A New Data Readout Method Based On Machine Learning

Trigger systems are used in particle physics experiments to select interesting events from overwhelming backgrounds and reduce the data rate to an acceptable level.With the deepening research on particle physics, modern collider experiments are operating at higher energy and higher luminosity, which yield an unprecedented amount of science data and bring great challenges to the readout systems. In these experiments,it requires a volatile trigger system and complex trigger algorithm which can hardly be performed with traditional trigger system whose first-level trigger system is implemented in dedicated hardware. When it comes to heavy ion collisions, the collective flow between particles cannot be ignored, whose systematic and complexity may be broken using traditional trigger algorithm. Moreover, the distribution of the global trigger signal is difficult when considering the large number of electronic channels. So trigger-less readout systems are widely used in modern particle physics experiments, which will transfer all the data acquired at the front-end to the back-end through lots of high-speed links. At the back-end, processor farms are used to perform software-based event filter to decrease the data rate. However, the current trigger-less readout system is not the overall optimal, since it will lead to the difficulties in both the high-speed data transmission at the front-end and mass data processing at the back-end.In modern particle physics experiments, it is too hard to obtain an accurate analytic theory to guide the trigger algorithm due to the complexity of the collider interactions.Machine learning is a subfield of artificial intelligence, through which a computer can improve the performance by studying existing experiences. Because of the advantages in solving data-driven problems, machine learning has been adopted in high energy physics to distinguish signals from large backgrounds and has exhibited much better performance.Therefore, based on the concept of trigger-less, a new readout method using machine learning method is proposed in this paper. The method can perform real-time data classification which is entirely based on the generalization ability learned from the characteristics of labeled data. For this reason, it has a strong versatility and can be applied to different physical experiments. With this method, a flexible triggering decision can be achieved to lower the requirements of the readout system without loss of any valid information. After several key points of the method have been discussed,a realization architecture is proposed, which can guarantee real-time classification and data readout. Based on the architecture, a prototype was designed and some preliminary tests were conducted. The test results on both the software and hardware platform verifies the feasibility of the method. This research on using machine learning in the readout system of the particle physics experiments from the global perspective is a useful attempt.