Interfacing experimental data with machine learned interatomic potentials

Construction of Machine Learned (ML) interatomic potentials from quantum mechanical energy and force data has become routine. To produce a new generation of validated interatomic potentials for simulating materials under extreme conditions, it is critical to incorporate experimental data. Experimental data can be used to validate potentials constructed from quantum mechanical data and may be incorporated into the training procedure itself. In this talk, we will present a recently developed YZn interatomic alloy potential constructed using our Active Learning Framework (ALF). ALF automates the process of constructing ML interatomic potentials using active learning, where ML uncertainties are used to augment training datasets in an optimal way. ALF is freely available and can be deployed on HPC resources such as Sierra and other GPU accelerated compute clusters. Dynamic properties of both pure Zn and YZn alloys will be computed and compared to EXAFS measurements obtained from the Dynamic Compression Center (DCS) at the Advance Photon Source (APS) on shock compressed Zn. Finally, future methods discussing possible ways to incorporate experimental data into the training procedure itself will be discussed.