A Descriptor Is All You Need: Accurate Machine Learning of Nonadiabatic Coupling Vectors.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Photochemistry

Background:

• Nonadiabatic couplings (NACs) are essential for modeling photochemical and photophysical processes.
• Current methods like fewest-switches surface hopping (FSSH) require accurate NACs.
• Machine learning (ML) NACs is challenging due to their complex nature.

Purpose of the Study:

• To develop accurate and efficient machine learning models for nonadiabatic couplings.
• To accelerate simulations of photochemical dynamics using machine learning.
• To overcome the challenges associated with learning vectorial and singular NACs.

Main Methods:

• Design of novel NAC-specific descriptors informed by domain expertise.
• Implementation of a new machine learning phase-correction procedure.
• Application to fully ML-driven FSSH simulations of fulvene and methylenimmonium cation.

Main Results:

• Achieved unprecedented accuracy in learning NACs with R² exceeding 0.99.
• Demonstrated efficient and robust ML-driven FSSH simulations.
• Accurate description of S₁ decay in fulvene and successful three-state simulation.

Conclusions:

• The developed ML approach significantly enhances the accuracy and efficiency of photochemical simulations.
• The method is generalizable to systems with multiple electronic states.
• Open-source implementation available in MLatom facilitates broader adoption.