FPGA implementation of a deep learning algorithm for real-time signal reconstruction in radiation detectors under high pile-up conditions

Abstract

The analog signals generated in the read-out electronics of radiation detectors are shaped prior to the digitization in order to improve the signal to noise ratio (SNR). The real amplitude of the analog signal is then obtained using digital filters, which provides information about the energy deposited in the detector. The classical digital filters have a good performance in ideal situations with Gaussian electronic noise and no pulse shape distortion. However, high-energy particle colliders, such as the Large Hadron Collider (LHC) at CERN, can produce multiple simultaneous radiation events, which produce signal pileup. The performance of classical digital filters deteriorates in these conditions since the signal pulse shape gets distorted. In addition, this type of experiments produces a high rate of collisions, which requires high throughput data acquisitions systems. In order to cope with these harsh requirements, new read-out electronics systems are based on high-performance FPGAs, which permit the utilization of more advanced real-time signal reconstruction algorithms. In this paper, a deep learning method is proposed for real-time signal reconstruction in high pileup radiation detectors. The performance of the new method has been studied using simulated data and the results are compared with a classical FIR filter method. In particular, the signals and FIR filter used in the ATLAS Tile Calorimeter are used as benchmark. The implementation, resources usage and performance of the proposed Neural Network algorithm in FPGA are also presented.