Deep Learning for Nonadiabatic Excited-State Dynamics.

Area of Science:

• Computational chemistry
• Quantum chemistry
• Machine learning in chemistry

Background:

• Exploring potential energy surfaces (PESs) of polyatomic molecules is crucial for understanding chemical reactions.
• Nonadiabatic dynamics, especially near conical intersections, present significant computational challenges.
• Accurate modeling of excited-state PESs is essential for photochemistry and photophysics.

Purpose of the Study:

• To demonstrate the efficacy of deep learning (DL) for modeling complex, nonlinear multistate potential energy surfaces.
• To accurately represent ground- and excited-state PESs of the CH2NH molecule using deep neural networks (DNNs).
• To investigate the ability of DL models to reproduce topological structures, photoisomerization paths, and conical intersections.

Main Methods:

• Utilized deep neural networks (DNNs) as accurate representations of CASSCF ground- and excited-state potential energy surfaces.
• Incorporated geometries near conical intersections into the training dataset for DNN models.
• Performed nonadiabatic dynamics simulations using the trained DNN models and compared results with ab initio methods.

Main Results:

• DNN models accurately reproduced excited-state topological structures and photoisomerization pathways for CH2NH.
• The DNN models successfully identified and characterized conical intersections.
• Nonadiabatic dynamics simulations using DNN models yielded results highly consistent with pure ab initio calculations.

Conclusions:

• Deep learning provides a powerful and accurate method for exploring complex excited-state potential energy surfaces.
• DNNs can effectively model nonadiabatic dynamics, including crucial features like conical intersections.
• This work encourages the broader application of machine learning in studying excited-state dynamics of polyatomic molecules.