Speaker
Description
The luminosity upgrade in high-energy physics experiment means large amount of data need to be processed and transferred. It brings challenges for design of vertex detectors both on the power consumption and readout rate. Our groups try to implement the compact data pre-processing module and artificial neural network into CMOS Monolithic Active Pixel Sensor (MAPS) to remove clusters generated by beam background and reduce the data load of system. The compact pre-processing algorithm and ANN model have been deployed in offline method, with validation conducted through MAPS beam test datasets.
Summary (500 words)
The CMOS Monolithic Active Pixel Sensor (MAPS) architecture has been successfully implemented in vertex detectors for both the STAR experiment (Ultimate-2) and ALICE ITS2 (ALPIDE). The vertex detector is located at the most closed to the interaction point. With increasing luminosity in high-energy physics experiments, the massive data poses challenges to legacy MAPS design regarding power consumption and readout bandwidth requirements. Our prior simulation studies revealed that substantial cluster in vertex detectors are generate by charged particles from beam background. These charged particles have large angles of incidence, producing elongated clusters. Our group try to develop a methodology based on artificial neural networks, reconstructing angles of incidence according to hit distribution within clusters, classifying source of charged particles, removing extra clusters and reducing data flow of system.
A pixel-level clustering algorithm is designed to identify hits and organize adjacent hits into a cluster. Features of a cluster related to angle of incidence and corresponding extraction operators are designed. These operators for different features are inspired by the Sobel operator which is employed in the image processing. A specific three layers artificial neural network is proposed to reconstruct the angle of incidence according to these features. In order to implement these algorithms into the MAPS ASIC, they are the trade-off between complexity and occupied resources.
TaichuPix-3 is a MAPS developed for the baseline vertex detector of the Circular Electron Positron Collider (CEPC). It is fabricated by a 180nm CMOS imaging sensor technology with a pixel pitch of 25 × 25μm2. It contains 1024×512 with a thickness of 150μm. Our colleagues have setup the system to collect the MAPS output under four different incident angles by beam test.
The acquired experimental data is processed, and features extracted are fed into our design to train the artificial neural network according to incident angles. A control group is set to reconstruct incident angles by a complex convolution neural network. We will compare the classification results and make discussion.
