Bayesian Optimization of a Molecular Simulation of H₂O

A joint project with Leqian (Tim) Tan applying Bayesian optimization and uncertainty quantification to classical molecular models of water. We trained a Local Gaussian Process surrogate on the oxygen-oxygen radial distribution function from 512 LAMMPS simulations, then performed inference over the molecular model parameters using Hamiltonian Monte Carlo. The HMC-optimized parameters yielded improved liquid structure prediction compared to TIP4P/2005, with full posterior uncertainty quantification on the parameters and RDF predictions.