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

We present recent work on the development of a fully-differentiable density functional theory simulator (DQC – Differentiable Quantum Chemistry) where the exchange-correlation functional is represented by a trainable neural network. We demonstrate how the computing paradigm of automatic differentiation, constrained by the Kohn-Sham framework, can provide a novel way to construct highly accurate exchange-correlation functionals using heterogeneous experimental data even for highly limited datasets. 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, atoms, and molecules, not present in the training dataset.