TBMaLT, a flexible toolkit for combining tight-binding and machine learning.

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Tight-binding (TB) methods, including Density Functional Tight-Binding (DFTB), offer efficient quantum mechanical simulations for large systems.
• These methods balance speed and accuracy by combining ab initio density functional theory with empirical approximations.

Purpose of the Study:

• To improve the accuracy and transferability of TB methods using machine learning.
• To develop an open-source toolkit for integrating machine learning with TB approaches.

Main Methods:

• Incorporating local atomic environment information into machine learning parameter tuning for TB models.
• Developing the Tight-Binding Machine Learning Toolkit (TBMLT) with modular interfaces.
• Utilizing existing Hamiltonians like GFN1-xTB within the toolkit.

Main Results:

• Machine learning significantly enhances the accuracy of TB methods with shorter, more transferable learning procedures.
• The TBMLT facilitates the implementation of combined ML-TB approaches.
• Proof-of-concept applications demonstrate the toolkit's functionality and the combined methods' potential.

Conclusions:

• Combining machine learning with TB methods provides a powerful and efficient approach for quantum mechanical simulations.
• The TBMLT is a versatile framework for developing and applying advanced computational chemistry tools.
• This approach offers accurate predictions and enables calculation of derived quantum mechanical quantities without additional learning.