Machine Learning for Liquid Argon TPCs: From Task-Specific Networks to Foundation Models

by Matteo Vicenzi (Brookhaven National Laboratory (US))

Monday 5 Oct 2026, 12:45 → 13:45 America/New_York

Liquid argon time projection chambers (LArTPCs) are the central detector technology of the current and next-generation neutrino program in the US. These detectors record neutrino interactions as high-resolution 2D images which are then combined into 3D pictures of each event. Over the past few years, traditional reconstruction frameworks have progressively integrated machine learning approaches for specific tasks. This has also been the case for Wire-Cell, a mature data-first reconstruction framework developed at BNL. These trends have accelerated rapidly. In this talk, I will review some recent applications of machine learning for LArTPCs, including full deep-learning reconstruction chains, surrogate models and differentiable simulations, which open new approaches to calibration and inference. Most of these methods rely on supervised training with simulated data. Foundation models offer a complementary approach, learning general-purpose representations directly from detector data through self-supervised training, which can then be adapted to a wide range of downstream tasks. I will conclude by discussing the ongoing development of a Wire-Cell Foundation Model, including data preparation, model architecture, initial training, and preliminary studies of its learned representations.

20261005_MV_SBU_MLforLArTPCs.pdf