DeePKS: A Comprehensive Data-Driven Approach toward Chemically Accurate Density Functional Theory.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Generalized Kohn-Sham density functional theory (DFT) is a powerful tool for electronic structure calculations.
• Developing accurate and broadly applicable energy functionals remains a significant challenge in DFT.
• Existing methods often struggle with accuracy across diverse chemical systems.

Purpose of the Study:

• To introduce a novel machine learning (ML) framework for constructing energy functionals.
• To enable the training of self-consistent models using large, diverse datasets.
• To achieve chemically accurate predictions for molecular properties.

Main Methods:

• A general ML-based framework was developed for energy functional construction.
• Self-consistent models were trained on extensive datasets encompassing various systems and labels.
• The framework integrates generalized Kohn-Sham DFT with advanced ML techniques.

Main Results:

• The resulting energy functional demonstrated high accuracy for energy, force, dipole, and electron density predictions.
• Chemically accurate results were obtained for a wide range of molecules.
• The functional's performance showed continuous improvement with increasing data availability.

Conclusions:

• The proposed ML framework offers a pathway to highly accurate and generalizable energy functionals.
• This approach significantly advances the capabilities of density functional theory.
• The method is scalable and can be further refined with more data.