Quantum-mechanical (QM) simulations, thanks to their predictive power, can provide significant insights into the nature and dynamics of defects such as vacancies, dislocations and grain boundaries. These considerations are essential in the context of the development of reliable, inexpensive and environmentally friendly alloys. However, despite significant progress in computer performance, QM simulations of defects are still extremely time-consuming with ab-initio/non-parametric methods. The two-centre Slater–Koster (SK) tight-binding (TB) models can achieve significant computational efficiency and provide an interpretable picture of the electronic structure. In some cases, this makes TB a compelling alternative to models based on abstraction of the electronic structure, such as the embedded atom model. The biggest challenge in the implementation of the SK method is the estimation of the optimal and transferable parameters that are used to construct the Hamiltonian matrix. In this paper, we will present results of the development of a data-driven framework, following the classical approach of adjusting parameters in order to recreate properties that can be measured or estimated using ab-initio or non-parametric methods. Distinct features include incorporation of data from QSGW (quasi-particle self-consistent GW approximation) calculations, as well as consideration of higher-order elastic constants. Furthermore, we provide a description of the optimisation procedure, omitted in many publications, including the design stage. We also apply modern optimisation techniques that allow us to minimise constraints on the parameter space. In summary, this paper introduces some methodological improvements to the semi-empirical approach while addressing associated challenges and advantages.
Read full abstract