Learning Multiple Potential Energy Surfaces by Automated Discovery of a Compatible Representation.

Area of Science:

• Computational Chemistry
• Chemical Dynamics
• Theoretical Chemistry

Background:

• Modeling electronically nonadiabatic processes requires accurate potential energy surfaces (PESs).
• Developing analytic representations for multiple coupled PESs presents a significant challenge in chemical dynamics.
• Existing methods often lack automation or accuracy for complex polyatomic systems.

Purpose of the Study:

• Introduce a novel, automatic method for learning multiple potential energy surfaces (PESs) and their gradients.
• Develop a deep neural network approach for efficient and accurate representation of coupled PESs.
• Provide a tool for the chemical dynamics community to model nonadiabatic processes.

Main Methods:

• Developed the compatibilization by deep neural network (CDNN) method.
• Utilized a database of geometries and potential energies as input.
• Generated a compatible potential energy matrix (CPEM) via a specialized CDNN architecture.
• Obtained analytic adiabatic PESs and gradients through diagonalization and automatic differentiation.

Main Results:

• Demonstrated the accuracy and automatic nature of the CDNN method.
• Showcased the ability of CDNN to discover the CPEM, which acts as an implicit electronic Hamiltonian.
• Validated the method's effectiveness for polyatomic systems.

Conclusions:

• The CDNN method offers an accurate and fully automatic approach for learning coupled PESs.
• This technique simplifies the modeling of electronically nonadiabatic processes in polyatomic systems.
• CDNN is expected to be highly valuable for practical applications in chemical dynamics research.