Real-time Anomaly Detection for Liquid Argon Time Projection Chambers

TL;DR

Real-time anomaly detection for LArTPCs is addressed by a two-stage framework: a high-capacity Teacher autoencoder provides unsupervised reconstruction-based anomaly scores, which are then distilled into a lightweight Student network for efficient hardware deployment. The Student reproduces the Teacher's outputs with markedly fewer parameters, enabling low-latency inference on FPGAs and CPUs and enabling online triggering and data filtering. Evaluations on the MicroBooNE Open Dataset show the Teacher's scores correlate with event topology (notably multi-track events), and the Student retains this sensitivity while delivering a fixed-point, hardware-friendly output; ROC-AUC reaches up to 0.934 for 864×64 segments. Hardware synthesis demonstrates functional equivalence with the original model and sub-5 ns inference on FPGA targets, highlighting the potential for AI-assisted data selection in future LArTPC experiments and defining a path toward robust, real-time anomaly-based triggers.

Abstract

We present a real-time anomaly detection framework for liquid argon time projection chambers (LArTPCs), targeting applications in particle physics experiments such as the Short Baseline Near Detector (SBND) or the future Deep Underground Neutrino Experiment (DUNE). These experiments employ detectors that generate and stream high-resolution but sparse images of neutrino and other particle interactions. Our approach utilizes anomaly detection with autoencoders, compressed through knowledge distillation (KD), to enable the detection of anomalous signals in the data through efficient inference on resource-constrained hardware. The framework is targeted for deployment on computing platforms equipped with field-programmable gate arrays (FPGAs), GPUs, or CPUs, allowing low-latency selection of relevant activity directly from the raw detector data stream. We demonstrate that our approach is suitable for the detection and localization of anomalously "high-multiplicity" activity, and outline promising applications for LArTPC online data filtering and triggering.

Figures (17)

• Figure 1: Example of a pixel intensity distribution for a single input image after downsampling and before thresholding.
• Figure 2: Example of a preprocessed MicroBooNE input image after downsampling and pixel intensity thresholding. The image is divided into smaller segments for network input. White dotted lines indicate a division into $864 \times 64$-pixel segments. The energy depositions have been artificially diluted by a $3\times3$ radius for a better visualization.
• Figure 3: Architectures of the Teacher autoencoder (a, b) and the quantized Student network (c), showing each layer type and activation function, with values used for quantization. Actual layer shapes are scaled corresponding to the input image size.
• Figure 4: Schematic overview of the knowledge distillation process. The Teacher network is an autoencoder, which computes a reconstruction loss (anomaly score) for each input image. The Student network is trained to reproduce this score directly, enabling lightweight inference pol2023kd.
• Figure 5: Histogram of Teacher anomaly scores across the dataset for different segment sizes. For each of the three cases, high anomaly scores are less frequent, as expected.