Data-Driven Dynamics: Machine Learned Interatomic Potential for Simulating Materials Under Extreme Shock Conditions

Extended X-ray Absorption Fine Structure (EXAFS) experiments provide high-resolution, quantitative insights into the local atomic structure—even under extreme shock conditions. However, while the data is rich and high-fidelity, its inherent complexity and accompanying noise can make structural analysis challenging. In this talk, we introduce our novel approach for refining machine-learned interatomic potentials (MLIPs) for zinc under dynamic compression by leveraging experimental EXAFS data. MLIPs have transformed molecular dynamics simulations by systematically bridging the gap between electron-aware quantum mechanical models and large-scale, atomistic simulations. Our method employs the Hierarchically Interacting Particle Neural Network (HIPPYNN), which is initially trained on quantum mechanical data through an active learning framework (ALF) and then iteratively refined using both ambient and shock-compressed EXAFS spectra. We will demonstrate that the refined HIPNN MLIP model accurately captures zinc's structural responses over a broad range of pressures and temperatures and discuss the challenges of fitting multiple spectra simultaneously. Although the QM-trained MLIP initially struggled to predict zinc's targeted thermodynamic properties using MD simulation, our refinement procedure improves a diverse range of physical properties predicted by the model.