Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 19: Poster Session
MM 19.9: Poster
Tuesday, March 10, 2026, 18:00–20:00, P5
Acquisition strategies in multi-objective Bayesian optimization — •Tatu Linnala1,2, Matthias Stosiek1, Joakim Löfgren2, and Patrick Rinke1,2 — 1Department of Physics, Technical University of Munich, Garching, Germany — 2Department of Applied Physics, Aalto University, Espoo, Finland
Bayesian optimization (BO) is a machine learning technique for optimizing expensive black-box functions, and it is increasingly used in computational and experimental materials optimization. Many BO applications involve multiple competing objectives, requiring multi-objective BO (MOBO) to approximate the Pareto front. We extended the Bayesian Optimization Structure Search (BOSS) code to support advanced MOBO methods, focusing on acquisition strategies. Specifically, we implemented three variants of the expected hypervolume improvement acquisition function: an exact form for bi-objective problems and Monte Carlo approximations for higher dimensions. Additionally, we included scalarization-based methods for greater computational efficiency. These methods were benchmarked on six test cases, including synthetic functions and a real-world lignin extraction problem. Results show that hypervolume-based methods yield the most accurate predictions at high computational cost, although scalarization methods may sometimes be sufficient. This highlights the trade-off between accuracy and computational cost, and the application specificity of the optimal strategy. We provide guidelines for selecting appropriate MOBO settings, and the extended BOSS code provides a flexible toolkit for multi-objective optimization.
Keywords: machine learning; Bayesian optimization; multi-objective optimization; acquisition functions