Uncertainty-Aware Reactive Force Field Potentials for Materials via Physics-Informed Bayesian Optimization

Bridging density functional theory (DFT) fidelity with reactive molecular dynamics scales remains a fundamental computational material science challenge. Although reactive force field (ReaxFF) potentials use numerical bond-order terms, their calibration is an ill-posed, high-dimensional inverse problem. This study introduces uncertainty-aware ReaxFF potentials via a physics-informed Bayesian optimization (PIBO) framework. This framework embeds physical bounds, positivity constraints, and stability priors directly into the optimization loop, employing an adaptive Gaussian process surrogate with a staged acquisition schedule to ensure data-efficient convergence toward the DFT energetics. Applied to MoS2, PIBO produced physically consistent energy decompositions, including bond, off-diagonal, angle, torsion, van der Waals, and Coulomb terms, together with parameter posteriors quantifying epistemic uncertainty. The profile-likelihood posterior, which quantifies parameter identifiability from the training data, was reported separately from the optimizer replication variance, and the two were broadly consistent in magnitude, while relative posterior widths ranged from 4% for the constrained parameters to 32% for the weakly identified high-order terms. In comparison with several established optimizers on the training data energy scans, the PIBO potential reasonably reproduced the DFT bond, angle, and dihedral energy profiles. On a single-layer 1H-MoS2 system excluded from the training, PIBO-calibrated ReaxFF potentials predicted the equilibrium geometry, uniaxial and biaxial stress-strain responses, and a sulfur monovacancy formation energy for which the DFT reference fell within the nominal 95% PIBO ensemble uncertainty band. These results demonstrate that the PIBO framework provides a reproducible route to uncertainty-aware ReaxFF parametrization for extrapolation data sets.