Pushing the frontiers of density functionals by solving the fractional electron problem.

Area of Science:

• Quantum Chemistry
• Computational Materials Science
• Artificial Intelligence in Science

Background:

• Density functional theory (DFT) is crucial for describing quantum mechanical properties of matter.
• Existing DFT approximations contain systematic errors due to violations of exact functional properties.
• These errors limit the accuracy of DFT for various chemical and material systems.

Purpose of the Study:

• To overcome fundamental limitations in traditional density functional theory approximations.
• To develop a novel functional that accurately describes systems with fractional charge and spin.
• To establish a pathway toward the exact universal functional for quantum mechanical calculations.

Main Methods:

• Training a neural network on molecular data.
• Incorporating fictitious systems with fractional charge and spin during training.
• Developing the DM21 (DeepMind 21) functional based on neural network outputs.

Main Results:

• DM21 demonstrates superior performance over traditional functionals in benchmarks for main-group atoms and molecules.
• The new functional accurately models challenging systems, including hydrogen chains, charged DNA base pairs, and diradical transition states.
• DM21 correctly describes artificial charge delocalization and strong correlation phenomena.

Conclusions:

• The DM21 functional represents a significant advancement in density functional theory.
• The data-driven, constraint-based methodology offers a viable route to the exact universal functional.
• This approach promises improved accuracy for a wide range of quantum chemical and materials science applications.