Excited-state nonadiabatic dynamics in explicit solvent using machine learned interatomic potentials.

Area of Science:

• Computational Chemistry
• Photochemistry
• Machine Learning in Science

Background:

• Nonadiabatic simulations are crucial for understanding photoinduced processes in complex environments.
• Traditional quantum mechanics/molecular mechanics (QM/MM) methods are computationally expensive for excited-state dynamics.
• High computational cost limits the application of QM/MM in trajectory surface hopping methods.

Purpose of the Study:

• To develop a computationally efficient machine learning (ML) approach for nonadiabatic excited-state simulations.
• To replace traditional QM/MM electrostatic embedding with a machine-learned interatomic potential (ML/MM).
• To validate the accuracy and applicability of the ML/MM method for modeling photoinduced processes.

Main Methods:

• Utilized FieldSchNet, a machine-learned interatomic potential, to incorporate electric field effects into electronic states.
• Implemented an ML/MM approach as a surrogate for QM/MM electrostatic embedding in nonadiabatic excited-state trajectories.
• Applied the ML/MM method to simulate the excited-state dynamics of furan in water, considering five coupled singlet states.

Main Results:

• The ML/MM model successfully reproduced the electronic kinetics and structural rearrangements observed in QM/MM surface hopping simulations.
• Sufficiently curated training data is essential for the accuracy of the ML/MM model.
• Identified robust performance metrics for validating the accuracy and interpretability of the ML/MM model.

Conclusions:

• The developed ML/MM method offers a computationally feasible alternative to QM/MM for nonadiabatic excited-state simulations.
• This approach enables more extensive studies of photoinduced processes in explicit molecular environments.
• The validated ML/MM model provides reliable insights into the dynamics of excited electronic states.