⚛️ nuclear experiments

How Architecture and Training Affect TPC Representations Across Experiments

This paper demonstrates that the architectural design of encoders, rather than just their training, is a primary source of reusable, task-relevant structure in Time Projection Chamber (TPC) event representations, as evidenced by the cross-experiment utility of embeddings from both trained and randomly initialized Sparse ResNet and PointNet models.

Original authors: Tyler Wheeler, Michelle P. Kuchera, Raghuram Ramanujan, William Sieland, Ryan Krupp, Daniel Bazin, Connor L. Cross, Hoi Yan Ian Heung, Andrew J. Jones, Ruchi Mahajan, Saiprasad Ravishankar, Pranjal Si

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tyler Wheeler, Michelle P. Kuchera, Raghuram Ramanujan, William Sieland, Ryan Krupp, Daniel Bazin, Connor L. Cross, Hoi Yan Ian Heung, Andrew J. Jones, Ruchi Mahajan, Saiprasad Ravishankar, Pranjal Singh, Benjamin Votaw, Chris Wrede

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the high-stakes world of nuclear physics, scientists often rely on massive, intricate detectors to catch fleeting glimpses of the universe's most elusive particles. Among the most versatile tools in this arsenal are time projection chambers, or TPCs. Imagine a large, gas-filled tank where a charged particle, like a proton or an electron, flies through and leaves a trail of ionized gas behind it, much like a jet plane leaving a vapor trail in the sky. Sensors on the walls of this tank capture the position and timing of these ionization points, allowing physicists to reconstruct the particle's full three-dimensional path and measure its energy. These devices are essential for studying rare nuclear reactions and exotic isotopes, but they generate a unique kind of data: a sparse, jagged cloud of points that varies in size and shape from one event to the next.

For decades, analyzing this data required building a custom computer model for every single experiment. If a scientist wanted to study a new type of particle or used a slightly different detector, they had to start from scratch, training a new artificial intelligence from the ground up. This process is slow, expensive, and inefficient. Recently, the field has begun to explore "foundation models," a type of artificial intelligence designed to learn general patterns from vast amounts of data so that the knowledge can be reused for many different tasks. The big question for physicists is whether a model trained on one specific detector can actually understand the data from a completely different detector, or if the unique geometry of each machine makes the models too specialized to be useful elsewhere.

A team of researchers set out to answer this question by testing how well different types of AI architectures could transfer their knowledge across distinct experimental setups. They focused on two very different time projection chambers: the GADGET II, a compact device designed to study the decay of radioactive ions, and the Active-Target TPC, a much larger cylinder used to observe nuclear reactions in reverse. These two machines differ drastically in size, the gas they use, and the types of particle events they record. The researchers trained two different styles of neural networks on data from one of these machines and then froze the models, preventing them from learning anything new. They then asked these frozen models to help solve classification tasks on data from the other machine, or even on different tasks within the same machine.

The study compared two distinct approaches to building these AI models. The first was a sparse ResNet, a deep network with over twenty million adjustable settings, designed to look for complex patterns in the data. The second was a PointNet-style model, a much simpler architecture with fewer than ninety thousand settings that treats each point in the data cloud independently before combining them. To their surprise, the researchers found that the architecture itself—the basic mathematical structure of the network—was responsible for a huge portion of the model's ability to understand the data. Even before the models were trained on any specific physics task, the PointNet-style network, with its random, untrained weights, was already able to distinguish between different types of particle events with remarkable accuracy.

When the researchers trained these models on a simple task, such as telling the difference between two types of particles, and then tested them on completely different tasks or on data from the other detector, the results were consistent. The models did not lose their usefulness when the experiment changed. The PointNet-style models, in particular, proved to be incredibly robust; their initial, untrained structure was already so good at organizing the data that training them further provided only a small boost in performance. The more complex ResNet models benefited more from training, but they also retained the ability to generalize across the different detectors. Crucially, the study showed that the two architectures organized the data in fundamentally different ways. The ResNet tended to cluster events into tight, separate groups, while the PointNet created a broader, more continuous map of the data. Despite these structural differences, both approaches successfully transferred their knowledge across the gap between the two very different detectors.

These findings suggest that the choice of the AI's underlying structure is just as important as the training data itself. The researchers demonstrated that a significant amount of task-relevant information is built into the architecture of the network before it ever sees a single labeled example. This means that scientists do not necessarily need to retrain massive models from scratch every time they switch to a new detector or a new physics question. Instead, they can rely on the inherent structure of well-designed networks to provide a strong foundation that works across different experimental conditions. By freezing the trained models and testing them on new tasks, the team showed that these representations are reusable and robust, offering a path toward more efficient and flexible data analysis in nuclear physics. The work indicates that the "intelligence" of these models comes not just from what they learn, but from how they are built, a discovery that could reshape how physicists approach the design of future experiments.

Technical Summary: How Architecture and Training Affect TPC Representations Across Experiments

Problem Statement
In experimental nuclear physics, Time Projection Chambers (TPCs) generate rich, sparse, variable-length point cloud data representing charged-particle trajectories. Traditionally, machine learning (ML) models for TPC data analysis are developed in a bespoke fashion for specific experiments, detector geometries, and physics tasks. This approach is inefficient and limits the reusability of learned representations. While the field of ML has shifted toward foundation models that learn generalizable representations for downstream tasks, it remains unclear to what extent representations learned in one TPC context (specific detector, specific experiment, specific task) transfer to others with different geometries, event topologies, and scientific goals. Furthermore, the relative contributions of the neural network architecture versus supervised training to the formation of these task-relevant representations are not well understood.

Methodology
The authors evaluate the reusability of TPC event representations across two distinct detector systems (GADGET II and Active-Target TPC) and four experimental datasets involving different isotopes and physics processes. The study employs a "frozen encoder" approach to isolate the information available in the latent space before any downstream adaptation.

Data and Tasks: The study utilizes data from the Facility for Rare Isotope Beams (FRIB).
- GADGET II: Used for $\beta$ -delayed charged-particle spectroscopy (implant-decay mode). Datasets include $^{20}\text{Mg}$ (binary and 3-class particle ID) and $^{21}\text{Mg}$ (3-class).
- AT-TPC: Used for inverse-kinematics nuclear reactions (active-target mode) in a solenoid magnetic field. Datasets include $^{16}\text{O} + \alpha$ (track multiplicity classification) and $^{10}\text{B} + d$ (track multiplicity).
- Inputs are variable-length sparse tensors of $(x, y, z, q)$ coordinates and charge.
Architectures: Two encoder families were implemented using the Minkowski Engine to handle sparse tensors natively:
- Sparse ResNet14: A shallow residual network with ~21.4 million parameters.
- PointNet-style: A shared point-wise MLP with global max pooling, containing ~83.8 thousand parameters.
- Both produce 512-dimensional embeddings.
Experimental Design:
- Source Training: Encoders were trained on a simple "source" classification task (e.g., binary particle ID or binary track multiplicity) for each dataset.
- Random Controls: Identical architectures were evaluated with randomly initialized weights to isolate the contribution of architectural inductive bias from supervised training.
- Frozen Probes: After source training, encoder weights were frozen. Linear (SVM) and non-linear (RBF kernel SVM) probes were trained on the frozen 512-dimensional embeddings to perform "downstream" tasks. These downstream tasks included:
  - Within-experiment: Same detector, different task complexity.
  - Cross-experiment: Same detector, different isotope/experiment.
  - Cross-detector: Different detector system entirely.
Evaluation: Performance was measured via accuracy and macro-averaged F1 scores on held-out test sets. Principal Component Analysis (PCA) was used to visualize the geometry of the latent spaces.

Key Results

Architecture as a Primary Source of Structure: Randomly initialized encoders (without any supervised training) produced embeddings that substantially outperformed basic baselines. For the PointNet architecture, the random representation was often highly informative, achieving near-ceiling performance on several tasks (e.g., 99.7% accuracy on the $^{20}\text{Mg}$ 3-class task).
Transferability Across Systems: The task-relevant structure induced by the architecture persisted across different experiments and even between the two distinct detector systems (GADGET II and AT-TPC). There was no large, systematic loss of utility when transferring representations between detectors.
Impact of Training: Supervised encoder training provided additional gains, but the magnitude varied by architecture. ResNet benefited more substantially from training than PointNet, which often started from a strong random baseline.
Linear vs. Non-linear Structure:
- PointNet representations were often already linearly separable, showing limited gains from non-linear probes.
- ResNet representations often required non-linear probes (RBF) to fully access the class structure, particularly in cross-detector settings.
Latent Geometry Differences: PCA visualizations revealed that while both architectures achieved similar downstream performance, they organized the embedding space differently. Random ResNet embeddings formed compact, separated clusters, while PointNet embeddings formed broader, continuous distributions. Training reorganized these spaces but the fundamental geometric differences between the architectures remained.

Significance and Claims
The paper argues that the architecture itself is a major source of task-relevant structure in TPC embeddings. The findings suggest that:

Reusability is Feasible: Fixed-dimensional TPC event representations can be reused across classification tasks, experiments, and detector systems without retraining the encoder, provided the architecture is suitable.
Architecture Choice Matters: The choice of architecture (e.g., ResNet vs. PointNet) dictates the initial organization of the latent space and the amount of information available before training. This structural bias is as critical as the training data itself.
Diagnostic Value of Frozen Probes: Using frozen probes with random-weight controls provides a necessary diagnostic tool. It allows researchers to distinguish between structure induced by the architecture and structure learned during training, a distinction that end-to-end fine-tuning results alone cannot resolve.

The authors conclude that when developing reusable detector models, the architectural contribution to representation learning must be treated explicitly, and frozen-probe diagnostics should be reported alongside fine-tuning results to provide a complete evaluation of scientific representations.

More like this