GTAttn-XC: Physically constrained attention for nonlocal density functionals.

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Machine-learning (ML) enhances electronic density functional theory (DFT) approximations.
• Current ML-DFT methods struggle with long-range interactions due to handcrafted descriptors.
• Scalability issues limit the modeling of complex electronic responses.

Purpose of the Study:

• Introduce GTAttn-XC, a novel ML model for exchange-correlation functional approximation.
• Enable unsupervised modeling of long-range electronic responses using attention mechanisms.
• Develop a unified approach for local accuracy and nonlocal interactions in DFT.

Main Methods:

• Coupled a multiscale real-space grid graph representation with attention operators.
• Developed a learnable attention mechanism for unsupervised learning of electronic responses.
• Avoided computational overhead of explicit high-order correlation terms.

Main Results:

• GTAttn-XC demonstrated high accuracy across diverse tasks.
• Evaluated on benchmark datasets including MGCDB84.
• Achieved consistent accuracy in weak interactions, reaction energies, barrier heights, and thermochemical energies.

Conclusions:

• GTAttn-XC offers a novel pathway for high-accuracy nonlocal exchange-correlation approximations.
• The model effectively captures both local and nonlocal electronic interactions.
• This work advances the applicability of DFT for complex systems.