High-Throughput Electronic Structure Prediction of Materials using Machine-Learned Extended Hückel Tight-Binding Models

Accurate knowledge of electronic structures is crucial for high-throughput materials screening and inverse design of materials for many applications, including thermal transport, tunable electronics, and sensing. Electronic structures play a crucial role in determining the magnetic, electrical, and optical responses across different classes of materials, such as transition metals, low-dimensional materials, and especially materials with defects. While first-principles calculations based on Density Functional Theory (DFT) have demonstrated the numerical accuracy required for predicting the electronic structure, the principal challenge lies in their scalability, particularly in systems with heterogeneity and sizes that expand much beyond their unit cell representation, which makes them intractable for high-throughput applications. The Extended Hückel (eH) tight-binding method is a semiempirical quantum chemistry method that can predict a qualitatively correct electronic picture of materials with computational ease and clarity. While earlier studies have demonstrated the efficacy of the eH model in predicting electronic structures (e.g., phases of the Sr–Ti–O family), there is currently no generalized parameterization workflow to accurately parameterize eH tight-binding models across a broad class of materials. In this work, we leverage machine learning to develop a workflow that parametrizes the eH tight-binding model by mapping complex atomic configurations to tight-binding Hamiltonians and comparing them with the DFT-predicted electronic structure from material databases. We demonstrate the efficacy of our workflow in predicting electronic structures across different elemental systems such as C, Ge, Si, and P of different phases and dimensionalities. The predicted Hamiltonian by the parameterized eH model allows us to evaluate the electronic structures of two-dimensional B in the presence of defects and large-scale phase change materials (GexSbyTez) in both crystalline, amorphous, and mixed phases.