Anomaly preserving contrastive neural embeddings for end-to-end model-independent searches at the LHC

The Big Picture

Imagine searching for a single misspelled word in every book ever printed, not knowing what word you’re looking for, or even what language it might be in. That’s roughly the challenge physicists face at the Large Hadron Collider. Every second, the LHC produces about 40 million proton collisions, and somewhere in that avalanche of data might be a particle or interaction that upends our understanding of fundamental physics. The catch: you don’t know what it looks like.

This is the core problem of model-independent anomaly detection, hunting for signs of new physics without a specific theory to guide the search. Standard approaches compress the LHC’s raw sensor data into a handful of physically meaningful numbers: particle energies, momenta, masses, and directions. But measurements hand-picked to capture known physics might throw away exactly the subtle patterns that would signal something genuinely new.

A new study from MIT and ETH Zürich, led by Kyle Metzger and collaborators at IAIFI, takes a different route: let the machine learn its own compact representation of collision events, built from the ground up to preserve anomalies rather than smooth them away.

Key Insight: Contrastive neural embeddings can learn low-dimensional representations of LHC collision data that are far more sensitive to rare new physics signals than classical feature engineering, but only when the right balance is struck between organizing background events and leaving room for anomalies to stand out.

How It Works

The core idea borrows from contrastive learning, a technique that teaches a neural network what “similar” and “different” mean without explicit labels. The network sees pairs or groups of events and learns to map them into a compact geometric space. Events that look alike cluster together; unusual events get pushed to the outskirts.

The team trained two flavors of contrastive embeddings, compact mathematical fingerprints of each collision event:

• Supervised contrastive learning, where physics labels like particle types and process categories guide the clustering.
• Self-supervised contrastive learning, where the network discovers structure on its own by randomly dropping or perturbing particle objects within an event to create different “views” of the same underlying collision.

Both were tested with two neural architectures. Multilayer perceptrons (MLPs) process a fixed set of particle-level features in a standard feed-forward way. Transformers, the same architecture behind modern language models, treat each collision event as a variable-length sequence of physics objects and capture relationships between particles more naturally.

The inputs are measured kinematic properties of reconstructed physics objects: jets (sprays of particles from quark or gluon scattering), leptons (electrons, muons, and their heavier cousins), and missing energy, a signature of particles that escape the detector unseen. The network compresses all of this into an embedding of just a few dimensions.

These embeddings then feed into signal-agnostic statistical detection methods, algorithms that scan for unusual clusters in the embedding space without being told what signal to look for. Tested against a suite of benchmark new-physics signals and rare Standard Model processes, the contrastive embeddings consistently outperform hand-crafted features, showing higher sensitivity to injected rare signals at fixed false-positive rates.

Why It Matters

The most interesting finding here isn’t a benchmark number. It’s a conceptual warning.

There’s a tension between two goals that might seem aligned: learning to classify background events accurately and learning to detect anomalies. When you optimize a representation to organize known background processes as cleanly as possible, you suppress the very irregularities that make new physics stand out. A perfectly organized background leaves no room for the unexpected. The best representation for known physics is not the best representation for discovering unknown physics. Future searches need to be designed with that tension in mind.

The sweet spot turns out to be a combination of data compression, forcing the network to be selective about what it preserves, and domain knowledge, using physics-motivated labels to guide which background structures get organized. Supervised labels without compression risk encoding too much background detail. Compression without guidance can lose signal-relevant structure. Put the two together and the embeddings land in a useful middle ground.

The LHC’s Run 3 is generating data at unprecedented rates, and the High-Luminosity LHC upgrade will push this further. Model-independent searches are essential complements to targeted searches, but their power depends entirely on the quality of the representation feeding into them. A single trained embedding can simultaneously search for diverse signals across different final states, covering the full complexity of LHC collision data at once.

Open questions remain. How do these embeddings behave under detector systematics and real experimental noise? Can self-supervised versions be deployed in real-time trigger systems to catch anomalies before data is even written to disk?

Bottom Line: By learning compact, anomaly-preserving representations of LHC collision events through contrastive training, the team has built a more sensitive and broadly applicable tool for new physics searches, one that doesn’t need to know what it’s looking for.