Dresden 2026 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MM: Fachverband Metall- und Materialphysik

MM 41: Development of Calculation Methods II

MM 41.7: Talk

Friday, March 13, 2026, 12:00–12:15, SCH/A251

Machine-Learning Augmented Electronic Structure Calculations via Adaptive DFTB Parameters — •Yihua Song, Christoph Scheurer, Karsten Reuter, and Chiara Panosetti — Fritz Haber Institut der MPG

As a semi-empirical DFT approximation, Density Functional Tight Binding (DFTB) combines reduced computational cost with accessible electronic structure, providing a compromise between the full physicality of ab initio methods and the speed of machine-learning interatomic potentials. Yet, its accuracy in realistic systems remains limited due to intrinsic approximations and dependence on parametrization quality. To this end, we propose that the precalculated Slater-Koster (SK) electronic interaction tables in DFTB should be adapted to local atomic environments and demonstrate proof-of-principle studies showing the improved electronic structure description. Building on the observed smoothness of SK integrals across chemical and spatial environments, we use machine learning based on atomic descriptors to continuously adapt the parametrization beyond discrete atomic types. As a prototype, we introduce the DFTB Orbital Verity Engine (DOVE), a machine-learning framework that learns subtle discrepancies in electronic parameters across diverse local environments from reference electronic structures. We successfully validate it against complicated systems with multiple coordination states. This machine-learning framework for environment-adaptive tight-binding parameterization enables scalable electronic structure simulations with enhanced fidelity, efficient speed, and physical interpretability.

Keywords: Machine Learning; Tight-binding parameterizations; Electronic Structure; Density Functional Theory