⚛️ nuclear experiments

Pulse-shape discrimination with machine learning for CZT detectors at the DA $Φ$ NE beam test facility

This study demonstrates that a machine learning-based pulse-shape discrimination method using an XGBoost classifier can effectively distinguish nominal photon signals from anomalous events in CZT detectors under high-flux conditions at the DA $\Phi$ NE facility, significantly improving spectroscopy performance with high accuracy and minimal training data.

Original authors: Simone Manti, Francesco Artibani, Leonardo Abbene, Massimiliano Bazzi, Manuele Bettelli, Giacomo Borghi, Damir Bosnar, Mario Bragadireanu, Antonino Buttacavoli, Mario Carminati, Alberto Clozza, France

Published 2026-09-23

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Simone Manti, Francesco Artibani, Leonardo Abbene, Massimiliano Bazzi, Manuele Bettelli, Giacomo Borghi, Damir Bosnar, Mario Bragadireanu, Antonino Buttacavoli, Mario Carminati, Alberto Clozza, Francesco Clozza, Luca De Paolis, Raffaele Del Grande, Kamil Dulski, Carlo Fiorini, Ivica Friščić, Gaetano Gerardi, Carlo Guaraldo, Mihai Iliescu, Masa Iwasaki, Alexander Khreptak, Johan Marton, Pawel Moskal, Hiroaki Ohnishi, Kristian Piscicchia, Fabio Principato, Alessandro Scordo, Francesco Sgaramella, Michał Silarski, Diana Sirghi, Florin Sirghi, Magdalena Skurzok, Antonio Spallone, Kairo Toho, Oton Vazquez Doce, Andrea Zappettini, Johann Zmeskal, Catalina Curceanu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of nuclear and particle physics, scientists often need to listen to the universe by detecting tiny flashes of light or energy released when atoms interact. To do this, they use detectors that act like highly sensitive ears, capturing signals from X-rays and gamma rays. For decades, a specific type of detector made from a crystal called cadmium zinc telluride has been prized because it can work at room temperature without needing bulky, expensive cooling systems. These detectors are excellent at measuring the energy of incoming particles, which allows researchers to identify exactly what kind of atoms are present. However, these detectors face a significant challenge when placed near powerful particle accelerators. The intense environment creates a chaotic background of noise and false signals that can drown out the faint, important messages the scientists are trying to hear. The signals from unwanted events often look very similar to the real ones, making it difficult to tell them apart using standard methods.

A team of researchers recently tackled this problem by teaching a computer to listen more carefully to the shape of the signals, rather than just their loudness. They set up a test at a facility in Italy where a beam of electrons was fired at a lead target. This collision produced a mix of useful X-rays from the lead atoms and a messy background of other particles and energy bursts. The researchers recorded the electrical waves generated by a detector placed near the target. They noticed that the waves coming from the desired X-rays had a smooth, predictable shape, rising steadily to a peak and then settling down. In contrast, the waves from the background noise were jagged, irregular, or contained multiple bumps, often caused by particles hitting the detector at the same time or by electronic glitches.

To separate the good signals from the bad, the team used a machine learning tool called XGBoost, which is a type of algorithm designed to find patterns in complex data. They first showed the computer thousands of examples of these waves, manually labeling them as either "normal" (the good X-rays) or "anomalous" (the noise). The computer learned to look for specific details in the wave's shape, such as how quickly it rose, how many bumps it had, and how spread out the energy was over time. After training on just a few hundred examples, the computer became very good at the task, eventually reaching an accuracy of nearly 97 percent when tested on new data. It learned to ignore the messy, multi-bump waves while keeping the clean, single-bump waves that contained the scientific information.

When the researchers applied this trained computer model to a massive dataset of over 700,000 recorded events, the results were striking. The model successfully filtered out nearly half of the total data, removing the vast majority of the background noise. What remained was a much cleaner picture of the energy spectrum. The characteristic lines of lead X-rays, which are the primary goal of the experiment, remained bright and clear. Even the distinct signal from electron-positron annihilation, a specific event occurring at a higher energy level, became much easier to see against the reduced background. The double and triple peaks caused by particles hitting the detector simultaneously were almost entirely removed, proving that the computer had learned to recognize the complex signatures of these unwanted events.

This work demonstrates that machine learning can act as a powerful filter for detectors operating in the harsh environments of particle accelerators. By focusing on the detailed shape of the electrical signals, scientists can now recover clear data from experiments that would otherwise be obscured by noise. The study suggests that with only a small amount of initial human guidance, these computer models can be trained to improve the quality of spectroscopy significantly. While the researchers note that further testing is needed to ensure the method works across different conditions and detector setups, the results provide a strong proof of concept. They show that the future of high-precision measurements in crowded accelerator environments may rely less on building better hardware and more on teaching software to understand the subtle language of particle interactions.

Technical Summary: Pulse-Shape Discrimination with Machine Learning for CZT Detectors at the DAΦNE Beam Test Facility

Problem Statement
Cadmium zinc telluride (CZT) detectors are valued for room-temperature X- and gamma-ray spectroscopy due to their high stopping power and compactness. However, their application in high-flux accelerator environments, such as the DAΦNE collider, is limited by intense electromagnetic backgrounds. In these settings, recorded data often contain not only single photon interactions but also multiple energy depositions, pile-up events, charged-particle interactions, and electronic artifacts. These background contributions frequently overlap in energy with signals of interest, rendering simple energy-based rejection ineffective. The challenge is to distinguish "normal" photon-like pulses from "anomalous" events using the information contained within the waveform morphology, which is not fully captured by reconstructed pulse amplitude alone.

Methodology
The study utilized data collected at the DAΦNE Beam Test Facility (BTF) using a 300-MeV electron beam impinging on a 1-mm-thick lead target. This setup generated characteristic Pb X-rays (Lα, Kα, Kβ) alongside a broad background continuum extending to the 511-keV electron-positron annihilation peak. The detection system comprised a quasi-hemispherical CZT crystal (10 × 10 × 5 mm³) coupled to a CAEN VX2740 digitizer (125 MS/s, 16-bit).

The core methodology involved a feature-based Pulse-Shape Analysis (PSA) enhanced by machine learning:

Feature Extraction: From 656-sample waveforms (approx. 5 µs window), a set of 25 physically motivated features was extracted. These quantified amplitude, temporal development, and regularity. Three representative features were highlighted:
- Rise Time: Calculated as $t_{90} - t_{10}$ .
- Number of Peaks: Counted in the first derivative of the waveform (after median filtering).
- DOIRMS: A normalized temporal-width proxy defined as $\sigma_t / (t_{90} - t_{10})$ , characterizing the spread of collected charge relative to the rise time.
Labeling and Training: A subset of 10,000 events was manually labeled as "normal" or "anomalous" based on derivative peak counts and rise time criteria, followed by visual inspection to correct errors.
Classification: An XGBoost classifier (300 boosted decision trees) was trained on 80% of the labeled data and validated on the remaining 20%. The model aimed to maximize the F1-score.

Key Results

Classification Performance: The XGBoost classifier achieved an Area Under the Curve (AUC) of 0.996 and an accuracy of 96.8% on the test set. This significantly outperformed a single-feature threshold approach using DOIRMS alone (AUC 0.85), demonstrating the value of multivariate correlation.
Data Efficiency: The study found that classification accuracy saturated near 97% using several thousand labeled examples, but a substantial performance level (approx. 93%) was reached with only a few hundred labeled events.
Spectral Impact: The trained model was applied to 790,700 events, retaining 48.2% of the dataset. The selection resulted in:
- A substantial reduction of the spectral continuum across the entire energy range.
- Strong suppression of double- and triple-coincidence structures (approx. 150 and 220 keV), consistent with the rejection of pile-up and multi-step waveforms.
- Preservation of characteristic Pb X-ray lines (Lα, Kα, Kβ) and the 511-keV annihilation peak, leading to an improved signal-to-noise ratio.

Significance and Claims
The paper presents a preliminary proof of concept for using machine-learning-assisted pulse-shape discrimination to improve CZT spectroscopy in accelerator environments. The authors claim that:

Feasibility: An interpretable, waveform-based background suppression strategy is feasible for CZT detectors in high-background conditions.
Efficiency: High classification accuracy can be achieved with a relatively small number of manually labeled training examples (a few hundred), reducing the annotation burden.
Utility: The method effectively suppresses background continua and pile-up events while preserving critical spectral features, addressing a key limitation for future precision measurements in projects like EXKALIBUR and SIDDHARTA-2.

The authors maintain a modest tone, noting that the classification metrics reflect agreement with morphology-based labels rather than independently verified interaction topologies. They identify the need for further validation across different beam conditions and detector setups to establish generalizability. Future work suggested by the authors includes exploring one-dimensional Convolutional Neural Networks (CNNs) for direct waveform analysis, integrating these methods into Data Acquisition (DAQ) systems for online event rejection, and utilizing Monte Carlo simulations to generate physically motivated training labels.

More like this