Contrastive Learning for Interpretable Anomaly Detection at Collider Experiments

Haoyi Jia; Sagar Addepalli; Julia Gonski

Source-linked AI summary

Contrastive Learning for Interpretable Anomaly Detection at Collider Experiments

Haoyi Jia, Sagar Addepalli, Julia Gonski

arXiv:2608.13652v1Aug 13, 2026cs.LGhep-exhep-ph

TL;DR

Collider anomaly detection faces difficult-to-interpret scores and representations affected by non-physics artifacts. ORCA structures a physics-informed embedding with supervised contrastive learning before applying an autoencoder, improving sensitivity and enabling likelihood-based interpretation through known processes.

Problem
Collider anomaly detection lacks well-constrained representations that simplify density estimation, disperse score-bloating artifacts, and expose subtle high-dimensional correlations.
Method
ORCA first learns a physics-process-based latent embedding with supervised contrastive learning, then uses it as input to a background-trained autoencoder for anomaly scoring.
Results
ORCA outperforms a baseline autoencoder across AUC and fixed-background-rejection metrics for a wide variety of Standard Model processes, while embedding fits outperform one-dimensional fits in signal injections.
Takeaways & Limitations
Likelihood fitting in the clustered embedding categorizes unknown events into template processes, providing an interpretable route to anomaly-detection searches at colliders.
Takeaways & Limitations
The template library cannot span all possible new-physics signals, so fitted yields project anomalous data onto the available reference processes.

Abstract

from arXiv · show

Generic event-level anomaly detection for collider physics has two recurring problems: anomaly scores are hard to interpret, and they correlate strongly with energy scale and object multiplicity. We present Organized Representation via Contrastive learning for Anomaly detection (ORCA), a two-stage framework that first learns an embedding space via supervised contrastive learning across a diverse set of physics processes, then runs a standard autoencoder in that space to generate event-level anomaly scores. On a simulated dataset consistent with conditions at the High-Luminosity Large Hadron Collider, ORCA delivers significant gains in both breadth and depth of sensitivity to new physics signals with respect to a baseline autoencoder architecture. Beyond improved sensitivity, the contrastive embedding makes the anomalous sample interpretable: because known processes occupy distinct regions of the space, a maximum-likelihood template fit to the embedding distributions can attribute events in an anomalous sample to template physics processes with quantified uncertainties. We demonstrate that the fit accurately recovers injected signal yields, including for signals excluded from the training of the embedding, and characterizes signals absent from the template library through the known processes they most resemble. These results establish ORCA as a route to interpretable anomaly detection-based searches at colliders, where the embedding geometry carries higher dimensional physics information compared to standard one-dimensional output fits, enhancing downstream statistical analysis.

I. INTRODUCTION

Collider anomaly detection offers model-agnostic searches with strong background suppression, but traditional implementations face limitations in learning capacity and interpretation. ORCA addresses both through a supervised contrastive, physics-process-based embedding followed by an autoencoder, enabling more accessible interpretation of anomaly-detection results.

Motivation: Anomaly detection searches for beyond-the-Standard-Model physics with signal-model agnosticism and strong background suppression.This broadens the range of possible BSM phenomena entering the measurement phase space.
Limitations: Traditional autoencoders trained directly on low-level inputs can memorize recurring entries instead of learning higher-order correlations.Their reconstruction objective makes them sensitive to the characteristics of the training population.
Limitations: Anomaly-score interpretation is typically limited to clustering and correlations with raw kinematic inputs, while Monte Carlo-based interpretations risk reflecting mismodeling.Existing signal extraction can also be performed in anomalous phase spaces where an additional interpretation challenge remains.
Approach: Contrastive learning maps similar events close together and dissimilar events apart in a fixed-dimensional vector space, addressing traditional anomaly-detection limitations.The representation is trained to encourage alignment and uniformity.
ORCA: ORCA first creates a supervised, physics-process-based latent embedding from object-level kinematics, then uses that embedding as input to a second-stage autoencoder.The method is designed so interpreting results requires no additional tooling or familiarity with the underlying machine-learning method.

II. METHODOLOGY · A. Samples

ORCA is trained and evaluated on a simulated proton-proton collision dataset representing High-Luminosity LHC conditions. Events are encoded as fixed-size 110-dimensional vectors from selected reconstructed objects and missing transverse energy.

A. Samples: ORCA is trained and evaluated on events from the simulated COLLIDE-2V LHC collision dataset.The dataset contains approximately 750 million proton-proton collisions.
A. Samples: 750 million proton-proton collisions are simulated at √s = 13.6 TeV.This is the approximate dataset size and collision energy reported for COLLIDE-2V.
A. Samples: The dataset includes over 50 distinct Standard Model processes generated with MadGraph and Pythia.These generators provide the simulated physics processes used by ORCA.
A. Samples: The simulated events correspond to High-Luminosity LHC conditions with an average of µ = 200 simultaneous interactions.A simplified detector response is emulated using Delphes with a CMS Phase II configuration.
A. Samples: Each event is represented by a 110-dimensional vector containing selected jets, electrons, muons, photons, and missing transverse energy.The representation uses the 10 leading jets, four leading electrons, four leading muons, four leading photons, and MET.
A. Samples: Objects of each type are ordered by pT and truncated to fixed multiplicities.The retained counts are 10 jets and four each of electrons, muons, and photons.
A. Samples: Events with fewer objects are zero-padded, while objects beyond the retained counts are discarded.These rules enforce the fixed-dimensional event representation.
A. Samples: Jets are reconstructed with the anti-kt algorithm using a radius of R = 0.4.The passage specifies the jet-reconstruction algorithm and radius parameter.

B. Machine Learning Models

ORCA uses a two-stage architecture: supervised contrastive learning first organizes collider events into a physics-informed embedding, then an autoencoder assigns anomaly scores from reconstruction error. A direct-input autoencoder provides a matched baseline for isolating the contrastive representation’s contribution.

ORCA Stage 1: Embedder: ORCA first trains an encoder and projection network with supervised contrastive learning to separate physics-process groups in the representation space.The encoder maps each event to a DE-dimensional representation, while object-wise networks embed inputs into a common DO-dimensional space.
ORCA Stage 1: Embedder: A variance regularizer prevents representational collapse, using a conservative 0.9 margin and λ = 1.The normalized event representations lie on the DE-dimensional unit hypersphere, and dimensions exceeding the target standard deviation do not contribute to the loss.
ORCA Stage 1: Embedder: The embedder is trained on 48 Standard Model processes merged into 14 physics groups, with same-group events defining positive pairs.Events are split per process into training and validation sets with an 80/20 ratio.
ORCA Stage 2: Anomaly Detection: In Stage 2, an autoencoder reconstructs standardized 128-dimensional embeddings, with mean squared reconstruction error serving as the event anomaly score.The autoencoder is trained exclusively on background embeddings; events unlike the background produce large reconstruction errors.
Standard Autoencoder Baseline: The standard autoencoder baseline uses the same standardized 110-dimensional input feature vector and matches ORCA’s architecture and training procedure.Its inputs include kinematic features for up to 10 jets, 4 electrons, 4 muons, and 4 photons, plus missing transverse energy.

III. RESULTS

The RESULTS section presents ORCA’s structured latent-space representations and explains how contrastive clustering supports both anti-background selection and interpretation. It also uses proximity to known signal classes to classify test events and contextualize anomalous findings.

Latent-space representations: Figure 4 visualizes the ORCA embedding space with a six-dimensional principal component analysis of representative training signal models.The visualization examines latent representations learned for various Standard Model processes.
Selection and interpretation: Contrastive learning clusters information in the latent space, supporting anti-background selection that drives anomaly-detection performance.The structured space also enables standard high-energy-physics interpretation techniques to exploit high-dimensional information.
Anomaly characterization: ORCA can classify a test set by proximity to other signal classes, providing insight into potential anomalous findings in data.This proximity-based classification supports understanding anomalous events through their relationships to known signal classes.

A. Anomaly Detection Performance

ORCA’s anomaly-detection performance is evaluated through process sensitivity using AUC and fixed-background-rejection TPR, compared with a standard autoencoder. It achieves significant gains across nearly all tested signals, including strong high-background-rejection performance for the Υ(ℓℓ) exception.

Performance metrics: Anomaly-detection breadth is assessed by thresholding output loss and measuring AUC and fixed-background-rejection TPR on a representative signal test set.The training signal models are intended to represent processes accessible at the LHC.
Model comparison: ORCA is compared with a standard autoencoder to isolate the impact of its contrastive-learning-informed embedding, with gains shown across AUC, fixed-background-rejection TPR, and representative ROC curves.The comparisons cover all signal models used in training and the full range of background acceptance.
Results: Significant gains occur across nearly all signals in both metrics, except for Υ(ℓℓ) in AUC, where its embedding is close to minimum-bias data.In the high-background-rejection regime measured by TPR, ORCA recovers a strong performance gain over the autoencoder for Υ(ℓℓ).

B. Interpretability

ORCA interprets anomalous events by fitting known-process templates in a contrastively organized embedding space, with uncertainties adjusted for the composite likelihood structure. Validation shows accurate recovery for represented signals, while absent signals are characterized through the processes they most resemble.

Template-fit interpretation: Contrastive learning separates known processes into sufficiently distinct embedding regions for quantitative per-process decomposition of anomalous samples.The interpretation is obtained through a template fit in embedding space.
Template-fit interpretation: The fit reduces the embedding with PCA and ICA, then models per-dimension bin counts as mixtures of process-specific template probability mass functions.Each bin count is treated as an independent Poisson variable, and process yields are fit parameters.
Statistical treatment: Because the factorized likelihood is composite and overstates information when events enter multiple dimensions, yield uncertainties use a robust sandwich covariance estimator.Intervals are asymptotic and are not quoted for yields at the physical boundary νi = 0.
Template limitations: The limited template library makes fitted yields projections onto available reference processes, so unrepresented signals are absorbed by the templates they most resemble.The method diagnoses anomalous-region composition rather than identifying the true origin of signals absent from the library.
Validation: In holdout tests, 200 injected events from each excluded signal group were correctly attributed to the simulated signal process within error in informed fits.Blind fits categorized Di Higgs primarily as top pair and Single Vector Boson primarily as γ + V when the correct templates were absent.
Validation: The CL multi-dimensional embedding fit provided the highest signal recovery across injected yields, outperforming both the CL anomaly score and the multi-dimensional baseline autoencoder latent-space fit.The CL anomaly score showed nearly negligible signal recovery, while extra baseline latent dimensions alone were insufficient for the best performance.

IV. CONCLUSIONS

The conclusions present ORCA as a high-performance, interpretable anomaly-detection framework built from a physics-structured contrastive embedding and an autoencoder anomaly score. ORCA outperforms a baseline autoencoder across diverse Standard Model processes while retaining deployment flexibility through interchangeable model components.

Conclusions: ORCA combines labeled signal events for physics-structured latent-space learning with an embedding-based autoencoder reconstruction objective for per-event anomaly scoring.The framework uses a two-stage architecture: contrastive embedding followed by anomaly scoring in the learned space.
Conclusions: ORCA outperforms a baseline autoencoder on AUC and fixed background rejection metrics across a wide variety of Standard Model processes.The reported performance advantage covers both discrimination quality and fixed-background-rejection evaluation.
Conclusions: Direct likelihood fitting of the ORCA embedding space provides a novel route to interpreting anomaly-detection results.This interpretation method is enabled by the embedding’s high performance.
Conclusions: The framework can retain interpretability while replacing the transformer embedder or embedding-based anomaly model to meet deployment constraints.A lightweight embedder could support low-latency environments such as hardware triggers, while the autoencoder can be replaced by another model operating on the embedding.