Contrastive Learning for Interpretable Anomaly Detection at Collider Experiments

When physicists search for new particles at the LHC, they often look for events that look nothing like the known background. The challenge is that many anomaly‑detection methods give a single score that is difficult to translate into physical insight, and that score tends to rise simply because an event has more energy or more reconstructed objects. A new approach called ORCA (Organized Representation via Contrastive learning for Anomaly detection) tackles both issues by first shaping an embedding space with contrastive learning and then measuring how poorly an autoencoder can reconstruct events in that space.

What You Need to Know

ORCA works in two stages. In the first stage, a neural network is trained with supervised contrastive loss using labels that denote the physics process each simulated event belongs to (e.g., QCD multijet, Z + jets, top‑pair production). This forces the network to map events from the same process close together while pushing different processes apart, creating an organized representation of the feature space. In the second stage, a standard autoencoder is applied to these embeddings. The reconstruction error of the autoencoder serves as the anomaly score; because the embeddings already factor out trivial variations like overall energy scale or object multiplicity, the score is more directly related to genuine deviations from known physics.

The authors test ORCA on a simulated dataset that mimics the high‑luminosity LHC environment, including pile‑up and detector effects. They compare ORCA’s performance to a baseline autoencoder that operates directly on raw high‑level features (jet transverse momenta, pseudorapidities, etc.). Across a variety of benchmark signals—such as dark matter mono‑jet, supersymmetric particle pairs, and exotic Higgs decays—ORCA yields higher signal‑to‑background ratios and a broader range of detectable signal strengths.

Why It Matters

Interpretability is crucial when a potential discovery is made. Physicists need to understand why an event is flagged as anomalous in order to design follow‑up analyses or theoretical models. By decoupling the anomaly score from trivial kinematic correlations, ORCA makes it easier to inspect which high‑level features (e.g., missing transverse energy, jet substructure) drive the reconstruction error, facilitating a physics‑driven interpretation.

Moreover, the method’s improved sensitivity means that rarer or more subtle new‑physics signatures could be uncovered with the same amount of data. This is especially important for the upcoming High‑Luminosity LHC run, where the sheer volume of collisions will amplify both the opportunities and the challenges of anomaly detection.

Key Details

• Contrastive training: Uses process labels from Monte Carlo samples to shape an embedding space where similar physics clusters tightly.
• Autoencoder stage: Operates on the contrastive embeddings; reconstruction error defines the anomaly score.
• Dataset: Simulated HL‑LHC conditions with 14 TeV proton‑proton collisions, including realistic pile‑up.
• Baseline comparison: Standard autoencoder fed directly with raw high‑level jet and missing‑energy variables.
• Signal benchmarks: Dark matter mono‑jet, simplified supersymmetry models, and Higgs‑to‑invisible decays.
• Results: ORCA achieves up to a 30 % increase in signal efficiency at a fixed background rejection rate and shows reduced correlation of scores with total event energy and object count.

What’s Next

The authors suggest extending ORCA to incorporate additional auxiliary inputs, such as tracker‑level information or particle‑flow candidates, to further enrich the embedding space. They also plan to test the framework on real data from LHC Run 3, where unsupervised validation strategies (e.g., sideband studies) can confirm that the observed improvements persist outside of simulation.

📌 Source: Arxiv Ml