Unsupervised particle separation in a LArTPC via interpretable latent representations

Liquid argon time projection chambers (LArTPCs) are widely used in neutrino physics due to their ability to image charged-particle trajectories at high resolution. In LArTPC data, particles can be distinguished by their characteristic energy deposition profiles (calorimetry) and spatial patterns (topology). We ask whether an unsupervised model can discover and structure these differences directly from real detector data. We train a convolutional $\beta$-VAE to reconstruct $9{,}419$ events from the LArIAT experiment and find that both calorimetry and topology are encoded in its eight-dimensional latent space. We then apply a Gaussian mixture model to the latent space to separate events into clusters corresponding to different particle species. Comparing these cluster assignments against particle-ID labels derived from auxiliary-detector information, we achieve a beamline-tag purity above 75\% across all particle species. This first application of $\beta$-VAEs to real LArTPC data demonstrates that physically meaningful particle representations can be learned directly from detector data without particle labels in the loss function.