GradDFT. A software library for machine learning enhanced density functional theory.

Area of Science:

• Computational Quantum Chemistry
• Materials Science
• Machine Learning

Background:

• Density Functional Theory (DFT) is widely used but lacks accuracy for strongly correlated systems.
• Machine learning (ML) offers potential to enhance DFT capabilities, but faces challenges.

Purpose of the Study:

• To develop a differentiable library for ML-enhanced DFT.
• To enable rapid prototyping and experimentation with novel exchange-correlation functionals.

Main Methods:

• Introduced GradDFT, a JAX-based, fully differentiable DFT library.
• Developed a novel parametrization of exchange-correlation functionals using neural networks.
• Compiled a dataset of experimental dimer dissociation energies for training and benchmarking.

Main Results:

• Demonstrated the generalization capabilities of ML-enhanced functionals across different systems.
• Assessed the impact of training data noise on model accuracy.
• Showcased a differentiable self-consistent iterative procedure for DFT calculations.

Conclusions:

• GradDFT facilitates the development of accurate ML-enhanced DFT functionals.
• The approach shows promise for improving calculations of strongly correlated systems.
• Further research can leverage GradDFT for advancing computational materials science.