Embedded Representations of TPC Data

Time Projection Chambers are a class of detector systems that produce 4π reconstruction of charged particle trajectories. TPCs typically generate large amounts of data, making them ripe for deep learning applications. These deep learning models embed raw data into a vector representation, which can then be used for tasks such as reaction channel classification, kinematic predictions, and noise removal.

This work investigates the structure of latent space representations learned by a foundation model trained on data collected from the ¹⁶O+α experiment using the Active-Target Time Projection Chamber (AT-TPC) at FRIB. Specifically, we examine the internal feature space of a foundation model architecture called TPCNet, which was pre-trained in a self-supervised manner on an event reconstruction task. The goal is to reveal physically meaningful structures, such as the number of particle trajectories or geometric clustering of events, within this latent space. We developed a flexible framework to analyze and visualize these latent representations using both supervised and unsupervised methods.

We used dimensionality reduction techniques including t-SNE, UMAP, and PCA to visualize how different event types cluster in the latent space. We also used linear probing to study the utility of the learned representations; a 0.84 F1 score on a downstream track-counting task suggests that TPCNet’s latent space encodes useful geometric and physics knowledge.