Compression and Deployment of Neural Networks in Trigger Systems for the Radio Neutrino Observatory in Greenland

The Radio Neutrino Observatory in Greenland (RNO-G) aims to detect ultra high energy (UHE) neutrinos, particles that could potentially carry information about the most exotic and violent phenomena in our universe. The concept of the detector is to observe radio signals in the Greenlandic ice cape, generated by collisions between neutrinos and particles in the ice. Due to thelow flux of UHE neutrinos and the extremely small cross section of neutrino interactions, the expected detection rate once the detector is fully operational is less than once per year. To increase this expected rate, one can design an intelligent trigger system based on a convolutional neural network (CNN) that can separate between possible neutrino signals and useless background noise picked up by the detector’s radio antennas. The CNNs must be deployed onto field programmable gate arrays (FPGAs), chips that allow for custom design after production by the manufacturer, which must filter between noise and signal in real time on the glacial surface of Greenland. Once installed on the detector, the FPGAs will be powered by a small renewable energy source in the form of a windmill or a solar panel. The energy produced by these sources are very limited and must be shared by all the components of a detector station, such as the data transmitter system which can transmit one short recording each second. Therefore, the CNN trigger system must be designed and compressed so that its computations are as fast and use as little energy as possible. This thesis explores several compression techniques for neural networks and demonstrates a complete tool-chain that can take non-experts from a defined neural network structure to hardware description language (HDL) files for deploying the network onto FPGA hardware within days.