Scaling and Cross-Detector Transfer of Panda V2 Representations

Characterize how Panda V2 sensor-level self-supervised representations scale with model size, pre-training data, and computational budget, and determine whether representations learned on one particle detector transfer to another detector or from simulated data to real data.

Background

Panda V2 is independently pre-trained with the same point-cloud encoder and self-distillation objective on liquid argon time projection chamber, collider TPC, and water Cherenkov data. The study evaluates downstream reconstruction using fixed-scale models and simulated datasets, but does not establish how performance changes with model capacity, pre-training-set size, or compute, nor whether a representation trained for one detector modality can be reused for another. The paper also leaves unresolved whether the learned representations generalize from simulation to experimental data, an important issue for deployment in real high-energy and nuclear-physics experiments.