Separation of track- and shower-like energy deposits in ProtoDUNE-SP using a convolutional neural network

Separation of track- and shower-like energy deposits in ProtoDUNE-SP using a convolutional neural network - Abed Abud, Adam et al - arXiv:2203.17053FERMILAB-PUB-22-240-AD-ESH-LBNF-ND-SCDCERN-EP-2022-077

A 7\,GeV/$c$ beam $\pi^+$ interaction in the collection view (W-view) in ProtoDUNE-SP data. The x axis shows the wire number. The y axis shows the time tick in the unit of 0.5 $\mu$s. The colour scale represents the charge deposition.

Examples of CNN input patches from a simulated ProtoDUNE-SP event. The inputs to the CNN are small $48\times 48$ pixel images created from patches of the full detector readout. Three examples are shown, each labelled with their appropriate class. The patch of the detector readout from which each patch was generated is emphasised.

The CNN architecture. In this case, the CNN processes 256 images in parallel. Each image is a $48\times 48$ pixel patch of the calibrated detector readout. A single convolutional layer, with 48 filters of size $5\times 5$, is used to extract features from the images. These are processed by two dense layers containing 128 and 32 neurons respectively, before being split into two branches which provide the track-shower-empty (TSE) and Michel outputs. The dimensions of the data after each operation are given next to the black arrows.

Evolution of the training and validation losses as a function of training epoch. The final weights of the network were taken from a checkpoint at the end of the fifth epoch, shown here as a vertical line. The overall loss; track, shower and empty loss; and Michel loss are shown in the top left, top right, and bottom left respectively. In calculating the overall loss, the track, shower and empty loss is weighted by 0.1 to be consistent with the smaller size of the Michel sample.

Shower classifier output distributions. The output of the shower classifier is shown for true shower hits in red and all other hits in blue. The blue line shows the F1 score as a function of classification threshold.

ROC curves for the shower classifier, showing the true positive rate against false the positive rate for varying classification threshold on the shower classifier output. The red (blue) line shows the ROC curve from ProtoDUNE-SP simulation with (without) SCE. The red curve is obscured by the blue due to close agreement.

Michel electron classifier output distributions.

The CNN shower score of each hit in reconstructed particles for the same 7\,GeV/$c$ $\pi^+$ interaction shown in Fig.~\ref{fig:r5815_e962}. This diagram shows the location of reconstructed hits in wire--time coordinates, and the hits are coloured based on the CNN shower score. Red hits are track--like, and blue hits are shower--like. A number of cosmic muon tracks can be seen, along with tracks and showers produced by the pion interaction. The small shower--like patches along the muon tracks are delta-ray electrons.

The CNN shower classifier scores for cosmic-ray muon hits from experimental data (black) and simulation (red). The error bars on the data are statistical.