Neural-Network Pauli Repulsion Potentials for Density-Functional Tight Binding for Large Molecular Systems

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). In our previous work [J. Phys. Chem. Lett. 11, 16 (2021)], we substantially improved the accuracy of DFTB method for the prediction of multiple properties of small molecules by developing ML repulsive potentials (NN_rep). However, these potentials do not account for long-range interactions which are crucial to investigate large/more flexible molecules and molecular dimers. Hence, we now employ a physics-inspired neural network architecture such as SpookyNet [Nat. Commun. 12, 7273 (2021)] to develop repulsive potentials supplemented with a treatment of many-body dispersion (MBD) interactions. In doing so, we show that the local and nonlocal interaction blocks in SpookyNet enhance the reliability of the MBD-corrected DFTB+NN_rep models in performing modeling techniques like (global) structure search or vibrational analysis of flexible molecules and molecular dimers. Our study thus provides valuable insights for the fast access to reliable property calculation of diverse molecular systems.