Hybrid Density Functional based Tight Binding-Molecular Mechanics Approach with Machine Learned Potentials for Large-Scale Materials Modeling

Density Functional based Tight Binding method (DFTB) has seen a rise in adoption in recent years as it offers both DFT-like accuracy and access to electronic structure as well as scalability for systems where appropriate parameters are available. DFTB is based on the parameterization of interatomic interactions that reduce the cost of calculation when compared to DFT. In DFTB, all interatomic interactions for all atoms of the model have to be pre-parameterized. Conventionally, these parameterizations are performed to multiple DFT calculations and or reference data. This approach, however, requires a significant number of comparatively high-cost DFT calculations and specific knowledge and skillsets to be performed appropriately, therefore reducing the applicability of DFTB for many heteroatomic materials. In this study, we propose a hybrid approach where interatomic interactions which have little influence on electronic properties are modeled with pairwise interatomic potentials while other interactions are modeled with the conventional Slater-Koster approximation. We demonstrate the usefulness of this method by modeling a large-scale silica-titania interface where the Si-Ti interactions are modeled at the force field level while other interactions are modeled at the DFTB level. The most appropriate functional representation of the potential and its parameters are obtained with a combination of machine learning and gradient-based fitting. The validity of the approximation is verified against DFT on small-scale models. The approach allows the expansion of DFTB application to new materials where Slater-Koster parameters are not yet available.