Particle Classifier
A calibrated classifier for identifying particle types from detector hit patterns, built to flag its own uncertainty rather than force a guess.
Problem
Detector output arrives as sparse, irregular hit patterns rather than clean images — a poor fit for standard convolutional architectures, and a domain where a confident wrong answer is worse than an honest 'unsure.'
Approach
Modeled each event as a graph over detector hits and trained a graph neural network to classify particle type directly from that structure, with an explicit calibration step so output confidence tracks true accuracy rather than raw softmax scores.
Architecture
Graph construction over raw hits → message-passing GNN encoder → pooled event representation → calibrated classification head, exported to ONNX for fast inference inside the existing detector software stack.
Results
- Matched baseline accuracy while running an order of magnitude faster than the prior simulation-based approach.
- Calibration curve stayed within 2% of ideal across the tested energy ranges.
- Deployed as an ONNX module for evaluation inside existing reconstruction pipelines.
Future work
Extending the graph representation to handle detector upgrades with higher hit density, and exploring self-supervised pretraining on unlabeled collision data to reduce dependence on simulated training sets.