Obtaining Electronic Properties of Molecules through Combining Density Functional Tight Binding with Machine...

Area of Science:

• Computational chemistry
• Materials science
• Quantum mechanics

Background:

• Density Functional Tight Binding (DFTB) is a cost-effective quantum mechanical method for large systems.
• Existing DFTB parametrizations often struggle to accurately reproduce diverse electronic properties.
• Optimization of electronic properties is crucial for accurate molecular modeling and material design.

Purpose of the Study:

• To introduce a novel machine learning (ML) workflow for optimizing electronic properties within the DFTB framework.
• To explore two distinct ML approaches for generating essential two-center integrals.
• To evaluate the performance of the ML workflow against existing DFTB parametrizations.

Main Methods:

• Developed an ML workflow to generate two-center integrals for DFTB.
• Implemented two training strategies: direct basis function parameter training and spline model training for diatomic integrals.
• Used generated integrals to construct Hamiltonian and overlap matrices for electronic structure calculations.

Main Results:

• Both ML approaches improved electronic properties like charge distributions, dipole moments, and polarizabilities compared to standard DFTB.
• Direct training on basis function parameters resulted in consistent Hamiltonians and overlap matrices within physically reasonable ranges.
• Simultaneous improvement of multiple electronic properties was achieved exclusively through the basis function parameter training method.

Conclusions:

• The proposed ML workflow effectively optimizes electronic properties in DFTB calculations.
• Directly training basis function parameters offers a more robust approach for achieving accurate and stable electronic structure calculations.
• This work provides a pathway for developing more accurate and reliable DFTB models for diverse chemical applications.