Essentiality of the Basis Function in Deep Learning Physical Chemistry Properties

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning

Background:

• Deep learning (DL) models often struggle with transferability due to inappropriate basis functions for feature generation.
• Selecting the right basis functions is crucial for effective dimension elevation in DL applications.

Purpose of the Study:

• To develop and evaluate novel basis functions for enhanced feature generation in DL for molecular and atomic interactions.
• To improve the transferability, interpretability, and performance of DL models in quantum physical chemistry.

Main Methods:

• Utilized associated Laguerre polynomials and spherical harmonics for feature generation.
• Applied these functions to model molecules and atomic interactions in both real and momentum spaces.
• Investigated the symmetry properties of spherical harmonics for calculating excited state orbitals.

Main Results:

• Demonstrated superior performance of the proposed basis functions compared to previous models.
• Significantly enhanced model transferability and interpretability to novel elements and structures.
• Enabled accurate calculation of energy and distribution of excited state orbitals by leveraging symmetry.

Conclusions:

• Associated Laguerre and spherical harmonic functions offer a robust approach for DL in chemistry.
• This method improves DL model generalization and interpretability in quantum physical chemistry.
• The approach shows potential for broader applications, including image classification.