报告题目：Nonadiabatic Dynamics and Machine Learning
Our works focus on the theoretical description of nonadiabatic dynamics on molecular excited states.
We introduced the supervised machine-learning (ML) approaches into the nonadiabatic dynamics simulation. The nonadiabatic Zhu-Nakamura TSH dynamics simulation was run using potential energy surfaces (PESs) based on kernel-ridge-regression technique. The TSH dynamics results using ML-PESs agree well with those based on pure ab-initio CASSCF PESs. This work provides the possibility to run a huge number of trajectories in nonadiabatic dynamics simulation of complex systems.
We show the possibility to analyze the geometrical evolution of trajectory-based nonadiabatic molecular dynamics by the unsupervised machine learning and big data analysis, particularly the dimensionality reduction techniques (Classical Multidimensional Scaling and Isometric Mapping). These approaches allow us to extract the major molecular motion from the very complicated time-dependent evolution from many trajectories without pre-knowledge of reaction pathway of excited state reactions. This opens a very interesting research topic in the future.
I will also discuss out recent work on the symmetrical quasiclassical dynamics in the basis of mapping Hamiltonian. After the transformation of Hamiltonian with many discrete electronic states to the mapping Hamiltonian with many coupled harmonic oscillators, it is possible to obtain a way to perform the semi-classical approximation. Our benchmark calculations show that the symmetrical quasiclassical dynamics based on this mapping Hamiltonian may give reasonable descriptions of the nonadiabatic dynamics of spin-boson systems.