Transforming Legacy Equation of State Databases into Interatomic Potentials via Machine Learning

To better understand and predict dynamic phenomena probed by Sandia’s Z Machine, we develop single element machine learned interatomic potentials (MLIPs) for metals. These studies leverage Density Functional Theory (DFT) simulations and generate on the order of 10⁶ configurations, many the result of prior EOS development. Turning this training data into MLIPs requires down sampling and sorting of the configurations. Within our MLIP code FitSNAP, we have developed capabilities to sort training data into groups to allow for finely tuned weighing of configurations, such that certain groups can be matched with higher fidelity while other groups can be included to improve transferability. We compare sorting DFT training data by density and temperature vs. automated binning using principal component analysis. We will also discuss tuning hyperparameters to allow for potentials that yield stable dynamics and match vapor dome points predicted by DFT. We will show MLIP comparisons to phase change behavior in DFT, predictions of the critical point, and demonstrations of exascale molecular dynamics simulations that can be achieved with such potentials. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.