Reinventing Density Functional Theory with Machine Learning on Integral Features

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Kohn-Sham density functional theory (KS-DFT) is crucial for electronic structure modeling.
• Accuracy of KS-DFT relies on density functional approximations.
• Existing methods face limitations in balancing accuracy and computational cost.

Purpose of the Study:

• Introduce a novel strategy for developing accurate density functionals.
• Leverage machine learning with integrated local descriptors.
• Improve computational efficiency while enhancing functional dependence.

Main Methods:

• Developed a multilayer perceptron learning nonlinear functional of "integral features".
• Formulated an "integral-features" approach applying the model once per calculation.
• Trained ML25@MN15 functional on 185 databases covering diverse chemical properties.

Main Results:

• ML25@MN15 achieved a mean unsigned error of 1.05 kcal/mol on 7232 energetic data points.
• Demonstrated superior performance over leading modern functionals across various categories.
• Maintained computational cost comparable to standard functionals like MN15.

Conclusions:

• Integral features offer a computationally efficient path to high-accuracy density functionals.
• Machine learning can determine complex density functional dependencies.
• This strategy advances electronic structure modeling, especially for systems with transition metals.