21.02.2022 | AI in particle physics

Artificial intelligence looks for unknown particles

Physicists from the Austrian Academy of Sciences (OeAW) use artificial neural networks to search for previously unknown particles in the data from particle accelerators. Their method, now published in the European Physical Journal C, saves a lot of time and could help discover new physics in the future.

A look inside the SuperKEKB accelerator and Belle II detector in Tsukuba, Japan. © KEK

The search for new particles with accelerators, such as at CERN, is complex. Because interference is ubiquitous and even the smallest deviations in the angle when two particles collide can lead to completely different results, physicists must observe an enormous number of particle collisions to determine through statistical analysis whether an unknown particle may have formed. Researchers at the Institute of High Energy Physics of the Austrian Academy of Sciences (OeAW) have developed so-called Punzi-nets to significantly accelerate this process using machine learning. The work is part of the InterLeptons project, funded by the European Research Council (ERC).

Machines recognize patterns

A Punzi-net is a neural network that first learns what a possible signal for a specific particle might look like. "We train the neural network with data simulating both the signal and the stochastic background noise of the detectors of an accelerator for a specific particle. This is how the network learns to distinguish between signal and noise," explains Gianluca Inguglia, physicist at the OeAW and head of InterLeptons. But that's not enough to accelerate the search for unknown particles.

In a second step, the Punzi-net is further optimized by learning how to fulfill a quality criterion for high-energy physics developed by the Italian physicist Giovanni Punzi as well as possible. "The network learns to minimize a new function that we call Punzi-loss. This is the reciprocal of Punzi's quality criterion. The smallest possible Punzi-loss means a high signal quality. In this way, the Punzi-net learns to check different possible masses of a particle at once, even those that were not included in the training data," Inguglia says.

Unknown mass

Especially when the exact mass of a particle is not known, Punzi-nets can speed up the search enormously. "In such cases, we previously had to individually test many different models for the various possible masses. If we have several hundred models, each test takes several hours. A Punzi-net can check all these models in the same amount of time," Inguglia says.

This makes the new tool interesting in the search for dark matter, for example. The researchers are currently in the process of applying their method for the first time to real data from the SuperKEKB electron-positron accelerator in Japan to search for the hypothetical Z boson, which could interact with dark matter. Here, too, the Punzi-net is trained using simulated data from a model and then used to search for traces in the real detector data. This approach makes the method highly customizable. In principle, it can be used to search for any particle for which a model for training the neural networks can be created.

In the future, scientists will be able to use Punzi-nets to search much more efficiently for new particles in accelerator experiments such as the LHC (Large Hadron Collider) at the CERN research center. "I expect that many colleagues will use our method and perhaps refine it. This is really great and the feedback from high energy physics is extremely positive. I am particularly pleased that two excellent doctoral students, Paul Feichtinger and Huw Haigh, mainly developed the method," says Inguglia.



„Punzi-loss: a non-differentiable metric approximation for sensitivity optimisation in the search for new particles“, F. Abudinén, M. Bertemes, S. Bilokin et al., The European Physical Journal C, 2022
DOI: https://doi.org/10.1140/epjc/s10052-022-10070-0