Learning the Exchange-Correlation Functional from Nature with Fully Differentiable Density Functional Theory.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Accurate prediction of molecular properties via ab initio simulations is crucial for materials discovery.
• Current deep neural network applications in quantum chemistry are hindered by limited and varied experimental data.

Purpose of the Study:

• To develop a novel machine learning approach for improving the accuracy of quantum chemistry simulations.
• To train a neural network to replace the exchange-correlation functional within the Kohn-Sham density functional theory framework.

Main Methods:

• Implemented a fully differentiable three-dimensional Kohn-Sham density functional theory framework.
• Trained an exchange-correlation neural network using a small dataset of eight experimental data points on diatomic molecules.

Main Results:

• The trained exchange-correlation networks significantly improved simulation accuracy.
• Achieved enhanced prediction accuracy for atomization energies across 104 diverse molecules, including those with novel bonds and atoms not in the training set.

Conclusions:

• This study demonstrates the potential of machine learning-driven exchange-correlation functionals to advance computational materials discovery.
• The method shows promise for accurate molecular property prediction with minimal experimental data, overcoming a key limitation in deep learning for quantum chemistry.