Model-independent searches and anomaly detection at the CMS experiment
This paper presents the CMS experiment's model-independent search for new physics using machine-learning-based anomaly detection on 138 fb of 13 TeV data, highlighting the successful unsupervised recovery of top quarks and the real-time deployment of unsupervised algorithms in the level-1 trigger during 2024.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
For more than a decade, the world's most powerful particle accelerator has been smashing protons together at speeds close to the speed of light, searching for the building blocks of a universe that extends beyond our current understanding. Scientists have long suspected that the Standard Model, the theory that successfully describes all known matter and forces, is incomplete. They expect to find new particles or hidden forces that could explain mysteries like dark matter or why the universe exists at all. However, despite the discovery of the Higgs boson, no new elementary particle has appeared in the data. This silence has forced researchers to rethink their strategy. Instead of building a specific net to catch a fish they think might be there, they are now trying to catch anything unusual that swims by, regardless of what it looks like. This shift relies on a powerful idea: if the universe is full of familiar patterns, then anything that breaks those patterns is worth investigating, even if we do not yet know what it is.
The researchers at the Compact Muon Solenoid, or CMS, experiment at CERN have embraced this approach by using machine learning to hunt for anomalies. In a recent study, they applied this technique to a massive dataset of 138 fb⁻¹ of proton-proton collision data recorded at an energy of 13 trillion electron volts. The team focused on a specific type of collision where two high-energy sprays of particles, known as jets, fly out in opposite directions. In the vast majority of these events, the jets are formed by common particles like quarks or gluons, creating a predictable, messy spray. The scientists looked for jets that were different—ones that carried a complex internal structure, perhaps containing heavier particles or multiple layers of decay, which would make them stand out from the ordinary background. They did not assume what these new particles might be, nor did they guess their mass or how they interacted. They simply asked the computer to learn what a normal jet looks like and then flag anything that did not fit the pattern.
To ensure this method worked, the team first tested it on data where the answer was already known. They asked the algorithm to find the top quark, a heavy particle that is well understood but difficult to spot in a sea of ordinary jets. Without being told what a top quark was or given any labels to identify it, the machine learning system successfully isolated events containing these particles. It found them by noticing that their internal structure was distinct from the rest. This validation proved that the system could recognize something new without a pre-written script. When the researchers then applied the same logic to the full dataset of 138 fb⁻¹ of proton-proton collision data, they scanned for any localized excess of strange events that might signal a new resonance, or a heavy particle decaying into two lighter ones.
The results were a testament to the power of this model-independent search, but they also confirmed the continued silence of the universe regarding new physics. The five different machine learning methods used in the study all performed better than traditional searches that look for specific types of particle signatures. They were able to spot subtle anomalies that older methods would have missed, particularly in complex scenarios where a heavy particle might break apart into many smaller pieces. Yet, despite this increased sensitivity, no single method found a deviation from the expected background that was strong enough to claim a discovery. The most significant fluctuations they saw were consistent with random statistical noise rather than a new particle. The team set new limits on what kinds of new particles could exist, effectively ruling out certain possibilities in mass ranges where no other search had looked before.
Beyond the offline analysis of recorded data, the experiment has pushed this technology into the realm of real-time decision making. The Large Hadron Collider produces collisions at a rate of 40 million times per second, far too fast for human operators or standard computers to review every single event. Usually, a system called a trigger filters out the vast majority of collisions, keeping only those that look interesting based on known physics. The CMS team has now installed two unsupervised machine learning algorithms directly onto the hardware that makes these split-second decisions. These algorithms, running on specialized chips, analyze the raw energy deposits in the detector without any prior knowledge of what to look for. During the 2024 data-taking period at an energy of 13.6 TeV, these systems selected more than four billion events for permanent storage, capturing rare and energetic processes that conventional triggers might have discarded. The data collected so far shows that the system is working as intended, highlighting well-known particles like the Z boson and the J/psi meson, proving that the machine can identify interesting physics without being told what to find.
The ultimate goal of this work is to expand the search beyond just the sprays of particles inside a single jet to the entire event. By looking at the global shape of a collision, scientists hope to find anomalies that are invisible when looking at individual pieces. While this specific study did not find a new particle, it successfully demonstrated a new way of looking at the data. It showed that the tools exist to find the unexpected, even in a universe that has so far remained stubbornly quiet. The researchers have moved from theoretical proposals to practical application, validating their methods on real data and proving that they can characterize the properties of a particle just by observing how it differs from everything else. The search continues, not for a specific ghost, but for any shadow that does not belong.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.