Generalized Density-Functional Tight-Binding Repulsive Potentials from Unsupervised Machine Learning.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• The approximate density-functional tight-binding (DFTB) method offers computational efficiency for large systems but often lacks accuracy and transferability.
• Parametrization of DFTB requires extensive data and manual effort, limiting its widespread application.
• Developing accurate and transferable interatomic potentials is crucial for molecular simulations.

Purpose of the Study:

• To enhance the accuracy and transferability of the DFTB method using unsupervised machine learning.
• To automate the parametrization process of DFTB by incorporating chemical environment information.
• To develop more specific and effective two-body potentials for molecular modeling.

Main Methods:

• Combined the approximate density-functional tight-binding (DFTB) method with unsupervised machine learning algorithms.
• Introduced generalized pair-potentials that account for the chemical environment during the learning process.
• Trained the model on energies and forces from 2100 equilibrium and non-equilibrium molecular structures.

Main Results:

• Achieved reproduction of atomization energies within an error of approximately 2.6 kcal/mol, a significant improvement over standard DFTB.
• Successfully tested the method on a large dataset of ~130,000 organic molecules containing O, N, C, H, and F atoms.
• Demonstrated improved transferability and accuracy of the machine learning-parametrized DFTB method.

Conclusions:

• Unsupervised machine learning provides an effective strategy for automating and improving DFTB parametrization.
• The developed generalized pair-potentials enhance the accuracy of DFTB for predicting properties of organic molecules.
• This integrated approach offers a powerful tool for large-scale molecular simulations in computational chemistry and materials science.