Search for Dark Matter with Long Lived Particles at the LHC – DMwithLLPatLHC

Machine-learning techniques (Deep or Graph Neural Networks) were used with the aim to improve on the calibration of objects in the ATLAS detector, either basing them on high-level jet input or on constituent inputs. Multiple studies were done on photon identification enhancement, especially through a different representation of the energy pattern in the electromagnetic calorimeter and asymmetry variables that can be built out of this representation, which are especially helpful in the identification of collimated photons. Machine learning techniques were also used to show the potential in reducing the cross-talk effect and thus improving on the timing resolution of the electromagnetic calorimeter. The project allowed to contribute to two published searches for dark hadrons using the ATLAS experiment and one searching for axion-like particles. The MadAnalysis framework was used to implement new reinterpretation codes relevant to these searches and also to perform phenomenology studies of dark matter models.

The project allowed to improve on the performances for photons and jets, by using expert knowledge and machine learning, with some of these improvements already available in the ATLAS workflow for the whole collaboration to use or study. The project also allowed to contribute significantly to analyses probing completely new BSM models / parameter space using ATLAS Run-2 and partial Run-3 data, resulting in original publications and paving the way to even better analyses for the full Run-3 dataset, also based on phenomenology studies that have been made possible thanks to reinterpretation codes that are public, thus propagating the knowledge related to these analyses. Some communication actions (in conferences, workshops, in a course in a thematic school, or interviews) were also done to disseminate these results.

There is a new COFECUB project with Brazil (SEMA-ML: Searching for Dark Matter with Advanced Machine Learning Modelling) which will continue the collaboration on these topics. There are also other means by which these are pursued. Jet and photon performance studies will continue within the relevant ATLAS groups. A search for anomaly detection search using jet-based signatures in the context of a new MIAI chair between LPSC and LPC/Clermont is underway. Finally, the phenomenology studies will continue with the reintrepretation codes, which can benefit from collaborations within the context of the LHC BSM working group.

Long-lived particles are predicted in many dark matter models, including two emerging categories which are the focus of this project: axion-like particles (ALPs) and dark hadrons. These currently have limited coverage at colliders: there is hence an opportunity to gain access to completely new regions of the new physics parameter space. For the ALP, we will focus on its decay into a photon pair either promptly or inside the calorimeters. For the dark hadrons, the peculiar characteristics of their jets will be exploited: different number of tracks, jet substructure, development of the shower in the calorimeters that could come from the late decay of long-lived dark sector hadrons,... These searches will not only be developed within ATLAS, but theoretical aspects allowing to link phenomenology to underlying theory parameters will be addressed.

These searches use photons and jets, which both interact with matter as a particle shower governed by stochastic processes, even if with different underlying physics. Advanced neural network techniques will be developed to use the shower constituent information to improve on the performances; to improve training the uncertainties on input variables will also be considered. By building such tools based on low-level information (from the photon and jet constituents) rather than high-level information (from the photons and jets directly), we aim to improve significantly the precision of the measurement of energy, direction and mass with respect to existing techniques. For the same reasons, such tools will allow for better identification of signal photons and jets and better background rejection.

A final aim of this project will be to make it possible to reinterpret the results of the searches described above in terms of as many other models as possible. Setting up a collaboration (as part of an ATLAS Short-Term Association) between theorists and experimentalists from the beginning of the inception of the analyses will allow for the most general possible interpretation of the results. This includes an optimal choice of benchmarks, and will at the same time allow for gathering all of the necessary information to reinterpret the results. To achieve this, we will develop an appropriate software framework in the initial stages of the proposal, and test it by recasting existing ATLAS long-lived particle analyses.