Permutationally Restrained Diabatization by Machine Intelligence.

Area of Science:

• Computational Chemistry
• Quantum Dynamics
• Machine Learning in Chemistry

Background:

• Simulating electronically nonadiabatic processes requires adiabatic or diabatic representations.
• Direct dynamics in the adiabatic basis is computationally expensive with accurate electronic structure theories.
• Fitting adiabatic potential energy surfaces is challenging due to cusps and singularities.

Purpose of the Study:

• To extend the diabatization by deep neural network (DDNN) method for improved nonadiabatic dynamics simulations.
• To develop a method capable of approximating global, permutationally invariant adiabatic potential energy surfaces.
• To generate diabatic potential energy matrices (DPEMs) and adiabatic surfaces simultaneously.

Main Methods:

• Utilized a machine learning approach, diabatization by deep neural network (DDNN), for generating diabatic representations.
• Extended the DDNN method to incorporate permutational invariance constraints, termed permutationally restrained DDNN.
• Leveraged the smoothness and nonuniqueness of diabatic bases for efficient learning of potential energy matrices.

Main Results:

• The permutationally restrained DDNN method successfully produces accurate approximations to global adiabatic potential energy surfaces.
• Simultaneous generation of diabatic potential energy matrices (DPEMs) and adiabatic surfaces is achieved.
• The method yields analytic forms for diabatic and adiabatic quantities, facilitating dynamics calculations.

Conclusions:

• The permutationally restrained DDNN method offers a computationally efficient and accurate approach for nonadiabatic dynamics.
• This extended DDNN technique overcomes the fitting challenges associated with adiabatic representations.
• The ability to generate permutationally invariant adiabatic surfaces alongside diabatic ones enhances simulation reliability.