (Tue,) studied this question.