🔭 astrophysics

Machine Learning for Event Reconstruction in Imaging Atmospheric Cherenkov Telescopes

This chapter reviews the critical role of machine learning in reconstructing particle properties for Imaging Atmospheric Cherenkov Telescopes, covering the standard supervised learning pipeline, recent innovations in timing-based features and ensemble methods, and the emerging potential of deep learning architectures.

Original authors: Antonio Pagliaro, Antonino La Barbera

Published 2026-06-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Antonio Pagliaro, Antonino La Barbera

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

Imagine the night sky as a massive, dark ocean. Most of the time, it's calm, but occasionally, a giant wave crashes down from space. These "waves" are high-energy particles, mostly harmless cosmic rays (like protons) that constantly rain down on Earth. But every now and then, a rare, precious "gold coin" falls with them: a gamma ray from a distant black hole or exploding star.

The problem? For every one gold coin, there are thousands of rocks. If you try to pick them up by hand, you'll spend your whole life sifting through rocks and never find the gold.

This is the job of Imaging Atmospheric Cherenkov Telescopes (IACTs). They are giant, high-tech mirrors on the ground that don't look at the stars directly. Instead, they watch the "splash" these particles make when they hit Earth's atmosphere.

Here is a simple breakdown of how the paper explains using Machine Learning to find that gold.

1. The Setup: The Atmosphere as a Giant Camera

When a particle hits the atmosphere, it creates a massive shower of smaller particles, like a pebble hitting a pond and sending out a ripple. This shower moves faster than light can travel through the air, creating a faint, blue flash of light (Cherenkov radiation) that lasts for a billionth of a second.

The telescopes are like giant cameras trying to snap a photo of this flash.

The Gold Coin (Gamma Ray): Creates a clean, smooth, oval-shaped ripple. It's orderly.
The Rocks (Cosmic Rays): Create messy, jagged, chaotic splashes. They are full of holes and irregular shapes.

The Challenge: The "rocks" are so common that they drown out the "gold coins." The telescope sees a messy photo and has to guess: "Is this a gold coin or just a rock?"

2. The Old Way: The "Shape Detective"

For a long time, scientists used a method called Hillas Parameterization. Think of this as a detective looking at a fingerprint and measuring just a few things:

How long is the oval?
How wide is it?
How bright is the center?

They would draw a line on a graph: "If the oval is this wide and this long, it's a rock. If it's this narrow, it's gold." This worked okay, but it was like trying to identify a person by only measuring their height and weight. You miss a lot of details.

3. The New Way: The "Super-Brain" (Machine Learning)

The paper explains that we now use Machine Learning (ML) to act as a super-detective. Instead of just measuring a few lines, the computer looks at the entire picture, pixel by pixel, and learns to spot patterns humans can't see.

The paper highlights three major upgrades to this "Super-Brain":

A. Listening to the "Sound" of the Flash (Temporal Features)

The old cameras only took a picture of how bright the flash was. The new approach listens to when the light hits each part of the camera.

The Analogy: Imagine a wave hitting a beach. A smooth wave (gamma ray) hits the sand in a perfect, rolling line. A chaotic splash (cosmic ray) hits the sand in a jumbled, scattered mess.
The Result: By adding "time" to the data, the computer can tell the difference much better, especially for the faint, low-energy flashes that were previously impossible to distinguish. It's like upgrading from a black-and-white photo to a high-speed video.

B. The "Panel of Experts" (Ensemble Learning)

Instead of relying on just one smart computer program, the paper suggests using a Stacking Ensemble.

The Analogy: Imagine you have a panel of experts trying to identify a suspect.
- Expert A is great at spotting low-energy events but bad at high-energy ones.
- Expert B is great at high-energy but makes mistakes on low-energy.
- Expert C is good at everything but slightly biased.
The Magic: A "Manager" (the meta-learner) listens to all three. If Expert A says "Rock" and Expert B says "Gold," the Manager knows how to weigh their opinions to get the right answer.
The Result: This method fixes a major problem where computers used to guess the energy of the particles slightly wrong (a "bias"). By combining experts with opposite biases, the final guess becomes incredibly accurate.

C. The Future: Deep Learning (The "Raw Data" Artist)

The paper looks ahead to Deep Learning (like Convolutional Neural Networks).

The Analogy: The old method was like asking a human to describe a painting by listing the colors and shapes. The new method is like handing the painting directly to an AI that has seen millions of paintings. The AI doesn't need a human to tell it what to look for; it learns the patterns itself directly from the raw pixels.
The Goal: This allows the computer to find subtle clues that human-designed rules might miss.

4. The Results: Why It Matters

The paper uses data from a test array called ASTRI Mini-Array to prove these ideas work:

Better Separation: Adding the "time" features improved the ability to spot gold coins by about 11%.
Low Energy Boost: For the faintest, hardest-to-see flashes (low energy), the improvement was huge—up to 50%. This is crucial because the most interesting cosmic events often happen at these lower energies.
Zero Bias: The "Panel of Experts" method almost completely eliminated the error in measuring the energy of the particles. This means scientists can trust the measurements to study the physics of the universe without worrying the computer is "lying" about the numbers.

Summary

The paper argues that to understand the universe's most violent events, we need to stop looking at the sky with simple rulers and start using "Super-Brains." By teaching computers to look at the shape, the timing, and to listen to a team of different algorithms, we can filter out the noise of the universe and finally hear the clear signal of the rare, powerful gamma rays.

The future isn't just about building bigger telescopes; it's about building smarter software to make sense of the data they collect.

Technical Summary: Machine Learning for Event Reconstruction in Imaging Atmospheric Cherenkov Telescopes

Problem Statement
Imaging Atmospheric Cherenkov Telescopes (IACTs) are the primary instruments for observing Very High Energy (VHE) gamma rays (hundreds of GeV to hundreds of TeV). The fundamental challenge in IACT data analysis is the extreme signal-to-background imbalance: for every gamma-ray event, telescopes detect hundreds to thousands of background events initiated by hadronic cosmic rays (protons and nuclei). While gamma-ray showers produce compact, regular electromagnetic cascades, hadronic showers are chaotic, producing irregular, extended images with distinct temporal structures. Additionally, cosmic-ray electrons and positrons produce electromagnetic showers indistinguishable from gamma rays on an event-by-event basis, requiring statistical background subtraction. The core problem is to reconstruct the physical properties of the primary particle (energy and arrival direction) and perform highly efficient "gamma/hadron separation" (classification) despite the stochastic nature of air shower development and the high-dimensional, noisy nature of camera data.

Methodology
The paper outlines a progression from traditional reconstruction pipelines to modern machine learning (ML) approaches, utilizing data from the ASTRI Mini-Array and other major observatories (H.E.S.S., MAGIC, VERITAS, CTAO).

Standard Pipeline & Feature Engineering:
- Image Cleaning: Raw pixel signals are calibrated to photoelectrons and cleaned using two-level thresholds to remove Night Sky Background (NSB) noise.
- Hillas Parameterization: The cleaned images are reduced to a set of geometric and photometric parameters (e.g., Size, Length, Width, Alpha, Distance, Conc) derived from the moments of the light distribution.
- Stereoscopy: Data from multiple telescopes are combined to resolve the "head-tail" ambiguity of single images, reconstructing the shower core, arrival direction, and maximum development height ( $H_{max}$ ).
- Baseline ML: These parameters serve as inputs for supervised learning. The Random Forest (RF) algorithm has historically been the workhorse, utilizing bagging and random subspace methods to create robust ensembles of decision trees.
Advanced Feature Engineering (Temporal Dimension):
- The paper introduces a "lexicon" of temporal parameters derived from the precise arrival times of photons in each pixel.
- Key parameters include Time RMS (temporal spread), Time Gradient (sweep speed across the camera), Lacunarity (temporal inhomogeneity/gappiness), and Temporal Coherence.
- These features exploit the physical difference that gamma-ray showers evolve coherently along an axis, while hadronic showers are fragmented by muons and sub-showers.
Advanced Ensemble Learning:
- The authors explore moving beyond single RF models to Gradient Boosting (XGBoost, LightGBM) and Extra Trees.
- Stacking Ensembles: A meta-learning approach is employed where diverse base models (RF, XGBoost, Extra Trees) are trained, and a second-level "meta-learner" combines their predictions. This is specifically applied to mitigate systematic energy bias, where different models exhibit opposing biases (over- vs. under-estimation).
Feature Importance Evaluation:
- The paper critically evaluates methods for determining feature importance. It argues that native metrics (Gini importance) and permutation importance can be misleading due to correlations and out-of-distribution artifacts.
- Ablation Studies are advocated as the rigorous standard: retraining the model without a specific feature to measure the direct drop in performance ( $\Delta Q$ ).
Deep Learning Frontiers:
- The paper surveys the shift toward end-to-end deep learning, which bypasses manual feature engineering.
- Convolutional Neural Networks (CNNs): Process raw 2D pixel images (and potentially time channels) to learn spatial patterns directly.
- Recurrent Neural Networks (RNNs/LSTMs): Model the time-ordered sequence of pixel triggers.
- Graph Neural Networks (GNNs): Treat the telescope array as a graph to handle variable numbers of triggered telescopes in stereoscopic reconstruction.
- Unsupervised Approaches: Autoencoders are mentioned for anomaly detection, training on background data to identify gamma-ray outliers without explicit gamma-ray labels.

Key Contributions and Results

Temporal Features: The inclusion of temporal parameters in Random Forest classifiers for the ASTRI Mini-Array increased the Quality Factor (QF) for gamma/hadron separation from 3.32 to 3.68 (~10.8% gain). Crucially, this improvement is most pronounced at low energies (< 1.5 TeV), where traditional morphological parameters lose discriminatory power, potentially improving sensitivity by up to 50% in this regime.
Ablation Insights: Ablation studies revealed that feature importance is model-dependent. For instance, while Width is dominant for Random Forests, Distance becomes critical for LightGBM. The study also showed that derived parameters like DENS are often redundant, and Conc can be negligible or harmful in simulation data.
Bias Mitigation via Stacking: A stacking ensemble (45% XGBoost, 27.5% Extra Trees, 27.5% RF) successfully balanced the opposing energy biases of its constituent models. This reduced the systematic energy bias to near-zero across the 2–100 TeV range, a critical step for precision spectral measurements.
Performance Gains: A stacked ensemble using morphological and stereo features achieved a QF of 6.74, a 13.5% improvement over a baseline Random Forest (QF 5.94).

Significance and Outlook
The paper posits that machine learning has evolved from a supplementary tool to the central framework of IACT analysis. The significance of these developments lies in two areas:

Scientific Reach: By improving low-energy sensitivity through temporal features, next-generation instruments can access transient phenomena and sources with spectral cutoffs below 1 TeV.
Precision Measurement: By using ensemble methods to self-correct systematic biases, ML transitions the field from mere detection to precision spectroscopy, essential for identifying spectral breaks and cutoffs.

The authors conclude that the future of IACT analysis, particularly for the Cherenkov Telescope Array Observatory (CTAO), will be dominated by deep learning paradigms (CNNs, GNNs) that learn features directly from raw data. However, they emphasize that the "domain shift" between Monte Carlo simulations and real data remains the most critical challenge to address before these advanced models can be fully deployed in production pipelines. The move toward open data and standardized software (ctapipe, Gammapy) is highlighted as essential for fostering reproducible science and cross-disciplinary innovation.