Abstract:The Cherenkov Telescope Array is the next generation of observatory using imaging air Cherenkov technique for very-high-energy gamma-ray astronomy. Its first prototype telescope is operational on-site at La Palma and its data acquisitions allowed to detect known sources, study new ones, and to confirm the performance expectations. The application of deep learning for the reconstruction of the incident particle physical properties (energy, direction of arrival and type) have shown promising results when conducted on simulations. Nevertheless, its application to real observational data is challenging because deep-learning-based models can suffer from domain shifts. In the present article, we address this issue by implementing domain adaptation methods into state-of-art deep learning models for Imaging Atmospheric Cherenkov Telescopes event reconstruction to reduce the domain discrepancies, and we shed light on the gain in performance that they bring along.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is: **How to reduce the domain differences between simulated data and actual observed data through deep unsupervised domain adaptation methods, thereby improving the performance of event reconstruction in the Large - Sized Telescope (LST) of the Cherenkov Telescope Array (CTA)**. Specifically, the paper focuses on the domain transfer problem encountered when using deep learning for event reconstruction in very - high - energy (VHE) gamma - ray astronomy. Since the training model depends on simulated data, and there are significant differences between the actual observed data and the simulated data (such as different background light distributions), this leads to a decline in the performance of the model in practical applications. Therefore, the author introduced three unsupervised domain adaptation methods to mitigate these differences, hoping to improve the performance of the model on real data. ### Key Point Summary: 1. **Research Background**: - The Cherenkov Telescope Array (CTA) is the next - generation observation device for very - high - energy gamma - ray astronomy. - Deep learning has shown good event reconstruction performance on simulated data, but its application on actual observed data faces challenges because of the domain transfer problem. 2. **Problem Description**: - The differences between simulated data and actual observed data lead to a decline in the performance of deep - learning models in practical applications. - Domain transfer problems include, but are not limited to: approximations of theoretical models, changes in the Night Sky Background (NSB) distribution, and sparse coverage of the parameter space by training data. 3. **Solution**: - Introduce three unsupervised domain adaptation methods: Domain Adversarial Neural Network (DANN), Deep Joint Distribution Optimal Transport (DeepJ - DOT), and Deep Correlation Alignment (DeepCORAL). - These methods aim to make the model perform well on data from different domains, thereby reducing the negative impact of domain differences. 4. **Experimental Results**: - Through experimental verification, these domain adaptation methods have achieved significant effects in reducing performance degradation, and in some cases, even approach the best - case performance. - Especially in terms of energy resolution and angular resolution, the improvement is particularly obvious. 5. **Future Work**: - The author plans to apply these methods to real - observed data to solve the domain transfer problem between Monte Carlo simulation data and telescope - recorded data. Through these efforts, the paper shows how to use domain adaptation techniques to improve the accuracy of CTA - LST event reconstruction, providing new directions and tools for future research.

Deep unsupervised domain adaptation applied to the Cherenkov Telescope Array Large-Sized Telescope

Deep Learning and IACT: Bridging the gap between Monte-Carlo simulations and LST-1 data using domain adaptation

Deep Domain Adversarial Adaptation for Photon-efficient Imaging

Domain Adaptation in Automatic Picking of Phase Velocity Dispersions Based on Deep Learning

Exploring deep learning as an event classification method for the Cherenkov Telescope Array

Application of Deep Learning Methods Combined with Physical Background in Wide Field of View Imaging Atmospheric Cherenkov Telescopes

Unsupervised Domain Adaptation with Adversarial Learning for Liver Tumors Detection in Multi-phase CT Images

Deep learning techniques for Imaging Air Cherenkov Telescopes

DeepAstroUDA: Semi-Supervised Universal Domain Adaptation for Cross-Survey Galaxy Morphology Classification and Anomaly Detection

Extracting gamma-ray information from images with convolutional neural network methods on simulated Cherenkov Telescope Array data

Domain adaptation in application to gravitational lens finding

Identification of High Energy Gamma Particles From the Cherenkov Gamma Telescope Data Using a Deep Learning Approach

Semi-Supervised Domain Adaptation for Cross-Survey Galaxy Morphology Classification and Anomaly Detection

Unsupervised Domain Adaptation for SAR Target Detection

Unpaired Image Translation to Mitigate Domain Shift in Liquid Argon Time Projection Chamber Detector Responses

Multi-View Deep Learning for Imaging Atmospheric Cherenkov Telescopes

Unsupervised Domain Adaptation for SAR Target Classification Based on Domain- and Class-Level Alignment: from Simulated to Real Data

Domain Adaptation via Minimax Entropy for Real/Bogus Classification of Astronomical Alerts

Unsupervised Domain Adaptation: from Simulation Engine to the RealWorld

Domain Adaptation for Satellite-Borne Multispectral Cloud Detection

Using deep learning methods for IACT data analysis in gamma-ray astronomy: A review