Every collision becomes a list of numbers
An event is recorded as roughly a hundred numbers: the jets, electrons, muons and photons it contains, plus the missing transverse energy. Simulated events from a handful of processes, Standard Model ones (A, B) and beyond-the-Standard-Model candidates (C, D), together with the Standard Model background that dominates real data, all pile on top of each other in this raw space.
Contrastive learning organizes the space
An encoder is trained with a supervised contrastive loss on the labelled processes: pairs of events from the same process are pulled together, events from different processes are pushed apart. The mixture separates into clusters, and the background, which was never labelled, settles into a region of its own. The separation is not perfect: every process has events that look like background and stay buried in it. The embedding is high-dimensional; three directions are shown here only to make it visible.
Train an anomaly detector on background only
The embedding vector of each event is the input to a small autoencoder: it is squeezed through a bottleneck and reconstructed. Trained on background events alone, epoch by epoch the reconstruction of background improves, and nothing else is ever shown to it. When an event from somewhere else in the embedding is pushed through, the reconstruction comes out wrong, and that error is the anomaly score. The two stages are coupled only through the embedding, so either can be swapped independently.
Score the data and select the anomalous sample
Real data arrives: mostly background, with the Standard Model processes A and B present at their expected small rates, and an admixture of BSM process C. Background events reconstruct well and score low. Events that land in the A, B or C regions reconstruct poorly and score high. A threshold on the score defines the anomalous sample. Raise it and fewer events pass, but background falls off faster than signal.
Interpret the anomaly with a template fit
The anomalous sample is histogrammed along the embedding coordinates and fit with templates for the background and for every labelled process, one Poisson term per bin. At a loose threshold the signal is a faint tail on a dominant background. Tighten it and the background is suppressed while the signal bumps stand out. The fit then says how much of the sample is ordinary background, how much is known Standard Model processes (A and B: nothing new), and how much is BSM process C, with a yield, an uncertainty and a significance. A signal outside the library would be described by the processes it most resembles.
Interpretable by construction
A single anomaly score can only say that something is odd. The embedding geometry carries higher-dimensional physics information, and the fit is structurally the same binned template fit used in conventional searches, so an ORCA result can be reinterpreted without any familiarity with the machine learning underneath.