Evolving symbolic density functionals.

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Developing accurate density functionals is a long-standing challenge in computational science.
• Existing machine learning (ML) functionals often have too many parameters, hindering interpretability and integration.
• A gap exists between complex ML functionals and traditional human-designed symbolic functionals.

Purpose of the Study:

• To introduce a novel framework, Symbolic Functional Evolutionary Search (SyFES), for automated symbolic functional development.
• To create density functionals that are explainable, computationally cheaper, and easier to integrate into existing software.
• To advance the systematic development of symbolic density functionals using computational power.

Main Methods:

• SyFES automatically constructs functionals in a symbolic form.
• The framework was tested by reconstructing a known functional without prior knowledge.
• SyFES evolved an existing functional (ωB97M-V) to discover new, improved functionals.

Main Results:

• SyFES successfully reconstructed a known functional from scratch, demonstrating its capability.
• A new functional, GAS22 (Google Accelerated Science 22), was discovered by evolving ωB97M-V.
• GAS22 exhibited superior performance across most molecular types in the MGCDB84 dataset.

Conclusions:

• SyFES provides an effective method for the automated, systematic development of symbolic density functionals.
• The SyFES framework bridges the gap between complex ML models and interpretable symbolic functionals.
• This approach opens new avenues for leveraging computational resources in functional development.