May 25 – 29, 2026
Chulalongkorn University
Asia/Bangkok timezone

Semi-Supervised and Manifold Learning on Multi-Modal Event Representations in the LZ Dark Matter Experiment

May 27, 2026, 3:15 PM
1h
Sala Phra Kieo, Second Floor

Sala Phra Kieo, Second Floor

Board: 331-Th
Poster Presentation Track 9 - Analysis software and workflows Poster

Speaker

John Winnicki

Description

LUX-ZEPLIN (LZ) searches for dark matter using a dual-phase xenon time projection chamber. Unrelated scintillation and ionization signals can be paired during reconstruction, producing a background that resembles low-energy nuclear recoils. We study the shapes of secondary electroluminescence (S2) pulses to determine whether they can help distinguish this background. We combine wavelet compression with manifold learning and density-based clustering, using tagged calibration neutrons as a reference for the expected recoil response. The unsupervised separation weakens as the pulse-area range is restricted to the low-energy search region. Supervision with neutron-calibration labels improves signal acceptance. Adding labels for an identifiable subset of background events in a semi-supervised analysis increases their rejection at the cost of signal acceptance. A scan of quality cuts shows that many of the separated events already fail S2 requirements, particularly those on transverse-position reconstruction. This is observed even though the embedding uses only the summed temporal waveform. The approach may be useful for investigating backgrounds whose distinguishing pulse characteristics can guide the development of effective data selections.

Author

Co-authors

Presentation materials

There are no materials yet.