Santiago Folgueras, Universidad de Oviedo
Our world has witnessed a massive explosion of data and a surge of machine learning (ML) and AI applications. The result is an ever-increasing need for higher throughput and real-time computing capabilities. The Large Hadron Collider (LHC) and its experiments provide the perfect benchmark for bringing recent industry developments and exploring innovative technologies and algorithms to process large data volumes in real time to extract features and classify patterns.
Only a handful of the collisions produced by the LHC contain potentially interesting physics. A dedicated filtering (trigger) system makes it possible to determine, in real time, whether a given event of interest should be saved for data analysis or not. With the discovery of the Higgs boson, several questions remain unanswered, such as the nature of dark matter. The answer to those questions could be linked to the production of particles beyond the standard model (BSM) that may have long lifetimes compared to that of standard model particles at the weak scale.
If these long-lived particles (LLPs) were to be produced at the LHC, they would yield rare signatures which require dedicated triggering algorithms. The general goal of the INTREPID project is to enhance the trigger capabilities to enable the discovery of LLPs in colliders and thus find evidence of BSM physics. Several years before the start of the High-Luminosity (HL)-LHC, it is now the perfect time to explore alternative architectures and technologies not considered in the plans of the experiments, and that could not be explored otherwise. We will use a multidisciplinary approach involving advanced machine-learning techniques and real-time processing platforms to propose an innovative solution that may bring unexpected improvements to the trigger systems. Furthermore, it might be the only way in which LLPs could be discovered at the HL-LHC. Any manifestation of such particles will guide the future of high-energy physics and shed light upon the energy scale and nature of BSM physics.
The need for an advanced trigger system
Only a handful of the collisions produced by the LHC contain potentially interesting physics. The trigger system makes it possible to determine, in real time, whether a given event of interest should be saved for data analysis or discarded. In the CMS experiment, this selection is achieved by a two-level system: the level-1 trigger (L1T) consisting of custom hardware processors that receive data from calorimeter and muon systems, generating a trigger signal with a latency of a few microseconds. The high-level trigger is implemented in software and reduces the rate down to ~1 kHz.
In view of the High-Luminosity LHC, the L1T system of the CMS experiment will be fully replaced to greatly extend the throughput and capabilities of the current system despite the harsher environment. The L1T system (CMS Collaboration, 2020) has been designed to process 63 Tb/s input bandwidth with state-of-the-art commercial field-programmable gate arrays (FPGAs) and high-speed optical links reaching up to 28 Gb/s using generic processing cards based on advanced telecommunications computing architecture (ATCA) technology.
Key components and innovations
The projected L1T upgrade muon architecture (CMS Collaboration, 2020) is the starting point of INTREPID. Such architecture consists of three layers of increasing complexity, each running a more complex pattern recognition algorithm sequentially. The first layer collects the point-like signal (hits) from different subdetectors to generate stubs that are then processed in the second layer, the track finders, which build standalone muon tracks. The last layer, the global muon trigger, correlates the muon-only information (standalone tracks and stubs) with tracker tracks for ultimate precision. The first two are critical for the quests of LLPs.
The first challenge we will address is to enable the capabilities of the foreseen muon trigger architecture to detect LLP signatures, such as slow-charged particles or hadronic showers in the muon system. These showers could be sensitive to BMS physics, leaving signatures in the form of emerging jets that do not point to the primary vertex and cannot be triggered by the existing solutions.
Muon showers: If a muon travelling through the steel of the magnet flux-return yoke has sufficiently large momentum, radiative energy losses become relevant, and even dominant, above the muon critical energy for iron (300 GeV). This radiation creates cascades of particles (showers) that lead to extra hits being reconstructed in the muon detectors with great impact on trigger performance.
In addition, we will be taking a step aside from existing CMS Collaboration plans and exploring alternative solutions that the existing algorithmic and architectural choices could have missed. Starting from hit-level information, I will investigate the discovery potential for LLPs of a hit-based pattern reconstruction algorithm for muon reconstruction at the L1T system. Having all the point-like information directly available in the same processing board will allow unprecedented performance and discovery reach. Two innovative technologies will be investigated:
- Machine learning and AI techniques: We use tracking detectors to estimate the kinematics of the particles produced in a collision. Charged particles ionise the material of these detectors as they travel through them, providing several position measurements (hits) along the trajectory. The state-of-the-art techniques for trajectory reconstruction may not be reliable in future facilities, but machine learning (ML)-based solutions, particularly graph neural networks (GNNs), could bring a lot of potential to solve this problem. GNNs can naturally deal with the sparse and irregular nature of collision data and provide a unique opportunity to reconstruct signals from LLPs.
- Adaptative compute acceleration platforms (ACAPs): A disruptive architecture recently released, and its applicability to the L1T system deserves to be explored. It combines best-in-class 7nm programmable logic with scalar processing engines and vector processing intelligent (AI) engines. It provides a foundational platform that allows targeting workloads to the right type of processors for optimal performance (15x better than FPGAs [Xilinx, 2020]) and may significantly impact particle physics applications.
INTREPID’s team
The INTREPID team is a multidisciplinary team of physicists and electronic engineers. The team includes PhD students and postdocs who are working together and in collaboration with the rest of the HEP team in Oviedo. The team consists of three PhD students (Pelayo Leguina, Javier Prado, and Daniel Estrada), one postdoc (Dr Clara Ramón, from February to August 2024 and Dr Andrea Cardini, since November 2024). The team is led by Prof. Santiago Folgueras.
First look at the Run-3 data
To highlight the importance of a good trigger strategy, with the first data collected by the CMS experiment in 2023, we were capable of searching for exotic LLPs decaying to final states with a pair of displaced, oppositely charged muons originating from a common vertex spatially separated from the proton-proton interaction point by distances ranging from several hundred microns to several metres. The sensitivity of the search benefits from new triggers for displaced dimuons developed for Run 3, which improves the previous Run-2 strategy up to a factor of four. This enhancement allowed us to achieve a similar sensitivity to the previous result of the CMS Collaboration with a fraction of the luminosity (CMS Collaboration, 2024). The main limitation of this analysis, however, remains the L1T, and it is the main line of action of the INTREPID project.
Preliminary results
After one year of work, our main achievement is the development of the first version of the algorithm for muon shower identification in the layer-1 barrel muon trigger of the CMS experiment. Such signatures could be sensitive to BSM physics, leaving signatures in the form of emerging jets that do not point to the primary vertex and cannot be triggered upon with the current existing solution, the so-called analytical method (AM) for muon reconstruction (Abbiendi et al., 2023). Additionally, such signatures appear naturally in events with energetic muons and result in inefficiencies or momentum misassignments (Sirunyan et al., 2020).
A simple hit-counting algorithm has been designed to run in parallel with the analytical method, enabling the identification of showers at a given muon station when the number of received hits exceeds a specified threshold. When this occurs, the shower tagging algorithm saves all the hits corresponding to a shower to measure its position, time and width. It then sends a signal to the AM to ignore its output and tag the shower instead for further processing at later stages of the trigger.
The algorithm can correctly tag 99% of the events while 0.07% are mis-tagged. On average, the algorithm recovers 88% of the hits corresponding to a shower. This result was shown at the ICHEP and TWEPP conferences this year.
Prospects for the future
During the next months, we will work on a proof of concept for the first model to reconstruct displaced muons in the overlap muon track finder using a GNN from both the software and firmware side. These developments will allow us to significantly enhance the trigger capabilities of the CMS detector.
References
Abbiendi, G. et al. (2023) ‘The Analytical Method algorithm for trigger primitives generation at the LHC Drift Tubes detector’, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1049, 168103. doi: 10.1016/j.nima.2023.168103.
CMS Collaboration (2020) The Phase-2 Upgrade of the CMS Level-1 Trigger. CERN-LHCC-2020-004; CMS- TDR-021. Available at: https://cds.cern.ch/record/2714892/files/CMS-TDR-021.pdf.
CMS Collaboration (2024) ‘Search for long-lived particles decaying to final states with a pair of muons in proton-proton collisions at √s= 13.6 TeV’, Journal of High Energy Physics, 2024(5), 47. doi: 10.1007/JHEP05(2024)047.
Sirunyan, A.M. et al. (2020) ‘Performance of the reconstruction and identification of high-momentum muons in proton-proton collisions at √s = 13 TeV’, Journal of Instrumentation, 15(02), P02027. doi: 10.1088/1748-0221/15/02/P02027.
Xilinx (2020) Versal: The First Adaptive Compute Acceleration Platform (ACAP). Available at: https://www.xilinx.com/support/documentation/white_papers/wp505-versal-acap.pdf.
PROJECT NAME
INTREPID
PROJECT SUMMARY
The discovery of the Higgs boson at the LHC closes a central chapter of the standard model of particle physics but raises questions about dark matter, neutrino masses and baryon asymmetry. Answers may involve long-lived particles (LLPs) with non-standard signatures, requiring advanced identification algorithms. Enhancing trigger capabilities to discover LLPs and evidence of beyond the standard model (BSM) physics is, therefore, a crucial task. INTREPID uses advanced machine learning and ultra-fast processing to improve future trigger systems, potentially revolutionising high-energy physics and benefiting industries needing real-time data processing.
PROJECT LEAD PROFILE
Santiago Folgueras is a professor at the University of Oviedo and has been working at the CMS experiment since 2009. He is the author of 57 peer-reviewed scientific publications. Folgueras has been invited to present his research results at 14 international conferences, as the CMS Collaboration has recognised him as a key expert in SUSY searches, muon performance and trigger development. During his research career, Professor Folgueras has participated in seven research projects funded under competitive national calls.
PROJECT CONTACTS
Santiago Folgueras, Prof. of Physics
Department of Physics
Universidad de Oviedo
Email: folguerassantiago@uniovi.es
Web: https://intrepid.uniovi.es
FUNDING
This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme under grant agreement No. 101115353.
Figure legends
Figure 1: A proton-proton collision at a centre-of-mass energy of 13 TeV, recorded by the CMS experiment, compatible with the production of a Higgs boson decaying to long-lived particles that shower (orange lines) in the cathode strip chambers of the muon detectors system (red boxes).
Figure 2: Working scheme of the proposed shower algorithm, which will run in parallel to the AM to identify showers in the muon system.
Figure 3: Non-radiating muon hit count remains below threshold; when a muon radiates, the muon hit count surpasses a given threshold and sends a signal to the AM to stop processing the event.
