From Reliable to Understandable - How to gain trust in ML for physics

by Prof. Ramon Winterhalder (University of Milan)

Tuesday Oct 14, 2025, 1:00 PM → 2:00 PM Europe/Brussels

E/3rd floor-E.349 - Seminar room (E.349) (Marc de Hemptinne (chemin du Cyclotron, 2, Louvain-la-Neuve))

Machine learning (ML) has become an indispensable tool across the high-energy physics (HEP) landscape. However, for precision collider physics, its value cannot be judged solely by raw performance. Two key questions remain: can we trust its predictions, and do we understand what it has learned? In this talk, I will present recent advances addressing both aspects. First, I will discuss how calibrated uncertainties for amplitude surrogates can be learned using Bayesian neural networks and ensembles, providing robust estimates that disentangle statistical from systematic effects. Second, I will turn to interpretability, showing how modern explainable-AI techniques uncover what quark–gluon taggers learn and how this connects back to QCD. Together, these results highlight the importance of reliability and interpretability as guiding principles for trustworthy ML in HEP.