Charting electronic-state manifolds across molecules with multi-state learning and gap-driven dynamics via efficient...

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning

Background:

• Machine learning (ML) is underutilized in excited-state simulations due to several challenges.
• Accurate modeling of electronic states is crucial for photophysical and photochemical processes.

Purpose of the Study:

• To present a robust and affordable protocol for learning electronic states to accelerate molecular simulations.
• To improve the accuracy and efficiency of excited-state simulations using ML.

Main Methods:

• Introduced a physics-informed multi-state ML model capable of learning numerous excited states.
• Developed gap-driven dynamics for accelerated sampling of critical low-energy gap regions.
• Implemented active learning with physics-informed uncertainty quantification for robust model generation.

Main Results:

• The multi-state ML model achieves accuracy comparable to or better than ground-state energy predictions.
• Excited-state energy information enhances the quality of ground-state predictions.
• The protocol enabled efficient active learning and uncovered long-time-scale oscillations in cis-azobenzene photoisomerization.

Conclusions:

• The developed protocol overcomes limitations of ML in excited-state simulations.
• The combination of multi-state learning and gap-driven dynamics provides robust models for surface-hopping simulations.
• This approach accelerates the discovery of complex molecular dynamics and photochemical processes.