Assessment of few-hits machine learning classification algorithms for low energy physics in liquid argon detectors

TL;DR

This work tackles the challenge of classifying few-hits, low-energy events in liquid argon TPCs by comparing deterministic and machine-learning approaches for separating single-beta from neutrinoless double-beta events. It evaluates a physics-driven blob method, a Convolutional Neural Network, and a Transformer-Encoder, using simulated MeV-scale events to assess performance across readout Granularity and energy-threshold configurations relevant to DUNE Phase II MoO. The results show that both CNNs and Transformer-Encoders surpass the blob baseline in most configurations, with the Transformer offering greater memory efficiency and robustness to overfitting, and the CNN performing best at certain fine-grained readouts. These findings inform detector optimization, indicating that reducing energy thresholds and leveraging ML-based classification may reduce the need for extreme readout granularity, thereby guiding the design choices for large-scale LArTPCs. Overall, ML-assisted β versus ββ separation provides a practical pathway to enhance low-energy LArTPC physics, including potential neutrinoless double-beta decay searches and related backgrounds in the DUNE Phase II program.

Abstract

The physics potential of massive liquid argon TPCs in the low-energy regime is still to be fully reaped because few-hits events encode information that can hardly be exploited by conventional classification algorithms. Machine learning (ML) techniques give their best in these types of classification problems. In this paper, we evaluate their performance against conventional (deterministic) algorithms. We demonstrate that both Convolutional Neural Networks (CNN) and Transformer-Encoder methods outperform deterministic algorithms in one of the most challenging classification problems of low-energy physics (single- versus double-beta events). We discuss the advantages and pitfalls of Transformer-Encoder methods versus CNN and employ these methods to optimize the detector parameters, with an emphasis on the DUNE Phase II detectors ("Module of Opportunity").

Figures (7)

• Figure 1: Workflow of the blob detection algorithm. The three-dimensional profile reconstructed at the LArTPC (top) is transposed into the corresponding graph. The red nodes represent the track endpoints, i.e. the candidate blob position (bottom left). We then integrate hit energies within a radius $r=2$ mm from the endpoints pair to determine $E_{b1}$ and $E_{b2}$ (bottom right).
• Figure 2: Blob candidate energy distributions for the $\beta$ class (top) and the $\beta\beta$ class (bottom). $E_{b1}$ and $E_{b2}$ are extracted from the three-dimensional LArTPC event reconstruction considering a pixel size of $1\times1\times1$ mm${}^3$ and a hit energy threshold of $50$ keV. As expected, the $\beta\beta$ distribution centroid appears closer to the bisector than the $\beta$'s one, allowing for class separation.
• Figure 3: Convolutional Neural Network scheme. A batch of LArTPC events split into three planar views are fed to two independent stacks of convolutional, batch normalization bnorm and dropout dropout layers. The stack outputs merge into a single array, which passes through a fully connected layer. The output layer is a single neuron with a sigmoid activation function, which returns a $[0, \, 1]$ bounded network predictive score. Each convolutional step comprises $25$ filters with $[3\times3]$ dimensions, and all hidden layers are equipped with LeakyReLU activations leakyrelu. Dropout layers were inserted to prevent overfitting of the model.
• Figure 4: Transformer-Encoder scheme. LArTPC events consist of a collection of hits ($h_1\, h_2, ...\, h_n$), where $n$ can change for different events. Every hit comes with four variables (three space coordinates and the hit energy), and the input vector size is $4m$, where $m$ is the largest hit number for the events in the batch. The encoder consists of three stacks of multi-head attention layers and fully-connected layers followed by batch normalization and dropout. Each multi-head step comprises four parallel self-attention heads. The encoder output is mapped into the final prediction, i.e. a single-neuron layer with a sigmoid activation function. All hidden layers are equipped with LeakyReLU activations.
• Figure 5: Learning curves for the Convolutional Neural Network (left) and the Transformer-Encoder (right), considering $w=1$ mm and $E_t=50$ keV. The dashed lines show the performance in the absence of dropout layers, while the solid lines include the effect of dropout layers (inserted accordingly to Fig. \ref{['fig:3']} and \ref{['fig:4']}), with a dropout rate of $0.15$ and $0.02$ for the CNN and the Transfomer-Encoder, respectively. We observe that in both cases the usage of dropout layers mitigates overfitting without compromising the asymptotical validation accuracy.