How Architecture and Training Affect TPC Representations Across Experiments
Abstract
Deep-learning efforts have increasingly shifted toward foundation model approaches. In experimental physics, this allows models and learned representations to be reused beyond the experiments in which they were developed. This work evaluates the reusability of representations across experiments and detector systems using probes on frozen encoders. These probes reveal task-relevant structure before downstream adaptation, complementing fine-tuning. Together with random-weight controls, they distinguish contributions from architecture and encoder training that downstream performance alone cannot resolve. Time projection chamber (TPC) data provide a useful testbed because events from TPC systems can be represented as variable-length sparse tensors, while detector geometries, event topologies, and scientific tasks can differ substantially. We investigate whether fixed-dimensional TPC event representations can be reused across classification tasks, experiments, and detector systems. Sparse ResNet and PointNet-style encoders produce 512-dimensional embeddings for four datasets from the GADGET II TPC and AT-TPC. Randomly initialized encoders isolate the contribution from architecture before supervised training. We then train each encoder on a classification task, freeze its parameters, and train a linear or nonlinear probe for each downstream task. We find that this architecture-induced structure remains useful across experiments and detector systems. The randomly initialized PointNet-style representation is highly informative on several tasks. The two architectures organize their embedding spaces differently, but neither exhibits a large, systematic loss of utility cross-detector. These results show that architecture is a major source of task-relevant structure in TPC embeddings and should be treated explicitly when assessing representation learning and developing reusable detector models.
Related
- Temporal Geometry of Deep Networks: Hyperbolic Representations of Training Dynamics for Intrinsic Explainability
- Beyond Compression: Training Latent Representations for Stable Long-Horizon Rollout in Neural Surrogate Solvers
- Learning shared hyperelastic constitutive representation from heterogeneous full-field experiments across multiple materials
- Binder migration across film consolidation and pore emptying in electrode drying: a continuous model validated against drying-rate, thickness, and schedule experiments
- A neural network architecture and training algorithm to predict viscoelastic stresses from vortical data
BibTeX
@misc{wheeler2026architecture,
title = {{How Architecture and Training Affect TPC Representations Across Experiments}},
author = {Tyler Wheeler and Michelle P. Kuchera and Raghuram Ramanujan and William Sieland and Ryan Krupp and Yassid Ayyad and Daniel Bazin and Connor L. Cross and Hoi Yan Ian Heung and Andrew J. Jones and Ruchi Mahajan and Saiprasad Ravishankar and Pranjal Singh and Benjamin Votaw and Chris Wrede},
howpublished = {arXiv},
year = {2026},
month = aug,
doi = {10.48550/arxiv.2608.21756},
eprint = {2608.21756},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2608.21756},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - How Architecture and Training Affect TPC Representations Across Experiments
AU - Tyler Wheeler
AU - Michelle P. Kuchera
AU - Raghuram Ramanujan
AU - William Sieland
AU - Ryan Krupp
AU - Yassid Ayyad
AU - Daniel Bazin
AU - Connor L. Cross
AU - Hoi Yan Ian Heung
AU - Andrew J. Jones
AU - Ruchi Mahajan
AU - Saiprasad Ravishankar
AU - Pranjal Singh
AU - Benjamin Votaw
AU - Chris Wrede
PY - 2026
DA - 2026/08/27
PB - arXiv
DO - 10.48550/arxiv.2608.21756
UR - https://arxiv.org/abs/2608.21756
AB - Deep-learning efforts have increasingly shifted toward foundation model approaches. In experimental physics, this allows models and learned representations to be reused beyond the experiments in which they were developed. This work evaluates the reusability of representations across experiments and detector systems using probes on frozen encoders. These probes reveal task-relevant structure before downstream adaptation, complementing fine-tuning. Together with random-weight controls, they distinguish contributions from architecture and encoder training that downstream performance alone cannot resolve. Time projection chamber (TPC) data provide a useful testbed because events from TPC systems can be represented as variable-length sparse tensors, while detector geometries, event topologies, and scientific tasks can differ substantially. We investigate whether fixed-dimensional TPC event representations can be reused across classification tasks, experiments, and detector systems. Sparse ResNet and PointNet-style encoders produce 512-dimensional embeddings for four datasets from the GADGET II TPC and AT-TPC. Randomly initialized encoders isolate the contribution from architecture before supervised training. We then train each encoder on a classification task, freeze its parameters, and train a linear or nonlinear probe for each downstream task. We find that this architecture-induced structure remains useful across experiments and detector systems. The randomly initialized PointNet-style representation is highly informative on several tasks. The two architectures organize their embedding spaces differently, but neither exhibits a large, systematic loss of utility cross-detector. These results show that architecture is a major source of task-relevant structure in TPC embeddings and should be treated explicitly when assessing representation learning and developing reusable detector models.
ER -