Predicting HSE band gaps from PBE charge densities via neural network functionals

Area of Science:

• Computational Materials Science
• Quantum Chemistry
• Machine Learning in Physics

Background:

• Density Functional Theory (DFT) is standard for material property calculations but struggles with excited-state properties like band gaps.
• Accurate methods (e.g., GW, hybrid DFT functionals) are computationally expensive for large systems.
• Predicting band gaps is crucial for understanding material behavior and designing new materials.

Purpose of the Study:

• To investigate the feasibility of using neural networks trained on DFT charge densities to predict accurate band gaps.
• To develop a computationally efficient method for predicting material band gaps.
• To enable high-throughput screening of complex materials.

Main Methods:

• Training neural networks on traditional DFT charge densities computed with a standard Perdew-Burke-Ernzerhof (PBE) functional.
• Using the trained network to predict HSE (Heyd-Scuseria-Ernzerhof) band gaps for various materials.
• Comparing the accuracy of the neural network approach with standard regression methods.

Main Results:

• A single neural network trained on PBE charge densities accurately predicted HSE band gaps for seven diverse materials.
• Achieved a Root Mean Square Error (RMSE) of 172.6 meV.
• Demonstrated a 34% improvement in accuracy compared to standard regression between PBE and HSE band gaps.

Conclusions:

• Neural networks trained on DFT charge densities offer a promising approach to accurately predict material band gaps.
• This method significantly reduces computational cost and increases prediction accuracy.
• The approach has the potential for accurate high-throughput screening of complex materials.