Neural-network Density Functional Theory Based on Variational Energy Minimization.

Area of Science:

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

Background:

• Deep-learning density functional theory (DFT) accelerates materials discovery but often uses isolated, data-driven supervised learning.
• Current methods lack integration between neural network development and DFT principles.

Purpose of the Study:

• To present a theoretical framework unifying neural network optimization and DFT variational computation.
• To enable physics-informed unsupervised learning for materials research.
• To demonstrate a new paradigm for developing advanced deep-learning DFT methods.

Main Methods:

• Developed a differential DFT code integrating a deep-learning DFT Hamiltonian.
• Implemented automatic differentiation and backpropagation algorithms within DFT.
• Created a physics-informed neural network architecture.

Main Results:

• The neural-network DFT framework successfully unifies optimization and computation.
• Demonstrated the capability of the developed differential DFT code.
• Achieved superior accuracy and efficiency compared to conventional approaches.

Conclusions:

• The proposed physics-informed neural-network DFT offers a novel paradigm for materials discovery.
• This integrated approach overcomes limitations of isolated supervised learning methods.
• Neural-network DFT significantly enhances accuracy and efficiency in computational materials science.