A Density Functional Tight Binding Layer for Deep Learning of Chemical Hamiltonians

TL;DR

This paper integrates self-consistent-charge DFTB theory into deep learning models as a differentiable layer, enabling more accurate predictions of molecular electronic properties by training on quantum chemistry data.

Contribution

It introduces a novel DFTB layer for deep learning models, allowing direct incorporation of quantum chemistry calculations into neural network training.

Findings

Spline model reduces energy error by 60% on test molecules.

Neural network model reduces dipole moment error by 59%.

Models trained on larger molecules perform significantly better.

Abstract

Current neural networks for predictions of molecular properties use quantum chemistry only as a source of training data. This paper explores models that use quantum chemistry as an integral part of the prediction process. This is done by implementing self-consistent-charge Density-Functional-Tight-Binding (DFTB) theory as a layer for use in deep learning models. The DFTB layer takes, as input, Hamiltonian matrix elements generated from earlier layers and produces, as output, electronic properties from self-consistent field solutions of the corresponding DFTB Hamiltonian. Backpropagation enables efficient training of the model to target electronic properties. Two types of input to the DFTB layer are explored, splines and feed-forward neural networks. Because overfitting can cause models trained on smaller molecules to perform poorly on larger molecules, regularizations are applied that penalize non-monotonic behavior and deviation of the Hamiltonian matrix elements from those of the published DFTB model used to initialize the model. The approach is evaluated on 15,700 hydrocarbons by comparing the root mean square error in energy and dipole moment, on test molecules with 8 heavy atoms, to the error from the initial DFTB model. When trained on molecules with up to 7 heavy atoms, the spline model reduces the test error in energy by 60% and in dipole moments by 42%. The neural network model performs somewhat better, with error reductions of 67% and 59% respectively. Training on molecules with up to 4 heavy atoms reduces performance, with both the spline and neural net models reducing the test error in energy by about 53% and in dipole by about 25%.