Speaker
Description
The COMET Phase-I experiment searches for a muon-to-electron conversion at a target sensitivity of $3\times10^{-15}$, which has never been observed. The event signature is the emission of a mono-energetic electron of 105 MeV from a muonic atom of aluminum. This electron is detected by a Cylindrical Drift Chamber (CDC) and a set of trigger counters (TC) in a 1 T solenoidal magnetic field.
A high intense muon beam is used to achieve our sensitivity goal. It leads to an unacceptable trigger rate of a few MHz. For stable data acquisition, a trigger system which can reduce it down to a few kHz is required. The total system latency of <5 $\mu$s is also required due to buffer sizes of readout electronics.
In order to fulfill these requirements, we are developing a fast online trigger system using a machine learning based event classification with Field Programmable Gate Arrays (FPGA). This system finds helical electron tracks from the aluminum target. It differs from finding tracks from a point source, and traditional methods such as Hough transform cannot be processed within the required latency. Therefore, we adopt a Gradient Boosted Decision Tree (GBDT) with using multivariate information from CDC; position, energy deposition, and timing. In this system, the trigger electronics collect the hit information from ~5000 wires of CDC and make a trigger decision every 100 ns. For the decisions, look-up-tables inside FPGA convert from it to GBDT outputs within a clock cycle.
From a simulation study, it is found that the classification can reject >90% of background events with a 99% of signal acceptance, which corresponds to the trigger rate of a few kHz in conjunction with the information of TC. The prototypes of trigger electronics were developed, and the total latency was measured to be 2.8 $\mu$s, which meets the requirement. Furthermore, we successfully took cosmic-ray data by using the trigger system installed in a CDC setup. We present these results and prospects.
