Exciting DeePMD: Learning excited-state energies, forces, and non-adiabatic couplings

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Materials Science

Background:

• Non-adiabatic dynamics simulations are crucial for understanding chemical reactions and excited-state processes.
• Existing machine learning methods often approximate non-adiabatic coupling vectors (NACVs), limiting accuracy.
• The DeePMD architecture is effective for ground-state properties but requires extension for excited states.

Purpose of the Study:

• To extend the DeePMD neural network architecture for accurate prediction of electronic structure properties in non-adiabatic dynamics.
• To develop a robust method for learning non-adiabatic coupling vectors (NACVs) from local chemical environments.
• To overcome approximations in existing machine learning approaches for NACVs.

Main Methods:

• Extension of the DeePMD neural network to predict excited-state energies and forces.
• Implementation of Richardson's method to learn the symmetric dyad of energy-difference-scaled NACVs.
• Utilizing local chemical environment descriptors within the DeePMD framework.

Main Results:

• Successfully learned the map between NACVs and local chemical environment descriptors.
• Demonstrated the efficiency and accuracy of the extended DeePMD architecture.
• Validated the approach using the methaniminium cation (CH2NH2+) as a test case.

Conclusions:

• The developed neural network architecture accurately predicts electronic structure properties for non-adiabatic dynamics.
• The method overcomes limitations of previous approximations for NACVs.
• This advancement enables more reliable simulations of excited-state chemical processes.