DFTB-ML Hybrid for Scalable Electronic Analysis and Charge Transfer Predictions

Machine-learned interatomic potentials (MLIPs) enable long-time, large-scale molecular dynamics (MD) bridging the accuracy of density functional theory or quantum chemical methods and the computational cost of classical forcefields. However, this computational speed advantage typically comes at the cost of losing explicit electronic structure information of the simulated system. Here, we present a hybrid approach combining MLIPs for evolving the atomic structure of the system with density functional tight binding (DFTB) for efficient electronic structure evaluation on MLIP MD snapshots.

Our DFTB Hamiltonian incorporates self-consistency in terms of Mulliken charges allowing for prediction of charge transfer. With total energies and forces for structural evolution provided by the MLIP, repulsive terms in the DFTB Hamiltonian are omitted. This allows us to evaluate the electronic structure at select time intervals. The omission of the repulsive term fits furthermore simplifies the DFTB parametrization as only charge-response and basis optimization within the Slater-Koster framework is required.

We demonstrate the effectiveness of the hybrid DFTB-ML approach for predicting electronic descriptors capturing the reactivity of transition metal surfaces exposed on evolving catalyst nanoparticles. Additionally, we validate the accuracy of our DFTB parametrizations in capturing charge transfer processes during interfacial chemical dynamics.