Artificial Neural Network as a FPGA Trigger for a Detection of Very Inclined Air Showers

TL;DR

This paper presents a FPGA-based artificial neural network trigger designed to detect very inclined air showers caused by neutrinos, enhancing the ability to identify rare neutrino events in the Pierre Auger Observatory.

Contribution

It introduces a novel FPGA implementation of an ANN for real-time detection of inclined air showers, trained on real and simulated data, improving detection capabilities.

Findings

The FPGA ANN achieved good pattern recognition efficiency for inclined showers.

The Levenberg-Marquardt training algorithm was most effective for this application.

The system can distinguish 'old' and 'young' showers with high accuracy.

Abstract

Neutrinos can interact in the atmosphere (downward-going {\nu}) or in the Earth crust (Earth-skimming {\nu}), producing air showers that can be observed with arrays of detectors at the ground. The surface detector array of the Pierre Auger Observatory can detect these types of cascades. The distinguishing signature for neutrino events is the presence of very inclined showers produced close to the ground (i.e., after having traversed a large amount of atmosphere). Up to now, the Pierre Auger Observatory did not find any candidate for a neutrino event. A very low rate of events potentially generated by neutrinos is a significant challenge for a detection technique and requires both sophisticated algorithms and high-resolution hardware. We present a trigger based on a pipeline artificial neural network (ANN) implemented in a large FPGA which after learning can recognize traces corresponding to special types of events. The structure of an ANN algorithm being developed on the MATLAB platform has been implemented into the fast logic of the biggest Cyclone V E FPGA used for the prototype of the Front-End Board for the Auger-Beyond-2015 effort. Several algorithms were tested, however, the Levenberg-Marquardt one seems to be the most efficient. The network was taught: a) to recognize "old" showers (learning on a basis of real very inclined Auger showers (positive markers) and real standard showers especially triggered by Time over Threshold (negative marker), b) to recognize "young" showers (on the basis of simulated "young" events (positive markers) and standard Auger events as a negative reference). A three-layer neural network being taught by real very inclined Auger showers shows a good efficiency in pattern recognition of 16-point traces with profiles characteristic of "old" showers.