Physicists use machine learning techniques to explore strange collisions that could hint at new physics

One of the main goals of the LHC experiments is to look for signs of new particles that could explain many of the unsolved mysteries of physics. Often searches for new physics are designed to look for one specific type of new particle at a time, using theoretical predictions as a guide. But what if we want to look for new particles that are unpredicted and unexpected?

It would be an enormous task for physicists to comb through the billions of collisions that occur in the LHC experiments without knowing exactly what to look for. So instead of scouring the data for anomalies, the ATLAS and CMS collaboration is turning to artificial intelligence (AI) to do the job.

At the Morion conference on March 26, physicists from the CMS collaboration presented their latest results using a variety of machine learning techniques to search for pairs of “jets” – focused sprays of particles arising from strongly interacting quarks and gluons – that are particularly difficult to analyze but may hide new physics.

ATLAS and CMS researchers are using several strategies to train AI algorithms in the search for jets: By studying the shape of the jet's complex energetic signature, scientists can pinpoint which particles generated the jet.

Using real collision data, physicists from both experiments are training an AI to recognize the characteristics of jets emerging from known particles, which will enable the AI ​​to distinguish between these jets and features of atypical jets that may suggest new interactions, which appear in the dataset as an accumulation of atypical jets.

Another approach is to instruct AI algorithms to consider the entire collision event and look for unusual characteristics in the various particles detected. These unusual characteristics could indicate the presence of new particles. This technique was demonstrated in a paper published by ATLAS in July 2023, marking one of the first uses of unsupervised machine learning in LHC results.

CMS takes a different approach, having physicists create simulated examples of potential new signals and then tasking AI with identifying collisions in the real data that are different from normal jets but similar to the simulations.

The latest results published by CMS showed that each AI training method showed different sensitivity to different types of new particles, with no single algorithm proving to be best.

The CMS team was able to limit the production rates of several types of particles that generate the anomalous jets, and also demonstrated that their AI-driven algorithms significantly improved sensitivity to different particle signatures compared to traditional techniques.

These results show that machine learning is revolutionizing the search for new physics. “We already have ideas about how to improve the algorithm further and apply it to different parts of the data to search for different types of particles,” says Oz Amram from the CMS analytics team.