Researchers trained a sparse vision model to identify particle collisions at CERN

Physicists at CERN are using a new type of artificial intelligence to better understand the behavior of neutrinos. These tiny particles travel at extreme energies, creating messy and overlapping signatures when they hit detectors. Interpreting this data has traditionally been a slow and difficult task for manual analysis.

To solve this, the team developed a sparse Vision Transformer, a model designed to focus on relevant spatial data. They trained it using self-supervised learning. This allows the model to learn by reconstructing masked segments of its own input data rather than relying entirely on human-labeled examples.

This training method included tasks like identifying the hierarchy of interactions and distinguishing between different types of particle deposits. By mastering these patterns, the model became highly effective at classifying flavors of neutrinos and reconstructing their momentum.

Performance tests showed that the model works well even with very few labeled examples. When given only 1,000 labeled events, it matched the accuracy of older models trained on ten times that amount of data. It also demonstrated the ability to transfer its knowledge between different types of detectors.

"The encoder is forced to infer non-local spatial correlations rather than copying visible voxels," noted the research team during their assessment of the model's diagnostic capabilities.

While the results are promising for particle physics, there are still limitations. This is not yet a universal foundation model for all detectors. The system has not been calibrated with real-world experimental data, and its performance remains dependent on the specific simulation generators used during training.