Tests of Artificial Neural Network-Based Diabatization Approaches on Simple 1D Models

Area of Science:

• Quantum Chemistry
• Computational Chemistry
• Theoretical Chemistry

Background:

• A novel diabatization scheme using artificial neural networks (ANNs) was recently introduced.
• This method primarily relies on adiabatic energies from ab initio calculations.
• Unanswered questions regarding the method's performance and optimization exist.

Purpose of the Study:

• To investigate and improve the performance of the novel ANN-based diabatization scheme.
• To address unanswered questions concerning the method's effectiveness.
• To explore optimizations for the diabatization process using ANNs.

Main Methods:

• Testing various activation functions, including nonlinear output layer functions.
• Analyzing the impact of regularization terms in the loss function.
• Proposing cost-effective methods for extending training datasets.

Main Results:

• Significant improvements in the performance of the original diabatization method were achieved.
• The study identified effective activation functions and regularization strategies.
• Computationally inexpensive training set extensions were found to be beneficial.

Conclusions:

• The optimized ANN-based diabatization scheme demonstrates enhanced performance.
• The findings provide practical guidance for applying and improving ANN-based diabatization.
• Further research can build upon these optimizations for broader applications in quantum chemistry.