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

T: Fachverband Teilchenphysik

T 7: Data, AI, Computing, Electronics I

T 7.4: Vortrag

Montag, 16. März 2026, 17:00–17:15, KH 00.024

Self-Supervised Pretraining of HPGe Waveforms for Pulse-Shape Discrimination for LEGEND — •Niko Lay¹, Christoph Vogl¹, Tommaso Comellato¹, Konstantin Gusev¹, Brennan Hackett², Baran Hashemi¹, Lukas Heinrich¹, Patrick Krause¹, Andreas Leonhardt¹, Béla Majorovits², Moritz Neuberger¹, Nadezda Rumyantseva¹, Mario Schwarz¹, Michael Willers¹, and Stefan Schönert¹ — ¹Technical University of Munich, Garching, Germany — ²Max Planck Institute for Physics, Garching, Germany

LEGEND searches for (0νββ)-decay using HPGe detectors enriched in ⁷⁶Ge and operated in instrumented LAr. For LEGEND-1000, underground-sourced Ar, depleted in ⁴²Ar, constitutes the baseline choice. In the event that such argon is unavailable, a well-established and experimentally validated mitigation strategy is under preparation. Using HPGe detectors operated in ⁴²Ar-enriched LAr at the SCARF test facility at TUM, we compare three self-supervised pretraining objectives: a transformer-based autoencoder, an autoregressive objective, and masked contrastive modeling. We finetune the pretrained models to classify signal-proxy vs. background, and bulk vs. surface interaction. The resulting pretrained backbones provide a basis for future likelihood-amortization and simulation-based inference workflows, while this talk focuses on their impact on these two PSD tasks. We acknowledge support from the DFG under Germany’s Excellence Strategy – EXC 2094 (ORIGINS), through the Sonderforschungsbereich SFB 1258, and by TUM MDSI Seed Funds.

Keywords: Self-supervised Learning; Foundation Model; Finetuning; Likelihood-Ratio Estimation; Pretraining