SlaKoNet: A Unified Slater-Koster Tight-Binding Framework Using Neural Network Infrastructure for the Periodic Table

Area of Science:

• Computational Materials Science
• Quantum Chemistry
• Solid-State Physics

Background:

• Accurate electronic band structure prediction is crucial for materials design.
• Existing machine learning models lack universality, and traditional Slater-Koster (SK) models have limited transferability and require manual parameterization.
• Training on low-fidelity data hinders traditional SK model performance.

Purpose of the Study:

• To develop a universal and efficient framework for predicting electronic band structures.
• To overcome the limitations of existing machine learning and traditional tight-binding models.
• To enable rapid electronic structure screening for novel materials discovery.

Main Methods:

• Introduced SlaKoNet, a parameter optimization framework learning SK-based Hamiltonian matrix elements.
• Utilized automatic differentiation for parameter learning across 65 elements.
• Trained on over 20,000 materials from the JARVIS-DFT database using the TBmBJ functional.

Main Results:

• Achieved a mean absolute error (MAE) of 0.74 eV for bandgap predictions against experimental data.
• Demonstrated improvement over standard GGA functionals (MAE = 1.14 eV).
• Showcased scalability with up to 8.4x speedup on GPUs.

Conclusions:

• SlaKoNet offers a computationally advantageous and interpretable alternative to traditional methods.
• The framework enables efficient and accurate electronic structure screening.
• SlaKoNet facilitates accelerated materials discovery with targeted properties.