Towards transferable parametrization of Density-Functional Tight-Binding with machine learning

Machine learning (ML) has been proven to be an extremely valuable tool for simulations with ab initio accuracy at the computational cost between classical interatomic potentials and density-functional approximations. Similar efficiency can only be achieved by semi-empirical methods, such as density-functional tight-binding (DFTB). One of the limiting factors in terms of the accuracy and transferability of DFTB parametrizations is the so-called repulsive potential, which plays a considerable role for the prediction of energetic, structural, and dynamical properties. Few attempts of using ML-techniques to address this issue have been proposed recently but, up to now, evidence of transferability and scalability is still scarce. Using the QM7-X database of small organic molecules, we demonstrate that the DFTB repulsive energy can be effectively learned by means of ML-approaches including neural networks and kernel ridge regression. We further show how the resulting DFTB+ML model can also be used for more complex systems like molecular dimers and crystals, and modeling techniques like (global) structure search or vibrational analysis. DFTB+ML thus opens a route to the simultaneous access to reliable electronic and structural/dynamical properties of diverse molecular systems.