Hadronic Mono-Z Dark Matter Sensitivity with Flow Matching on CMS Open Data
This paper presents a projected sensitivity study for hadronic mono- dark matter using CMS Open Data, demonstrating that a conditional flow-matching model incorporating detailed extra-jet kinematics achieves expected significances up to 7.62 while addressing methodological challenges like missing-object imputation and in-sample bias.
Original paper licensed under CC BY 4.0 (https://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
The universe is filled with matter that we cannot see. Astronomers know this invisible stuff, called dark matter, exists because its gravity pulls on stars and galaxies, yet it refuses to interact with light or ordinary matter in any other way. Finding out what this substance is made of remains one of the biggest mysteries in physics. One way scientists try to catch a glimpse of it is by smashing protons together at incredibly high speeds inside massive machines called particle colliders. If dark matter particles are created in these collisions, they would fly away undetected, carrying energy with them. This missing energy would leave a tell-tale imbalance in the debris of the crash, a gap in the accounting that suggests something invisible has escaped.
To find this missing energy, researchers look for specific patterns where a known particle, like a Z boson, is produced alongside the invisible dark matter. The Z boson is a heavy carrier of the weak nuclear force, and it can decay into a pair of jets, which are sprays of particles that look like a single, heavy object in the detector. This specific setup, where a Z boson appears with nothing else but missing energy, is a promising hunting ground. However, the real challenge is that the background noise from ordinary particle collisions is overwhelming. It is like trying to hear a whisper in a hurricane; the signal is there, but the storm of ordinary events makes it nearly impossible to distinguish without incredibly sophisticated tools to separate the two.
In a recent study using data from the Compact Muon Solenoid (CMS) experiment at the Large Hadron Collider, researchers tackled this challenge by applying a new kind of artificial intelligence to the search for dark matter in these hadronic, or jet-based, collisions. The team, working with open data from 2015, focused on a specific dataset containing over 20 million recorded collisions. From this massive pool, they selected about 1.44 million events that fit the criteria for a Z boson produced with missing energy. Instead of relying on traditional statistical formulas to guess what the background noise should look like, they trained a flexible computer model to learn the shape of the background directly from the data itself. This model acts like a highly sensitive filter, learning the complex patterns of ordinary collisions so well that it can spot when a new event looks strange enough to be a sign of dark matter.
The researchers used a technique called flow matching, which allows the computer to map out the probability of different collision outcomes without being forced into rigid mathematical boxes. They taught the model to recognize the background by feeding it the selected collision events, but they kept a separate group of events hidden from the training process to test the model later. This ensured the computer wasn't just memorizing the data but actually learning the underlying patterns. To handle the messy reality of particle physics, where some measurements might be missing if no extra particles are produced, the team developed a special method to mark these gaps clearly rather than filling them with fake numbers that could confuse the model. They also applied a safety check to ensure that any signal they claimed to find was based on a solid number of background events, preventing them from being fooled by random statistical flukes in the tail end of the data.
When they tested this system against simulated dark matter signals, the results were striking for certain types of theoretical models. For two of the three scenarios they tested, the model projected that the experiment could have achieved a discovery-level sensitivity, reaching a significance that would be considered a discovery in the field if such a signal were observed. For the third, lighter scenario, the projected sensitivity was lower, but still present. Crucially, the study revealed that the extra jets of particles often produced alongside the main collision were not just noise; they carried vital information. When the researchers removed the details about these extra jets from their analysis, the projected ability to spot the dark matter signal dropped dramatically, by more than half in some cases. This suggests that the specific shape and arrangement of these extra sprays of particles provide a unique fingerprint that helps distinguish a dark matter event from the chaotic background of ordinary collisions.
The study does not claim to have found dark matter, as it is a projection based on simulations and a specific slice of past data, not a new observation of a real discovery. However, it demonstrates that using advanced, flexible machine learning models on complex jet data can significantly improve the ability to search for these elusive particles. By moving away from rigid formulas and letting the data teach the model what normal looks like, the researchers showed that the hadronic channel, once considered too noisy to be useful, holds substantial promise. The work highlights that the subtle details of how particles scatter and the extra debris they leave behind are key to unlocking the secrets of the invisible universe, offering a clearer path forward for future searches in the high-energy frontier of physics.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.