Machine Learning Decoherence Time Formulas with Force-Projected Kinetic Energies for Nonadiabatic Scattering Dynamics

Abstract

The framework of mixed quantum-classical dynamics is promising for realizing efficient and reliable simulations of general nonadiabatic dynamics processes. In particular, the surface hopping method based on independent trajectories has attracted extensive interest over the past decades. In practical applications, however, its accuracy is often limited by the overcoherence problem. To address this limitation, we here utilize a machine learning approach to reveal the optimal decoherence time formula for high-dimensional systems and consider the kinetic energy projected along various force directions as the feature inputs. Remarkably, the obtained formula consistently distinguishes itself within the training set across four distinct descriptor spaces. The systematic benchmark confirms the high reliability of the formula based on the kinetic energy projected along the force direction of the nonactive potential energy surface. In fact, the vast majority of average population errors achieved by surface hopping with the new formula are below 0.01 and 0.02 in the investigated one- and two-dimensional systems, respectively. These results thus highlight the high performance of the new decoherence time formula in nonadiabatic scattering dynamics and demonstrate the feasibility of projecting the total kinetic energy onto a proper force direction to uncover the intricate decoherence effect in high-dimensional applications.