Should we look elsewhere for new physics?

For the first time, CMS deploys two new anomaly detection algorithms in its real-time event selection.

Display of a candidate event recorded by the CMS experiment.

For the first time, CMS has deployed in its real-time event selection system, and collected data with, two new anomaly detection algorithms – machine-learning frameworks designed to spot unusual patterns in data autonomously – enabling the collaboration to explore an even wider range of exotic collision events; one such event is shown above.

The two algorithms, AXOL1TL and CICADA, expand the reach of the CMS trigger system to a broad range of unusual events that may not fit the traditional physics signatures used to design trigger strategies. By looking for patterns that deviate from what is expected from standard signatures, the algorithms offer a way to search for signs of new physics without having to know in advance exactly what those signs might look like.

At the LHC, proton bunches cross up to 40 million times every second. The CMS Level-1 trigger (shown in the figure below) has just a few microseconds to decide which events to pass on for more detailed processing. Limits on data readout and processing capacity allow only about 0.3% of events to pass this first stage of selection. Conventional trigger algorithms look for features such as energetic particles or particular combinations of particles. These criteria capture a broad range of physics processes, but unexpected signals could fall outside them.

“The question is, if we don’t know what we’re looking for in those 99.7% of unsaved events, how can we design a trigger to select them?” asks Melissa Quinnan, postdoctoral fellow at UC San Diego, and member of the AXOL1TL team.

Above: Diagram showing the flow of data through the CMS trigger system, from the initial detector readout to the final data analysis. Inputs for the algorithms are shown on the left, with CICADA operating in the “calorimeter trigger”, and AXOL1TL in the “global trigger”.

The machine learning (ML) technique of anomaly detection is well suited to the problem of identifying interesting phenomena without knowing their precise characteristics. Anomaly detection algorithms learn the patterns of typical collision events and flag those that deviate from them. Recent advances in compression and implementation techniques of ML models on hardware like Field Programmable Gate Arrays (FPGAs) have made it possible to deploy more complicated algorithms such as autoencoders, within the strict timing constraints of the CMS Level-1 trigger.

“Getting these complex algorithms to run in tens of nanoseconds meant designing the model and training around the hardware from the start,” says Elliot Kauffman, graduate student at Princeton University, and member of the CICADA team. “You can’t just take a traditional neural network off the shelf and hope it fits.”

CMS collected over 700 million unique events with these two anomaly detection algorithms. Known rare features like the Z-boson peak in the distribution of reconstructed mass of muon pairs (figure below) can be seen enhanced in the events collected with the AXOL1TL dataset without any prior knowledge, demonstrating the potential of this model-independent approach.

“Dedicated analyses are underway to search for signs of new physics in this unique dataset selected by anomaly detection”, said Sabrina Giorgetti, Project Graduate at CERN and a member of the AXOL1TL team.

Above: Distribution of reconstructed di-muon mass in data collected with AXOL1TL and for a random sample of CMS collision data (zero bias). The Z-boson peak is enhanced in AXOL1TL data.

AXOL1TL and CICADA demonstrate how ultrafast anomaly detection can broaden the physics reach of CMS. They have already collected millions of events that conventional triggers would have missed, creating new opportunities to search for unexpected phenomena. Whether these events contain signs of new physics remains an open question, and the CMS team are eager to find out!

Written by: Emily Smith, for the CMS Collaboration
Edited by: Muhammad Ansar Iqbal

Explore more background image

Explore more

No posts were found. Try to change the category or the date filters.