Berlin 2024 – wissenschaftliches Programm

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

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 2: Machine Learning & Physics

AKPIK 2.5: Vortrag

Mittwoch, 20. März 2024, 16:00–16:15, MAR 0.002

Humans in the loop for more trustworthy Bayesian optimization of materials — •Armi Tiihonen¹, Louis Filstroff², Petrus Mikkola¹, Emma Lehto¹, Samuel Kaski^1,3, Milica Todorović⁴, and Patrick Rinke¹ — ¹Aalto University, Espoo, Finland — ²ENSAI, CREST, Rennes, France — ³The University of Manchester, Manchester, United Kingdom — ⁴University of Turku, Turku, Finland

Bayesian optimization (BO) is a machine learning method for global optimization of black-box functions, e.g. the composition of perovskite materials for more stable solar cells. BO can be coupled to automated sample preparation and characterization pipelines, which introduces the challenge of ensuring sufficient sample quality during the optimization. Low quality samples are hard to detect automatically, but could obscure the optimization process. To make automated materials optimizations more robust and trustworthy, we add humans into the BO loop (HITL) as an additional data source. We implemented three HITL schemes, two based on data fusion and one on multifidelity BO. They query human opinion on sample quality only when necessary and guide the sampling away from composition regions with a lot of low-quality samples. We tested them in simulations based on previously obtained experimental perovskite data. Our data fusion HITL BO queries from humans on average 7% of the samples when the run is repeated 25 times. This leads to only 2% of low quality samples, in contrast to the 25% with the reference method without humans. Thus, HITL ensures more consistent sample quality during BO.

Keywords: machine learning; Bayesian optimization; materials; sample quality