(501a) Structure Optimized Potential Refinement: Pair Potentials from Experimental Scattering Data for Molecular Fluids

We have developed a convergent iterative algorithm that produces transferrable atomic pair potentials from experimentally determined pair distribution functions obtained from neutron scattering. The algorithm, Structure Optimized Potential Refinement (SOPR), was recently demonstrated to produce transferable pair potentials for monatomic fluids, predicting both fluid microstructure and ensemble thermodynamic properties. However, this initial procedure was only validated and benchmarked for noble gas liquids, significantly limiting its application to molecular systems. In this work, we expand the SOPR procedure and apply it to three molecular fluids, benzene, methane, and water. Briefly, the procedure iteratively compares simulated and experimental pair distribution functions and introduces an empirical potential to ensure convergence between the model and experiments. Molecular systems exhibit anisotropic and complex structural features and multiple atomic types, resulting in a substantially more challenging system to investigate. As pair distribution functions are co-related, convergence requires simultaneous tuning of all pair interactions, introducing numerical stability and over-fitting concerns. To overcome these challenges, the modified SOPR procedure for molecular systems introduced an inverse squared separation distance cutoff function with an acceptance factor to slowly update the empirical potential, avoid overcorrection, and mitigate fitting in local minima. Preliminary results demonstrate that the method can accurately represent the fluid micro-structure and predict thermodynamic properties (e.g., density, vapor-liquid equilibrium) with comparable or improved accuracy to conventional transferrable pair potentials (e.g., OPLS-AA, TraPPE-EH). The expansion of SOPR to molecular systems unlocks several new applications. First, the procedure can be used to directly include experimental scattering experiments into force field optimization. Additionally, the method produces model-independent pair potentials that could be utilized to probe necessary (or missing) physics from classical models in order to improve predictions.