Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 8: Development of Computational Methods: Diverse Topics and Machine Learning
MM 8.6: Vortrag
Montag, 27. März 2023, 17:15–17:30, SCH A 251
Physics-inspired Machine Learning for Predicting Ionization Energies of Electronically Localized Systems — •Ke Chen1,2,3, Christian Kunkel1, Bingqing Cheng3, Karsten Reuter1, and Johannes T. Margraf1 — 1Fritz-Haber-Institut der MPG, Berlin, Germany — 2Technische Universität München, Garching, Germany — 3Institute of Science and Technology, Klosterneuburg, Austria
Machine learning (ML) has been successfully applied to predict many chemical properties, most prominently energies and forces in molecules and materials. The strong interest in predicting energies in particular has lead to a ’local energy’-based paradigm for modern chemical ML models, which ensures size extensivity and linear scaling of computational cost. However, some electronic properties (such as excitation energies or ionization potentials) are not size-extensive and may even be spatially localized. Using extensive models in these cases can lead to large errors. In this work, we explore different strategies for predicting intensive and localized properties, using ionization energies in organic molecules as a test case. In particular, we compare size-intensive aggregation functions and effective, machine-learned Hamiltonians. The physical interpretability and cost/benefit ratios of the approaches will be discussed.