Learning to Use the Force: Fitting Repulsive Potentials in Density-Functional Tight-Binding with Gaussian Process...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Density-Functional Tight Binding (DFTB) offers a cost-effective alternative to Density Functional Theory (DFT) for large-scale simulations.
• DFTB provides access to electronic properties like band structure, unlike empirical potentials.
• Parameter transferability in DFTB is limited, necessitating project-specific adjustments, especially for the repulsive potential (Vrep).

Purpose of the Study:

• To develop a machine learning approach for fitting the repulsive potential (Vrep) in DFTB.
• To overcome challenges associated with traditional Vrep parameterization methods.
• To enable more accurate and transferable DFTB simulations through data-driven parameter optimization.

Main Methods:

• Utilizing Gaussian Process Regression (GPR) to reconstruct Vrep.
• Training the GPR model using force residues from DFT-DFTB calculations.
• Applying the method to organic molecules containing carbon, hydrogen, and oxygen.

Main Results:

• Successfully reconstructed Vrep using GPR, bypassing the need for complex optimization.
• Demonstrated the method's flexibility in handling multiple elements simultaneously.
• Showcased a data-driven philosophy for DFTB parameterization, shifting focus from functional form to data quality.

Conclusions:

• The proposed ML approach offers a robust and flexible way to parameterize DFTB's repulsive potential.
• This method enhances the accuracy and transferability of DFTB simulations.
• The data-driven strategy facilitates routine parameter set adjustment for specific computational projects.