Building the Next-Gen ML/AI Toolkit for LArTPC Experiments

Liquid argon time projection chambers (LArTPCs) deliver high-resolution images of neutrino interactions and are the chosen technology for several current and upcoming experiments, including the Short-Baseline Neutrino program and DUNE. Reconstruction in these detectors has traditionally relied on sophisticated pipelines of hand-engineered pattern-recognition algorithms, often augmented with task-specific neural networks. Across the field, there is now growing interest in approaches that integrate deep learning more deeply into the chain and learn directly from data with less per-task engineering.

In this talk I will survey three machine-learning developments along this direction. First, the LANTERN reconstruction chain, which integrates deep CNNs alongside hand-engineered algorithms, has been demonstrated on real MicroBooNE data with CC νₑ and νμ inclusive selections competitive with the experiment's Wire-Cell and Pandora frameworks — and is well positioned to play a complementary role in analyses that draw on the strengths of all three. Second, recent self-supervised work shows that a single point-native encoder, pretrained on raw 3D charge clouds, can be reused across semantic segmentation, particle identification, and interaction-level partitioning — matching or beating prior state-of-the-art with up to 1000× fewer labels. Third, score-based diffusion models can, for the first time, generate LArTPC-like images of comparable quality to Geant4, opening a path toward fast surrogate simulation, conditional event generation, and ML-based unfolding. I will close with the role of differentiable detector simulation in tying these components together, and outline how this toolkit will be applied to the remaining MicroBooNE data and to the SBN and DUNE programs.

Event resources