Constructing Diabatic Potential Energy Matrices with Quantum Dynamic Accuracy: A Neural Network Based Δ-Machine...

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning in Chemistry

Background:

• Accurate potential energy matrices (PEMs) are crucial for understanding molecular dynamics.
• Calculating high-level diabatic PEMs is computationally expensive.
• Existing methods struggle with global accuracy and efficiency for coupled states.

Purpose of the Study:

• To propose a new neural network (NN)-based Delta-machine learning (Δ-ML) approach.
• To construct global diabatic potential energy matrices (PEMs) for molecular systems.
• To reduce computational costs associated with high-level energy calculations.

Main Methods:

• Utilized inexpensive low-level energy data combined with a few high-level energies.
• Developed two NN-based Δ-ML schemes (A and B) for training adiabatic energy data.
• Applied the approach to nonadiabatic reactions (Na + H₂) and photodissociation (NH₃).

Main Results:

• Achieved effective and accurate global diabatic PEMs for both example systems.
• Reduced high-level calculation costs by approximately 87%.
• Successfully reproduced nonadiabatic reaction probabilities, absorption spectra, and product branching ratios.

Conclusions:

• The NN-based Δ-ML approach provides a computationally efficient and accurate method for PEM construction.
• Scheme B demonstrated superior training efficiency in the NH₃ system due to higher degrees of freedom.
• This method holds significant promise for advancing studies in molecular dynamics and spectroscopy.