Regensburg 2022 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 29: Data Driven Materials Science: Design of Functional Materials
MM 29.3: Vortrag
Donnerstag, 8. September 2022, 10:45–11:00, H45
Uncertainty Modelling for Property Prediction of Double Perovskites — •Simon Teshuva1, Mario Boley1, Felix Luong1, Lucas Foppa2, and Matthias Scheffler2 — 1Monash University, Melbourne, Australia — 2Fritz Haber Institute, Berlin, Germany
Statistical predictive models for double perovskite properties are of high interest, because the perovskite structure allows relatively accurate property prediction and at the same time provides enough flexibility to yield a huge number of different materials of which some are likely relevant for important applications. Existing results published for this class of materials typically refer only to the predictive performance as, e.g., measured by the root mean squared error. However, active learning strategies for effective materials screening also rely on adequate uncertainty estimates as provided by probabilistic models.
Here, we study the predictive performance of two popular machine learning models, Gaussian processes and random forests, together with the quality of their uncertainty estimates. This study is based on a dataset of over 800 single (ABO_3) and double (A_2BB'O_6)cubic perovskite oxides with computed bulk modulus, cohesive energy, and bandgap. We show that Gaussian processes, while providing sound Bayesian uncertainty estimates, can have inferior performance when their assumption of isometric smoothness of the target property is not met. In this case, as exemplified by the double perovskite bandgaps, random forests provide a better alternative, despite their rather ad-hoc uncertainty estimates. Improving these estimates thus appears to be a promising direction for future research.