Can exact conditions improve machine-learned density functionals?

Area of Science:

• Computational chemistry
• Quantum mechanics
• Machine learning in physics

Background:

• Density functional theory (DFT) relies on approximations for functionals.
• Analytic conditions historically guided functional development.
• Machine learning (ML) offers a data-driven approach to functional approximation.

Purpose of the Study:

• To investigate the impact of exact conditions on ML-approximated functionals.
• To analyze the performance of ML models for the non-interacting kinetic energy functional.
• To determine factors influencing the effectiveness of ML functional approximations.

Main Methods:

• Developed ML approximations for the non-interacting kinetic energy functional.
• Utilized exact conditions to guide the ML model.
• Tested performance on a simple one-dimensional system.
• Analyzed the role of the interpolation manifold in ML model accuracy.

Main Results:

• Incorporating exact conditions improved learning curves for the ML functional.
• The degree of improvement was sensitive to the nature of the ML functional's interpolation manifold.
• Demonstrated a hybrid approach combining analytic constraints and ML.

Conclusions:

• Exact conditions can enhance the accuracy of ML-based DFT functionals.
• The choice of interpolation manifold is crucial for successful ML functional development.
• This work suggests a promising direction for improving DFT accuracy through ML.