What If Machine Learning Could Learn the Electrons Too? Towards ML-Enhanced DFTB for Electronic Structure Prediction

Machine learning (ML) is transforming condensed matter physics and materials science, enabling the development of interatomic potentials that achieve near–first-principles accuracy at a fraction of the computational cost. However, while many ML frameworks have achieved remarkable success in predicting total energies, forces, and stress tensors, the explicit description of electronic properties remains a largely unexplored frontier. This gap limits the scope of ML-based methods when the target quantities depend directly on the electronic structure, such as in defect physics, charge transport, or optical processes.

Density Functional Tight Binding (DFTB) provides an appealing bridge between electronic-structure theory and efficient atomistic modeling. As an approximation to Density Functional Theory (DFT) based on precomputed Hamiltonians, DFTB inherently encodes electronic information. Yet, its accuracy and transferability depend heavily on the underlying Slater–Koster (SK) parameter sets, which are specific to element pairs and typically hand-tuned for narrow chemical environments. This rigidity often confines DFTB to the role of a configurational generator for subsequent ab initio refinement rather than a predictive method in its own right.

In this work, we explore how integrating ML with DFTB can help overcome these long-standing limitations. By employing neural-network models to learn and generalize trends in confinement potentials, we fine-tune the effective Hamiltonians governing electronic behavior while preserving the accuracy of energies and forces. Our results demonstrate that this ML-enhanced DFTB framework can systematically capture variations in electronic band structures and polaronic states across diverse configurations, opening the door to fast, transferable, and electronically aware simulations.