DNN-based Signal Processing for Liquid Argon Time Projection Chambers

TL;DR

The paper tackles robust ROI identification in LArTPC data, where bipolar induction signals and detector effects complicate signal extraction. It introduces a three-channel 2D segmentation approach using a U-ResNet that exploits cross-plane coincidences to distinguish real charge deposits from noise. Across SBND and ICARUS, the method outperforms traditional thresholding in ROI efficiency, purity, and subsequent energy reconstruction, and shows resilience to detector variations, even without extensive augmentation. The approach is already adopted in SBN and offers a foundation for signal processing in future experiments like DUNE, with a publicly released detector-variation dataset to support broader validation.

Abstract

We investigate a deep learning-based signal processing for liquid argon time projection chambers (LArTPCs), a leading detector technology in neutrino physics. Identifying regions of interest (ROIs) in LArTPCs is challenging due to signal cancellation from bipolar responses and various detector effects observed in real data. We approach ROI identification as an image segmentation task, and employ a U-ResNet architecture. The network is trained on samples that incorporate detector geometry information and include a range of detector variations. Our approach significantly outperforms traditional methods while maintaining robustness across diverse detector conditions. This method has been adopted for signal processing in the Short-Baseline Neutrino program and provides a valuable foundation for future experiments such as the Deep Underground Neutrino Experiment.

Figures (5)

• Figure 1: Training scheme using a 3-layer U-ResNet architecture. Input channel images are constructed by applying a low-pass frequency filter and identifying cross-plane coincidences. Example input and target images and waveforms are shown for an elongated track, a case which the signal reconstruction is challenging.
• Figure 2: Image augmentations used to mimic detector variations, exaggerated for clarity. From left to right: nominal image, image with masked wires, image with tick-direction smearing, image with pixel-wise random scaling, and image with event-wise random scaling.
• Figure 3: ROI efficiency and purity for track-like charge depositions in ICARUS, shown as a function of the track angle relative to the drift field ($\hat{x}$). Results are based on 10,000 test samples, with a charge threshold of $10^4$e$^-$. Right: Reconstruction performance for BNB-like electromagnetic showers in SBND, evaluated on 20,000 test samples.
• Figure 4: ROI Efficiency$\times$Purity for plane 0, Right: for plane 1 in ICARUS, from simulated neutrino interactions in BNB with simulated cosmic rays. Results are shown under different detector variations, comparing traditional ROI and DNN ROI models.
• Figure 5: Left: DNN ROI score prediction from a network trained with augmentation, Right: without augmentation, applied to an SBND event. The gray band indicates a detector dead region.