Speaker
Description
Deep learning, particularly techniques adapted from modern computer vision, has substantially advanced the analysis capabilities of particle and nuclear physics experiments, neutrino experiments in particular. Most of these applications, however, rely on supervised learning: the model is trained on simulated events with known labels, and therefore inherits the assumptions and imperfections of the simulation. Self-supervised learning (SSL) is an emerging alternative that learns directly from data without labels, and can therefore also capture features of real data that are not yet modeled in simulation. A successfully trained SSL model provides an encoder that converts each event into a compact, general-purpose numerical summary (a "representation") that can serve many, ideally all, tasks on the same dataset. Because of this generality, such a model is called a foundation model.
If the representations capture the information in the data in a simple form, downstream tasks, such as particle type identification, kinematics regression, panoptic segmentation (grouping pixels into individual particles and interactions), and parentage reconstruction (determining which particle produced which), can each be solved by a small model trained on a small dataset. Representations also open the door to new inference techniques. For example, comparing representations of data and simulation can quantify and help reduce their mismatch; in an oscillation experiment, comparing near- and far-detector representations can quantify distributional shifts; and clustering events in representation space may reveal new physics or anomalous detector behavior.
Despite this potential, applying SSL to particle detector data is challenging because of the nature and scale of the data. The data are globally sparse but locally dense: fewer than 1% of pixels are non-zero, yet signals are densely sampled where present. The data are also large: each event is recorded as multiple 2D projections or a single 3D image containing 100,000 to 1,000,000 non-zero pixels. Together, these properties yield very low information density per pixel, unlike the natural images for which SSL methods were developed.
In this talk, I will present the status of sensor-level foundation models, i.e., SSL applied directly to sensor-level detector data. Recent results demonstrate that a single model architecture and training method can be applied effectively across multiple detector technologies, including liquid argon (DUNE, SBN) and high-pressure gaseous xenon (NEXT) time projection chambers (TPCs), water Cherenkov detectors, and the sPHENIX silicon tracker (using its open dataset). Building on this progress toward mastering SSL for high energy and nuclear physics experiments, I will briefly discuss next steps in AI R&D and the development of new physics inference methods that exploit learned representations.