2D Convolutional Neural Network on FPGA for High-Resolution TPC Images
Video
Team Information
Team Members
Lukas Arnold, Engineer, Nevis Laboratories, Department of Physics, Columbia University
Yeon-Jae Jwa, PhD Candidate, Nevis Laboratories, Department of Physics, Columbia University
Faculty Advisor: Georgia Karagiorgi, Assistant Professor of Physics, Faculty of Arts and Sciences
Abstract
Low-energy particle detection with Liquid Argon Time Projection Chambers (LArTPCs) requires near-perfect accuracy from LArTPC data selection algorithms, for which Machine Learning algorithms are considered. Convolutional Neural Networks (CNNs), a type of Machine Learning algorithm, are studied for viability for employment at the Deep-Underground Neutrino Experiment (DUNE), a large-scale LArTPC experiment to be built in South Dakota. A CNN was found that both yielded very good physics accuracy, and was found to be well-implementable onto Field-Programmable Gate Array (FPGA) resources available at the experiment. Implementation has been undertaken both for the data preprocessing part, and the CNN itself, onto a demonstration FPGA board. While testing and implementation, such as integrating the different parts of the dataflow chain, are still ongoing, it is found that CNNs are a viable solution for data selection tasks at DUNE.
Team Lead Contact
Lukas Arnold: la2789@columbia.edu
