Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

CPP: Fachverband Chemische Physik und Polymerphysik

CPP 7: Emerging Topics in Chemical and Polymer Physics, New Instruments and Methods I

CPP 7.3: Vortrag

Montag, 9. März 2026, 12:00–12:15, ZEU/0255

Optimizing EquiDTB potentials for large-scale molecular simulations — •Zekiye Erarslan, Gianaurelio Cuniberti, and Leonardo Medrano Sandonas — Chair of Materials Science and Nanotechnology, TUD Dresden University of Technology, 01062 Dresden, Germany

Density Functional Tight-Binding (DFTB) is a semi-empirical method that enables efficient large-scale simulations at moderate computational cost compared to first-principles quantum-mechanical methods. However, its generalizability is limited by the use of parameterized pairwise repulsive potentials. The recently developed EquiDTB framework [chemRxiv, 10.26434/chemrxiv-2025-z3mhh] has shown that replacing this repulsive term with a many-body ΔTB potential---trained using equivariant neural networks---significantly improves both the accuracy and transferability of the DFTB method. As a result, EquiDTB achieves hybrid DFT-PBE0 level accuracy across diverse electronic, structural, and vibrational properties of large molecules and molecular dimers containing C, N, O, and H atoms. In this work, we extend the EquiDTB framework by training a more general ΔTB potential on the chemical space covered by the newly generated QCML dataset. This expansion broadens EquiDTB beyond its original four-element scope and enables accurate simulations of neutral molecular systems containing C, N, O, H, P, S, Na, and Cl. We validate the performance of the optimized EquiDTB model through extensive calculations on large molecular dimers, RNA systems, and organic periodic materials.

Keywords: Quantum-mechanics; Machine learning; Molecular simulations; Computational modeling