Toward the Exact Exchange-Correlation Potential: A Three-Dimensional Convolutional Neural Network Construct

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning

Background:

• Accurate exchange-correlation potentials are crucial for density functional theory (DFT) calculations.
• Developing universally applicable and accurate exchange-correlation functionals remains a significant challenge in quantum chemistry.

Purpose of the Study:

• To develop a deep neural network (DNN) capable of predicting exact exchange-correlation potentials.
• To demonstrate the DNN's ability to generalize from small molecular data to larger systems and interactions.

Main Methods:

• Construction of a deep neural network architecture.
• Training and validation using electron density data from small molecules (H2, HeH+, He2).
• Application of the trained DNN to predict potentials for larger/complex systems (stretched HeH+, linear H3+, H-He-He-H2+).

Main Results:

• The DNN successfully predicts exchange-correlation potentials with high accuracy.
• The model accurately captures van der Waals interactions, as shown with He2.
• The DNN demonstrates good transferability to systems not included in the training set.

Conclusions:

• Deep neural networks offer a promising route to achieving accurate and transferable exchange-correlation potentials in DFT.
• This data-driven approach can overcome limitations of traditional functional development.
• The method shows potential for improving the accuracy of electronic structure calculations for diverse molecular systems.