CUORE Data Release for ML Applications: Pulse Shape Analysis Dataset

Technical Summary: CUORE Data Release for ML Applications

Problem Statement
The CUORE (Cryogenic Underground Observatory for Rare Events) experiment searches for neutrinoless double-beta ($0\nu\beta\beta$) decay of $^{130}\text{Te}$. A significant background source in this search arises from "pile-up" events, where two or more separate particle interactions deposit energy within the same 10-second event window of a single crystal calorimeter. These pile-up events can mimic the energy signature of a single-particle event or distort the background near the expected decay peak ($Q_{\beta\beta}$). While the collaboration has established physics-based methods to identify these events, there is a growing demand for open scientific datasets to develop and benchmark Artificial Intelligence and Machine Learning (AI/ML) algorithms for pulse shape analysis and classification in cryogenic calorimetry.

Methodology and Dataset Construction
This paper presents a curated public dataset derived from CUORE calibration data, specifically designed to support supervised learning approaches for pulse classification. The dataset addresses the binary classification problem of distinguishing between:

• Class 0 (Clean Pulses): Single-pulse events exhibiting the standard thermal pulse shape within a 10-second window.
• Class 1 (Pile-up Pulses): Events containing two or more pulses within the same window, which may be time-separated and have varying amplitudes.

Data Specifications:

• Source: 988 $\text{TeO}_2$ crystal calorimeters operated at $\sim10$ mK.
• Format: HDF5 files containing 10,000 total events (5,000 clean, 5,000 pile-up).
• Waveform Structure: Each event is a 1-dimensional time-series array of length 10,000, sampled at 1 kHz. The window is defined such that the triggered pulse rise midpoint occurs near index 3000, providing $\sim3$ seconds of pre-trigger baseline.
• Labeling: Labels are derived from the number of flagged pulses ($n_{\text{pulses}}$) and the Normalized Reconstruction Error ($n_{\text{rei}}$) from the CUORE 2 tonne$\cdot$yr pulse shape discrimination cut. A label of 0 is assigned if $n_{\text{pulses}}=1$ and $|n_{\text{rei}}| < 10$; a label of 1 is assigned if $n_{\text{pulses}} > 1$ and $|n_{\text{rei}}| > 10$.
• Normalization: To facilitate AI/ML preprocessing, each event includes metadata for normalization: an offset (mean of the first 2,500 samples), a maximum value, and a scale factor. The paper provides Python scripts demonstrating how to construct normalized waveforms $\tilde{p}_i(t)$ and clip them to a $[0, 1]$ range.

Key Contributions

1. Public Dataset Release: The collaboration has released a subset of its data via Zenodo (DOI: 10.5281/zenodo.20721645) in HDF5 format, split into 90% training and 10% test sets.
2. Standardized Benchmarking: The paper proposes a specific pulse classification challenge. Participants are invited to submit predictions on the held-out test set (where labels are withheld) to be evaluated on:
   • Area Under the Receiver Operating Characteristic Curve (AUC-ROC).
   • Classification Accuracy.
   • False Positive Rate (FPR) at a fixed True Positive Rate (TPR) of 90%, a metric relevant for physics analyses requiring controlled signal efficiency.
3. Reproducibility: The authors provide minimal Python examples using h5py and numpy to load, normalize, and format the data for Conv1D neural network inputs.

Results and Claims
The paper does not present the results of a specific machine learning model trained by the authors, nor does it claim a new record in classification accuracy. Instead, the "results" presented are the successful curation and release of the dataset itself. The authors claim that this release serves as an "excellent resource" for the AI/ML community by combining high-quality, time-series cryogenic data with well-understood, physics-based labels.

Significance
The significance of this work lies in its facilitation of open science and the advancement of analysis techniques for rare-event searches. By providing a standardized dataset with clear binary labels (clean vs. pile-up), the CUORE Collaboration enables the testing and development of supervised learning algorithms for pulse shape analysis. This supports the broader goal of improving background rejection in $0\nu\beta\beta$ decay searches and other applications involving cryogenic calorimeters. The authors explicitly request that users cite the relevant CUORE physics results (Ref. [4]) and this specific arXiv paper in any publications resulting from the use of this data.