dxtb-An efficient and fully differentiable framework for extended tight-binding.

Area of Science:

• Computational Chemistry
• Machine Learning
• Quantum Chemistry

Background:

• Automatic differentiation (AD) is crucial for machine learning optimization.
• AD's utility is now recognized in quantum chemistry for derivative calculations.
• Semiempirical extended tight-binding (xTB) methods offer a balance of accuracy and computational cost.

Purpose of the Study:

• To present dxtb, an open-source, fully differentiable framework for xTB methods.
• To enable efficient computation of molecular properties and facilitate machine learning integration.
• To provide a foundation for physics-inspired, end-to-end differentiable models.

Main Methods:

• Developed in Python utilizing PyTorch for array operations.
• Implemented comprehensive code vectorization and optimization for computational efficiency.
• Leveraged automatic differentiation for calculating arbitrary-order derivatives.

Main Results:

• dxtb achieves performance comparable to compiled xTB programs for small molecules.
• Energy evaluations are on par with existing programs.
• Automatically differentiated nuclear derivatives are 2-5 times slower than analytical counterparts.
• Demonstrated utility in calculating molecular and spectroscopic properties.

Conclusions:

• dxtb streamlines optimization and property evaluation in quantum chemistry.
• It facilitates seamless integration of semiempirical quantum chemistry with machine learning.
• The framework advances semiempirical methods and supports hybrid machine learning applications.