Panda: Self-distillation of Reusable Sensor-level Representations for High Energy Physics

Panda teaser figure showing the pipeline from raw LArTPC data through self-supervised learning to downstream tasks

Panda overview. Raw charge depositions corresponding to particle trajectories recorded by a time projection chamber (TPC) (top left) are passed through a point-native hierarchical encoder pre-trained via self-distillation to produce a shared embedding (top right). The same pretrained features are used for three downstream tasks with lightweight heads (bottom): semantic segmentation; particle-level clustering; and interaction-level clustering that groups causally related particles.

Abstract

Liquid argon time projection chambers (LArTPCs) provide dense, high-fidelity 3D measurements of particle interactions and underpin current and future neutrino and rare-event experiments. Physics reconstruction typically relies on complex detector-specific pipelines that use tens of hand-engineered pattern recognition algorithms or cascades of task-specific neural networks that require extensive, labeled simulation.

We introduce Panda, a model that learns reusable sensor-level representations directly from raw unlabeled LArTPC data. Panda couples a hierarchical sparse 3D encoder with a multi-view, prototype-based self-distillation objective. On a simulated dataset, Panda substantially improves label efficiency and reconstruction quality, beating the previous state-of-the-art semantic segmentation model with 1,000× fewer labels. We also show that a single set-prediction head 1/20th the size of the backbone with no physical priors trained on frozen outputs from Panda can result in particle identification that is comparable with state-of-the-art reconstruction tools.

Dimensionality reduction visualization of learned embeddings showing particle type clustering — t-SNE visualization of per-point embeddings taken from the output of the backbone from 1,000 images.

Learned Representations

The outputs of the backbone, cast to 2D using t-SNE, capture both inter-class diversity and intra-class multi-modality. For example, electrons manifest as showers, Michel electrons, or Delta rays, and the model learns to separate these naturally.

Some overlap between photon/electron and muon/pion clusters reflects genuine physical ambiguities in LArTPC data. For example, photon- and electron-initiated electromagnetic showers can be indistinguishable should there be no resolvable conversion gap and unreliable energy deposition patterns.

Semantic segmentation mean F1 vs number of labeled events — Semantic segmentation mean F₁ as a function of images used for pre-training and fine-tuning.

Label Efficiency

Full fine-tuning on 1,000 events matches the previous SOTA trained on 1M labels. Pre-training greatly outperforms the previous SSL method, PoLAr-MAE, by a large margin at all label counts.

Citation

@misc{young2025pandaselfdistillationreusablesensorlevel,
    title={Panda: Self-distillation of Reusable Sensor-level Representations for High Energy Physics}, 
    author={Samuel Young and Kazuhiro Terao},
    year={2025},
    eprint={2512.01324},
    archivePrefix={arXiv},
    primaryClass={hep-ex},
    url={https://arxiv.org/abs/2512.01324}, 
}

Panda
Self-distillation of Reusable Sensor-level Representations for High Energy Physics

Abstract

Learned Representations

Label Efficiency

Interactive Examples

Paper

Citation