CMS Uses Neural Flow to Spot Anomalous Dijet Events
Researchers applied an unsupervised neural spline flow model to CMS Open Data, identifying dijet events with low likelihood scores that deviate from
A new study uses a neural spline flow model to hunt for anomalous dijet events in proton-proton collision data without specifying a target signal. The unsupervised method, applied to CMS Open Data, flags events with low likelihood scores as potential deviations from the dominant Standard Model background.
Researchers trained a normalizing flow model on a high-dimensional set of jet, dijet, and event-level observables. The model learns the complex probability distribution of typical collision events directly from the data. Any event that falls into a low-probability region under this learned density is considered a candidate anomaly.
Methodology and Validation
The analysis was performed on a dijet sample from CMS Open Data. After training, the team scrutinized events with extreme anomaly scores. They conducted a thorough validation to ensure the findings were strong. This included statistical comparisons at the feature level, tests to decorrelate the anomaly score from dijet mass, permutation-based null tests, and checks on training stability.
The selected anomalous events showed significant departures from the background-only expectation. These deviations were most pronounced in jet-substructure observables. Crucially, the identified anomalies remained stable under several known biases that can plague unsupervised learning techniques.
Characteristics of the Anomalies
The anomalous events are scattered across the kinematic phase space. They do not cluster into a narrow peak in the dijet invariant-mass spectrum, which would be a classic signature of a new particle resonance. Instead, the anomalies display correlated deviations across multiple observables.
This pattern suggests a multivariate difference in jet substructure and overall event topology. The anomalies are consistent with a structured, but diffuse, departure from the Standard Model rather than a simple, localized signal.
Implications for New Physics Searches
The study, detailed in a paper on arXiv, makes no claim of discovering new physics. Its primary aim is to demonstrate the technique's sensitivity. The neural spline flow-based approach proves capable of detecting rare, structured deviations in collider data.
This method offers a model-independent tool for exploratory searches. It could complement traditional, hypothesis-driven searches for physics beyond the Standard Model at the Large Hadron Collider. The work shows that density estimation can flag interesting events for further, targeted investigation.
The research provides a new avenue for sifting through vast datasets. It looks for the unexpected without being told what to look for. The final events identified by the flow model present a puzzle for physicists to decipher.
