Machine learning has become a popular instrument for the search of undiscovered particles and mechanisms at particle collider experiments. It enables the investigation of large datasets and is ...therefore suitable to operate directly on minimally-processed data coming from the detector instead of reconstructed objects. Here, we study patterns of raw pixel hits recorded by the Belle II pixel detector, that is operational since 2019 and presently features 4 M pixels and trigger rates up to 5 kHz. In particular, we focus on unsupervised techniques that operate without the need for a theoretical model. These model-agnostic approaches allow for an unbiased exploration of data while filtering out anomalous detector signatures that could hint at new physics scenarios. We present the identification of hypothetical magnetic monopoles against Belle II beam background using self-organizing kohonen maps and autoencoders. These two unsupervised algorithms are compared to a Multilayer Perceptron and a superior signal efficiency of the Autoencoder is found at high background-rejection levels. Our results strengthen the case for using unsupervised machine learning techniques to complement traditional search strategies at particle colliders and pave the way to potential online applications of the algorithms in the near future.
Celotno besedilo
Dostopno za:
DOBA, IZUM, KILJ, NUK, PILJ, PNG, SAZU, SIK, UILJ, UKNU, UL, UM, UPUK
2.
Track finding at Belle II Bertacchi, Valerio; Bilka, Tadeas; Braun, Nils ...
Computer physics communications,
02/2021, Letnik:
259
Journal Article
Recenzirano
Odprti dostop
This paper describes the track-finding algorithm that is used for event reconstruction in the Belle II experiment operating at the SuperKEKB B-factory in Tsukuba, Japan. The algorithm is designed to ...balance the requirements of a high efficiency to find charged particles with a good track parameter resolution, a low rate of spurious tracks, and a reasonable demand on CPU resources. The software is implemented in a flexible, modular manner and employs a diverse selection of global and local track-finding algorithms to achieve an optimal performance.
The upcoming Belle II experiment is designed to work at a luminosity of 8×10 35 cm -2 s -1 , 40 times higher than its predecessor. The pixel detector of Belle II with its ~ 8 million channels will ...deliver ten times more data than all other sub-detectors together. A data rate of 22 Gbytes/s is expected for a trigger rate of 30 kHz and an estimated pixel detector occupancy of 3%, which is by far exceeding the specifications of the Belle II event builder system. Therefore a realtime data reduction of a factor > 30 is needed. A hardware platform capable of processing this amount of data is the ATCA based Compute Node (CN). Each CN consists of an xTCA carrier board and four AMC/xTCA daughter boards. The carrier board supplies the high bandwidth connectivity to the other CNs via Rocket-IO links. In the current prototype design, each AMC board is equipped with a Xilinx Virtex-5 FX70T FPGA, 4 GB of memory, Gbit Ethernet and two bi-directional optical links allowing for a bandwidth of up to 12.5 Gbits/s. IPMI control of mother and daughter board is foreseen. One ATCA shelf containing 10 mother boards/40 daughter boards is sufficient to process the data from the 40 front end boards. The data reduction on the CN is done in two steps. First, the event data delivered by the front end electronics via optical links is stored in memory until the high level trigger (HLT) decision has been made. The HLT rejects >2/3 of these events. In a second step, pixel data of positively triggered events is reduced with the help of regions of interest (ROI), calculated by the HLT from projecting trajectories back to the pixel detector plane. The design allows additional ROI inputs computed from hit cluster properties or tracklets from the surrounding silicon strip detector. The final data reduction is achieved by sending only data within these ROIs to the main event builder.
We present an FPGA-based online data reduction system for the pixel detector of the future Belle II experiment. The occupancy of the pixel detector is estimated at 3%. This corresponds to a data ...output rate of more than 20 GB/s after zero suppression, dominated by background. The Online Selection Nodes (ONSEN) system aims to reduce the background data by a factor of 30. It consists of 33 MicroTCA cards, each equipped with a Xilinx Virtex-5 FPGA and 4 GiB DDR2 RAM. These cards are hosted by 9 AdvancedTCA carrier boards. The ONSEN system buffers the entire output data from the pixel detector for up to 5 seconds. During this time, the Belle II high-level trigger PC farm performs an online event reconstruction, using data from the other Belle II subdetectors. It extrapolates reconstructed tracks to the layers of the pixel detector and defines regions of interest around the intercepts. Based on this information, the ONSEN system discards all pixels not inside a region of interest before sending the remaining hits to the event builder system. During a beam test with one layer of the pixel detector and four layers of the surrounding silicon strip detector, including a scaled-down version of the high-level trigger and data acquisition system, the pixel data reduction using regions of interest was exercised. We investigated the data produced in more than 20 million events and verified that the ONSEN system behaved correctly, forwarding all pixels inside regions of interest and discarding the rest.
Machine learning has become a popular instrument for the identification of dark matter candidates at particle collider experiments. They enable the processing of large datasets and are therefore ...suitable to operate directly on raw data coming from the detector, instead of reconstructed objects. Here, we investigate patterns of raw pixel hits recorded by the Belle II pixel detector, that is operational since 2019 and presently features 4 M pixels and trigger rates up to 5 kHz. In particular, we focus on unsupervised techniques that operate without the need for a theoretical model. These model-agnostic approaches allow for an unbiased exploration of data, while filtering out anomalous detector signatures that could hint at new physics scenarios. We present the identification of hypothetical magnetic monopoles against Belle II beam background using Self-Organizing Kohonen Maps and Autoencoders. The two unsupervised algorithms are compared to a convolutional Multilayer Perceptron and a superior signal efficiency is found at high background rejection levels. Our results strengthen the case for using unsupervised machine learning techniques to complement traditional search strategies at particle colliders and pave the way to potential online applications of the algorithms in the near future.
The PANDA detector is a state-of-the-art general-purpose detector for physics with high luminosity cooled antiproton beams, planed to operate at the FAIR facility in Darmstadt, Germany. The central ...detector includes a silicon Micro Vertex Detector (MVD) and a Straw Tube Tracker (STT) or Time Projection Chamber (TPC). The electromagnetic lead tungstate calorimeter(EMC) provides almost 4π spatial coverage, good granularity and high energy resolution for electromagnetic showers measurement. A DIRC Cherenkov detector serves for particle identification. A novel trigger-less data push data architecture for the PANDA trigger and data acquisition system is proposed requiring the data from readout module to be processed in real-time to reconstruct charged tracks, electromagnetic showers and calculating PID parameters. This presentation shows results from the development of online high level trigger algorithms. A track finding algorithm for helix track reconstruction in the solenoidal field and a cluster finder for searching clusters in the EMC have been developed with special considerations for the implementation on the FPGA based Compute Node platform which has been developed for PANDA. Performance parameters such as momentum and spatial resolution for the helix track finder, energy and spatial resolution for the EMC cluster finder will be presented. With respect to the FPGA implementation, the partition strategy based on the readout electronics layout and the Compute Node processing architecture will be presented.
The upgrade of the Belle experiment and the KEKB accelerator aims to increase the experimental data set of the experiment by the factor 50. This will be achieved by increasing the luminosity of the ...accelerator which requires also a significant detector upgrade. A new pixel detector based on the DEPFET 1 technology is one of the detector upgrade requirements to handle the increased reaction rate and provide better vertex resolution. One of the features of the DEPFET detector is a long integration time of 20 μs. With the expected detector occupancy of about 2%, the detector will generate about 22 GB/s of data. A custom two-level read-out system consisting of Data Handling Hybrid and Online Selection Node is built to control the detector, process data and eventually reduce amount of data by one factor of magnitude. The DHH system synchronizes all detector modules to an external clock, provided by the clock and trigger distribution system of the experiment. It performs initial data processing that includes pixel remapping, sub-event building and cluster reconstruction. Then the data are further processed in the Online Selection Node and reduced by accepting only those hits which are associated to regions of interest pointed by other tracking detectors with better time resolution. The functionality, architecture and current state of the Belle II pixel detector readout system are discussed.
We present an FPGA-based online data reduction system for the pixel detector of the future Belle II experiment. The occupancy of the pixel detector is estimated at 3 %. This corresponds to a data ...output rate of more than 20 GB/s after zero suppression, dominated by background. The Online Selection Nodes (ONSEN) system aims to reduce the background data by a factor of 30. It consists of 33 MicroTCA cards, each equipped with a Xilinx Virtex-5 FPGA and 4 GiB DDR2 RAM. These cards are hosted by 9 AdvancedTCA carrier boards. The ONSEN system buffers the entire output data from the pixel detector for up to 5 seconds. During this time, the Belle II high-level trigger PC farm performs an online event reconstruction, using data from the other Belle II subdetectors. It extrapolates reconstructed tracks to the layers of the pixel detector and defines regions of interest around the intercepts. Based on this information, the ONSEN system discards all pixels not inside a region of interest before sending the remaining hits to the event builder system. During a beam test with one layer of the pixel detector and four layers of the surrounding silicon strip detector, including a scaled-down version of the high-level trigger and data acquisition system, the pixel data reduction using regions of interest was exercised. We investigated the data produced in more than 20 million events and verified that the ONSEN system behaved correctly, forwarding all pixels inside regions of interest and discarding the rest.