ANALYSIS OF LArTPC DATA USING MACHINE LEARNING METHODS

Deep Convolutional Neural Networks (CNNs) have exhibited remarkable efficacy in data analysis across various domains of physical research. Within the realm of particle and high-energy physics, imperative challenge revolves around the scrutiny of particle tracks acquired through track detectors, exemplified by liquefied argon time projection chambers (LArTPCs). These cutting-edge neural networks have showcased unparalleled performance in tackling this intricate task, demonstrating their profound potential in enhancing particle track analysis. Since the scalability requirement is imposed on neutrino detectors, it is also imposed on CNN model implementations. The scalability for LArTPC data processing is achieved due to the sparsity of the data, which have the form of thin trajectories. To process sparse data, subspecies sparse convolutional networks (SSCNs) and sparse tensor networks (STNs) have been proposed. In this paper, we present the results of semantic segmentation of the LArTPC-simulated PILArNET data using STNs and test various modifications of the classical U-Net architecture: Attention U-R,esNet, U-R,esNet3+, U-R,esNet with an additional deep supervision block, as well as loss functions: focal balanced cross-entropy to improve the accuracy of Michel electron identification. The best results were obtained for balanced cross-entropy. Further improvement is possible by combining several methods.