Augmenting Reverse Monte Carlo with machininteratomic potentials and uncertainty quantificatione learning

Reverse Monte Carlo (RMC) is a technique to interpret experimental diffraction data by applying the Metropolis Monte Carlo algorithm to derive an atomistic model that fits the data. This optimization method is inherently stochastic and underconstrained, so it tends to yield highly disordered and distorted structures. There is no guarantee the result is physically sensible, so in practice it can be challenging to distinguish between unintuitive findings and artifacts of experimental limitations. These concerns can be mitigated by applying an interatomic potential (IAP) to constrain RMC to structures that are both physically sensible and consistent with experimental data. We present methods to improve the accessibility of IAPs, especially for experimentally-focused RMC users. We have updated the RMCProfile code to be compatible with most IAPs supported by LAMMPS, including models archived in OpenKIM. For situations where an IAP is not available for the materials system of interest, we propose a methodology to train machine learning IAPs on structures generated by RMC, leveraging its tendency to generate disordered and distorted structures as an advantage. Using Gaussian process machine learning IAPs with uncertainty quantification, we can accelerate the acquisition of diverse training data and assess whether the IAP is robust enough for use in RMC. These tools aim to improve the accessibility of analysis techniques that bridge theory and experiment.