Machine learning accurate exchange and correlation functionals of the electronic density.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning applications

Background:

• Density functional theory (DFT) is crucial for atomic-scale electronic structure studies.
• The accuracy of Kohn-Sham DFT relies on approximations for exchange-correlation functionals.
• Balancing accuracy and computational cost is a key challenge in DFT.

Purpose of the Study:

• To introduce NeuralXC, a framework for developing density functionals using supervised machine learning.
• To enhance the accuracy of existing approximate functionals while preserving computational efficiency.
• To create transferable machine-learned functionals applicable across diverse chemical systems.

Main Methods:

• Development of the NeuralXC framework utilizing supervised machine learning.
• Training and optimization of machine-learned density functionals.
• Evaluation of functional performance on molecular systems, including bond-breaking scenarios.

Main Results:

• NeuralXC functionals learn meaningful physical information from training data, demonstrating transferability.
• A NeuralXC functional optimized for water outperforms other methods in describing bond breaking.
• The developed functionals achieve accuracy comparable to more computationally intensive methods.

Conclusions:

• NeuralXC represents a significant advancement in designing accurate and efficient density functionals.
• Machine learning offers a promising avenue for creating universal functionals for both molecules and solids.
• This approach has the potential to accelerate electronic structure calculations in various scientific domains.