Regensburg 2025 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 3: Data-driven Materials Science: Big Data and Worksflows
MM 3.10: Vortrag
Montag, 17. März 2025, 12:45–13:00, H10
Towards Multi-Fidelity Machine Learning Using Robust Density Functional Tight Binding Models — •Mengnan Cui1,2, Karsten Reuter1, and Johannes T. Margraf2 — 1Fritz-Haber-Institut der MPG, Berlin, Germany — 2University of Bayreuth, Physical Chemistry V: Theory and Machine Learning
Machine learning has revolutionized the atomistic simulation of molecules and materials, offering unparalleled computational speed with high accuracy. However, its performance depends heavily on the quality and quantity of training data, presenting challenges due to the scarcity of high-fidelity datasets (beyond semilocal DFT). This study investigates transfer learning (TL) across multiple fidelities for molecules and solids, examining the role of fidelity levels and configuration/chemical space overlap in pre-training and fine-tuning. This reveals negative transfer driven by noise from low-fidelity methods like DFTB, which can significantly impact fine-tuned models. Despite this, multi-fidelity approaches consistently outperform single-fidelity learning and, in some cases, even surpass TL based on foundation models by leveraging an optimal overlap of pre-training and fine-tuning chemical spaces.
Keywords: DFTB; Transfer Learning; Multi-Fidelity; Machine learning