Physical Prior Mean-Driven Bayesian Committee Molecular Dynamics (BCMD): From Born–Oppenheimer Dynamics to Curvature-Guided Non-Adiabatic Dynamics

Molecular dynamics (MD), when combined with high-accuracy quantum mechanics for energy and atomic force evaluations, is an indispensable tool extensively used to uncover in-depth atomistic details of chemical processes. However, the computational cost of first-principles predictions can be prohibitive. Here, we introduce a direct dynamics method, Bayesian Committee molecular dynamics (BCMD), which integrates Bayesian learning with a physically informed prior mean to efficiently construct potential energy surfaces (PESs) on the fly. This approach reduces quantum mechanical (QM) evaluations while maintaining high accuracy in nuclear motion propagation. The learned surrogate surfaces enhance the efficiency of the dynamics by guiding further atomic motions. We apply BCMD to both ground-state thermal chemistry and nonadiabatic photochemistry, demonstrating its versatility across different electronic structure regimes. Our approach is highly data-efficient and embeds chemical dynamics knowledge into the learning, enabling domain knowledge-informed, surrogate-based ab initio MD for mechanism exploration, photochemistry, and atmospheric chemistry.