Predictive tight-binding model for materials design and understanding

Parameterized tight-binding is a potentially ideal method for large-scale materials design applications, as it combines the speed of

model calculations with built-in physics and chemistry, including quantum mechanics. However, in practice the method is often limited by

the availability of relevant and predictive interaction parameter datasets. In this work, we discuss progress in fitting and testing of

the ThreeBodyTB.jl datasets, which employ both two-atom and three-atom interactions to predict the tight-binding Hamiltonian from a set of

atomic positions. We use an active-learning feedback mechanism to systematically generate new density functional theory calculations to

test and improve our fitting. In addition to our previous work on 64 elemental systems and their 2016 binary combinations, we discuss

recent progress in fitting datasets appropriate for systems with three or more different atomic species, which significantly broadens the model

applicability. As initial applications, we study the accuracy of the model for predicting phonons and defect energies.