Pattern matching with neural networks for the PANDA at FAIR experiment

The upcoming next-generation antiproton experiment PANDA at FAIR in Darmstadt, Germany will be the first accelerator-based experiment where the data selection relies entirely on a software filter. This paradigm shift is driven by the constantly increasing intensity of accelerators and has a great chance to be adopted by other cutting-edge experiments in the future. The software trigger in our experiment will need to cope with incoming data rates of up to 15 MHz, corresponding to a raw data rate of up to 200 GB/s.

Of particular interest for Uppsala are hyperon reactions. Hyperons are baryons, in which one or several of the light quarks have been replaced with heavier ones. Due to their relatively long-lived nature, their decay vertices can be separated from the beam-target interaction point by up to several metres. This poses a particular challenge on the track and event reconstruction. In order to filter interesting data, the Uppsala group is, in collaboration with other international groups, developing algorithms to reconstruct physical observables from the raw incoming detector signals. One of these algorithms employs a pattern matching technique within one if PANDA's subdetectors.

The goal of this project is to augment the algorithm with modern techniques related to machine/deep learning, i.e. neural networks. This involves the extraction of physical observables, e.g. momenta, from incoming hit and track patterns using Monte Carlo training data. A further requirement would be the implementation of a database through which the training data can be efficiently accessed in a high event rate environment.