PaperlayerView exact revision on arXiv

Source-linked AI summary

Panda Diplomacy: Foundation Model Pre-training across Particle Imaging Detectors for High Energy and Nuclear Physics

Samuel Young, César Jesús-Valls, Kazuhiro Terao

arXiv:2609.00611v1hep-excs.CV

TL;DR

Existing sensor-level foundation-model approaches are tied to individual detectors, leaving it unknown whether one representation-learning mechanism can span fundamentally different modalities. Panda V2 applies a shared point-cloud self-distillation recipe across LArTPC, collider-TPC, and water-Cherenkov data, achieving competitive multi-task reconstruction with 1,000 labeled events and revealing physically meaningful latent structure.

  • Problem

    Existing sensor-level reconstruction systems are closely tied to individual instruments, and whether one representation-learning mechanism works across fundamentally different detectors remains unknown.

  • Method

    Panda V2 independently pre-trains a shared point-native architecture with EMA-teacher self-distillation on LArTPC, collider-TPC, and water-Cherenkov data, then uses lightweight frozen or LoRA-adapted heads.

  • Results

    Using 1,000 labeled events per task, Panda V2 supports competitive reconstruction across all three detectors and reveals latent directions associated with particle progression, track curvature, and transverse momentum.

  • Takeaways & Limitations

    The results support a common sensor-level pre-training framework for heterogeneous particle imaging detectors within the evaluated settings.

  • Takeaways & Limitations

    The study does not characterize scaling with model size, pre-training data, or compute, and does not test cross-detector or simulation-to-real transfer.

Abstract

from arXiv · show

Foundation models are increasingly being pursued in particle and nuclear physics, but existing approaches remain strongly tied to individual experiments through detector-specific architectures or pre-training objectives, limiting their reuse across sensing modalities. We show that a point cloud self-distillation framework yields a substantially more general sensor-level pre-training recipe. We show that the same refined architecture and objective can be independently pre-trained with minimal changes on three qualitatively different detector modalities: liquid argon time projection chamber (LArTPC), collider TPC, and water Cherenkov. Using 1,000 labeled images for downstream task adaptation, Panda V2 matches or exceeds specialized foundation-model baselines trained with orders of magnitude more supervision, matching state-of-the-art particle-clustering performance with 70x fewer labeled events on sPHENIX while substantially improving particle identification, and on LArTPC data matching Panda (arXiv:2512.01324) particle reconstruction with up to 1,000x fewer labels. Beyond reconstruction, simple linear probes reveal physically meaningful latent structure associated with particle causality and track curvature.

1 Introduction

Panda V2 addresses whether one sensor-level representation-learning mechanism can span fundamentally different particle detectors. It applies a shared pre-training recipe across LArTPC, collider TPC, and water Cherenkov data, then evaluates reconstruction and encoded physical structure with limited supervision.

  • Motivation: Particle images differ radically across LArTPC, collider TPC, and water Cherenkov experiments.These modalities respectively image ionization trajectories, dense curved trajectories in magnetic fields, and projected Cherenkov-photon patterns.
  • Motivation: Existing reconstruction methods are highly detector-specific, whereas foundation models aim to reuse representations learned from abundant unlabeled data across downstream tasks.The open question is whether the same representation-learning mechanism can operate across fundamentally different particle detectors.
  • Approach: Panda V2 independently pre-trains the same architecture and self-distillation objective on three detector modalities, changing only input-scale-related hyperparameters.The objective contains no detector-specific propagation or reconstruction rules and uses generic spatial augmentations.
  • Results: Using 1,000 labeled events per downstream task, lightweight readouts recover point-, particle-, and event-level observables across all three detectors.The representations remain competitive with specialized reconstruction methods trained with substantially more supervision.
  • Results: Simple latent directions are associated with the arrow of time for individual LArTPC particles and transverse momentum in sPHENIX.These findings extend the evaluation beyond reconstruction to physically meaningful structure encoded in the learned representations.

2 Method

Panda V2 uses a shared point-native encoder and observable prototype self-distillation recipe across detector modalities. Lightweight task-specific heads then provide parameter-efficient downstream adaptation for reconstruction.

  • Panda V2 encoder: Each detector image is represented as a sparse set of voxelized 3D points with associated sensor measurements.The shared encoder processes native voxelization, combines convolution and attention across scales, and produces contextual features for each input point.
  • Panda V2 encoder: The four-stage point-native encoder uses Point Transformer V3 and LitePT components across all three modalities.Sparse 3D convolutions operate at high spatial resolution, while local self-attention is used at coarser scales.
  • Self-distillation: Panda V2 self-distillation uses an EMA teacher and student that encode independently augmented views while the student predicts teacher assignments over learned prototypes.Masked measurements are removed, and only observable locations are distilled.
  • Parameter-efficient adaptation: Downstream reconstruction uses lightweight task-specific heads, with the backbone either frozen or adapted using LoRA.The heads include linear classifiers, a two-stage GNN, coordinate–feature cross-correlation, and a DETR-style set-prediction module.

3 Experiments

Panda V2 is independently evaluated on LArTPC, collider-TPC, and water-Cherenkov data using 1,000 labeled events per downstream task. It supports competitive reconstruction and exposes physically meaningful latent directions.

  • Experimental setup: 1,000 labeled events per downstream task are used across PILArNet-M, TPCpp-10M, and WAND.The experiments cover semantic segmentation, particle clustering, and reconstruction tasks across LArTPC, collider-TPC, and SK-like water-Cherenkov data.
  • LArTPC reconstruction: On PILArNet-M, Panda V2 reaches particle-clustering ARI 0.973 versus Panda V1’s 0.914 at 1,000 labels, and semantic-segmentation macro-F1 0.976 versus 0.952.The comparison uses Panda V1’s strongest published frozen-backbone result for each task.
  • Collider-TPC reconstruction: On TPCpp-10M, track-clustering ARI 0.943 closely matches FM4NPP’s 0.945 result obtained with 70,000 labeled events.Particle-identification macro-F1 is 0.838 versus 0.713, while noise-tagging macro-F1 is 0.942 versus 0.912.
  • Water-Cherenkov reconstruction: On WAND, single-particle probes reach electron vertex resolution 98.7 cm, angular resolution 5.75°, and momentum resolution 4.80% using 1,000 labeled events.The single-particle resolutions remain worse than the highly optimized fiTQun and APFit algorithms; multi-ring recalls are 80.4%, 44.4%, and 96.7% for one-, two-, and ≥3-ring events.
  • Physical structure: A single latent direction orders LArTPC points along particle trajectories with Spearman ρ = 0.89, while another encodes TPC track curvature and transverse momentum with ρ = 0.957.These results show physically meaningful structure in frozen representations beyond downstream reconstruction.

4 Conclusion and Limitations

The study shows that a common sensor-level self-supervised framework can span three substantially different particle imaging detectors and support multiple reconstruction tasks with 1,000 labeled events. Linear probes also reveal physically meaningful structure, while transfer and scaling questions remain open.

  • Conclusion: A common sensor-level self-supervised framework supports semantic segmentation, particle clustering, event-property reconstruction, and ring counting across three detector types.The resulting representations can be adapted with only 1,000 labeled events.
  • Conclusion: Simple linear probes reveal particle progression, track curvature, transverse momentum, and forwardness in the learned representations.These structures are observed in LArTPCs, TPCpp-10M, and water-Cherenkov events.
  • Limitations: The study does not characterize scaling with model size, pre-training data, or compute, and does not test cross-detector or simulation-to-real transfer.The authors identify these as unresolved questions for future work.

A.1 Reconstruction in high energy and nuclear physics experiments

Particle-physics reconstruction maps low-level detector measurements to physical observables, but sensor-level foundation models must learn both local hit semantics and global structure. Existing methods often use modality-specific representations or objectives, motivating evaluation of reusable point-cloud self-distillation across detector processes.

  • Reconstruction maps ionization deposits, photodetector hits, or calorimeter signals to particle identities, positions, directions, energies, and momenta.
  • Sensor-level methods operate directly on detector measurements, learning both local semantics of individual hits and global event structure.
  • Most early particle-physics foundation models operate on reconstructed collider objects using masked modeling, re-simulation-based self-supervision, or renormalization-group-motivated invariances.
  • Sensor-level approaches include masked point modeling, contrastive learning, self-supervised IceCube pre-training, and heterogeneous-detector masked reconstruction.
  • Modality-specific methods encode detector priors through trajectory-oriented tokenization, module-aware fusion, collider-coordinate serialization, or propagation-constrained neighbor prediction.
  • Particle-detector point clouds are stochastic, globally sparse, locally structured, and physically annotated at active measurements, motivating tests across distinct sensing processes.

B Dataset definitions

PILArNet-M is an openly released, fully labeled simulated LArTPC dataset containing 1,199,200 events. Its sparse point clouds encode energy deposits and multiple particle-level labels.

  • PILArNet-M contains 1,199,200 simulated, fully labeled LArTPC events spanning multi-particle particle-bomb and single-particle samples.
  • Primary particles are generated with MPV/MPR and propagated through liquid argon using Geant4 within the LArG41 simulation framework.
  • The sparse point clouds include deposited energy and labels for semantic class, particle identity, particle instance, interaction membership, momentum, and vertex.
  • Events are divided into 1,082,400 training, 66,800 validation, and 50,000 test events and publicly distributed in HDF5 format.

B.2 Water Cherenkov: WAND

WAND is a fully simulated water Cherenkov dataset modeled on a Super-Kamiokande-like detector. This work retains 13 configurations and forms a 119,917-event training pool after per-configuration holds and caps.

  • WAND simulates Super-Kamiokande-like cylindrical water Cherenkov events with approximately 11,000 photomultiplier tubes.
  • The simulation includes Geant4 particle propagation, LUCiD optical transport, wavelength-dependent scattering and absorption, and PMT response effects.
  • The dataset includes single-particle electrons, muons, charged pions, and neutral pions, plus multi-particle, multi-ring, pile-up, particle-gun, and GENIE neutrino interactions.
  • 119,917 events remain for training after retaining 13 configurations, holding out 1,000 events per configuration, and capping each configuration at 10,000 training events.

B.3 Collider TPC: TPCpp-10M

TPCpp-10M is an open simulated dataset of minimum-bias proton–proton collisions in the sPHENIX TPC at RHIC. It provides 10 million unlabeled pre-training events alongside labeled downstream data.

  • TPCpp-10M contains simulated minimum-bias proton–proton collisions in the sPHENIX Time Projection Chamber at RHIC.
  • The simulation uses sPHENIX geometry and a measured 1.4 T solenoidal magnetic field, modeling energy loss, scattering, secondary production, and decays.
  • The detector-response chain digitizes TPC ionization signals with gain, noise, shaping, and zero suppression before reconstructing three-dimensional spacepoints.
  • The release provides 10 million unlabeled events for foundation-model pre-training and 70,000 labeled events for track finding, particle identification, and noise tagging.

C Pre-training setup

Panda V2 uses a four-stage encoder architecture, with features from each stage combined into per-point representations for reconstruction. Separate backbones and clustering modules are trained for each detector-data modality.

  • Reconstruction setup: Features from every encoder stage are concatenated to produce one feature for each input point.These per-point features are passed to a clustering module for image reconstruction.
  • Reconstruction setup: A separate backbone and clustering module is trained for each detector-data modality rather than using one shared model across all three.
  • Backbone architecture: The encoder has four stages, using CNN-only processing in the first two and attention-only processing in the latter two.It uses local serialized attention with patch size 1024 and voxel resolutions (1, 2, 2, 8).

C.2 Training setup

Panda V2 uses a shared pre-training setup across detector modalities, then pairs frozen or lightly adapted backbone features with task-specific reconstruction heads and equivariant geometric summaries.

  • Training setup: The pre-training setup changes only grid size, mask size, and feature transforms to match each detector’s numerical and spatial scale.Mask sizes are 40 voxels for LArTPC, 12.5 for water Cherenkov, and 13.3 for sPHENIX.
  • Training setup: 10M image presentations are used for pre-training each detector encoder, with modality-dependent epoch counts of 10, 2, and 100.The counts correspond to PILArNet-M, TPCpp-10M, and WAND, respectively.
  • Equivariant probe: Equivariant probes combine approximately invariant backbone features with spatial coordinates to predict particle vertices and directions under rotations and translations.The coordinate–feature moment H(1) supplies vector-like outputs that transform with the event.
  • Equivariant probe: In water Cherenkov events, Figure 6 compares mean frozen event features with coordinate–feature moments for predicting electron and muon positions and directions.The figure reports 68th-percentile position and angular resolution, where lower is better.
  • Equivariant probe: The learned spatial moments enable vector-like predictions from weakly invariant features by exploiting their organization in three-dimensional space.This provides the geometric mechanism used by the equivariant probe.
  • Clustering: Clustering uses a two-stage coarse-to-fine GNN that first forms local prefragments and then merges them using invariant geometric features.The edge representation retains relative directional information through centroid separation, moment alignments, and pairwise moment correlations.

E.1 LArTPC

On LArTPC data, Panda V2 achieves strong particle clustering and semantic segmentation with only 1,000 labeled events, matching or surpassing Panda V1 with far more supervision. Its lightweight adaptation also supports probes of physically meaningful particle progression.

  • Adaptation setup: Panda V2’s LArTPC adaptation uses LoRA with lightweight task heads rather than a large decoder, reducing the trainable adaptation setup.Particle clustering uses a two-stage GNN with LoRA, while semantic labels use a linear classifier with LoRA.
  • Physical latent structure: The particle progression probe assigns each point a normalized ordered rank, with τ = −1 at the trajectory start and τ = +1 at its end.Performance is measured using within-particle Spearman rank correlation, and τ represents normalized progression rather than physical time or propagation distance.
Loading 2609.00611v1…