CIDER: An Expressive, Nonlocal Feature Set for Machine Learning Density Functionals with Exact Constraints.

Area of Science:

• Computational chemistry
• Materials science
• Quantum mechanics

Background:

• Machine learning (ML) shows promise for developing accurate exchange-correlation (XC) functionals in density functional theory (DFT).
• Existing ML-developed functionals require improvements in accuracy, numerical stability, and chemical space transferability.
• Accurate XC functionals are crucial for predicting molecular and material properties.

Purpose of the Study:

• To introduce the CIDER formalism for nonlocal density features.
• To train a Gaussian process model for exchange energy using CIDER.
• To develop a more accurate and transferable ML-based exchange functional.

Main Methods:

• Developed the CIDER formalism to capture nonlocal density features.
• Trained a Gaussian process model for exchange energy using CIDER.
• Ensured the model obeys the uniform scaling rule for exchange.
• Tested the functional's accuracy and transferability across main-group molecules.

Main Results:

• The CIDER exchange functional demonstrates significantly higher accuracy than tested semilocal functionals.
• The functional exhibits good transferability across main-group molecules.
• The CIDER formalism provides a robust, physics-informed approach for ML-based XC models.

Conclusions:

• The CIDER exchange functional represents a significant advancement in ML-driven XC functional development.
• This work introduces effective techniques for creating robust, physics-informed XC models.
• The findings pave the way for more accurate and reliable DFT calculations.