Predicting Molecular Photochemistry Using Machine-Learning-Enhanced Quantum Dynamics Simulations

Area of Science:

• Computational Chemistry
• Quantum Dynamics
• Photochemistry

Background:

• Light-driven molecular processes are crucial for technologies like photocatalysis, sunscreens, photovoltaics, and fluorescent probes.
• Accurate simulation of molecular photochemistry is a significant challenge due to the need for precise treatment of electronic structure, nuclear dynamics, and nonadiabatic couplings.
• Existing experimental techniques range from ultrafast transient absorption spectroscopy to advanced X-ray free-electron lasers.

Purpose of the Study:

• To describe the development of direct quantum dynamics methods for simulating complex molecular systems undergoing photochemical reactions.
• To present algorithmic advancements enabling on-the-fly potential energy surface generation and accurate quantum propagation.
• To showcase the application of these methods in modeling excited-state chemical dynamics.

Main Methods:

• Development of 'direct' quantum dynamics methods combining machine learning of potential energy surfaces (PESs) and nonadiabatic couplings with quantum propagation schemes like the multiconfiguration time-dependent Hartree (MCTDH) method.
• Implementation of active learning strategies for generating PESs during grid-based quantum chemical dynamics simulations.
• Development of novel diabatization schemes to facilitate direct grid-based simulations of photochemical dynamics.

Main Results:

• Enabled accurate on-the-fly simulations of molecular photochemistry by generating PESs in tandem with wave function propagation.
• Circumvented the need for extensive ab initio electronic structure data fitting by demanding energy evaluations only when required.
• Demonstrated the effectiveness of the developed methods through benchmark molecular simulations of systems with multiple nuclear degrees of freedom and coupled electronic states.

Conclusions:

• The developed direct quantum dynamics tools represent a significant advancement in modeling excited-state chemistry.
• These methods facilitate the study of complex photochemical processes such as photodissociation, proton/electron transfer, and ultrafast energy dissipation.
• The approach integrates computational methods with state-of-the-art experimental capabilities for a deeper understanding of light-driven molecular dynamics.