<br/>The high computational cost of ab initio electronic structure calculations remains a challenge for the computational design of molecules and materials. Semi-empirical models, such as Self Consistent Charge Density Functional Tight Binding Theory (DFTB), can compute electronic structure at a greatly reduced cost. However, the accuracy of such models is insufficient for many applications. We will present a means to systematically improve the accuracy of such models while maintaining their low computational cost. The key enabling technology is the implementation of DFTB as a layer that can be integrated into the deep learning models of machine learning. This DFTB layer takes, an input, DFTB Hamiltonian matrix elements generated from previous layers, such as a feed-forward neural net. The DFTB layer generates, as output, electronic properties from self-consistent solutions of the corresponding DFTB Hamiltonian. The DFTB layer allows backpropagation, such that the system can be trained efficiently to data on electronic properties. We will present results on the use of this approach to predict electronic properties of organic molecules and other molecular systems.

Deep learning of quantum chemical Hamiltonians

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.