Learning the exchange-correlation functional from nature with fully differentiable density functional theory

TL;DR

This paper demonstrates that training neural networks to replace the exchange-correlation functional within a differentiable DFT framework can significantly enhance the accuracy of molecular property predictions, even with minimal experimental data.

Contribution

The authors introduce a method to train neural networks as exchange-correlation functionals in a fully differentiable DFT framework, improving molecular property predictions with limited data.

Findings

Neural network-based functionals outperform traditional methods in predicting atomization energies.

Only eight experimental data points are needed to train effective functionals.

The approach generalizes well to molecules with unseen bonds and atoms.

Abstract

Improving the predictive capability of molecular properties in ab initio simulations is essential for advanced material discovery. Despite recent progress making use of machine learning, utilizing deep neural networks to improve quantum chemistry modelling remains severely limited by the scarcity and heterogeneity of appropriate experimental data. Here we show how training a neural network to replace the exchange-correlation functional within a fully-differentiable three-dimensional Kohn-Sham density functional theory (DFT) framework can greatly improve simulation accuracy. Using only eight experimental data points on diatomic molecules, our trained exchange-correlation networks enable improved prediction accuracy of atomization energies across a collection of 104 molecules containing new bonds and atoms that are not present in the training dataset.