DIRECTA (Deep learnIng in REal time for the Cherenkov Telescope Array) is a project funded by the french ANR (Agence Nationale pour la Recherche) for a duration of 3 years. It aims at developing deep learning methods for the real-time analysis of the Cherenkov Telescope Array (CTA) data. The project is led by the LAPP (Laboratoire d’Annecy de Physique des Particules) and involves the LISTIC (Laboratoire d’Informatique, Systèmes, Traitement de l’Information et de la Connaissance) as partner.

Scientific context

Low-energy gamma rays in the 20 GeV to few TeV range are important for probing cosmological distances and exploring new physics. These rays come from transient sources like Active Galactic Nuclei and Gamma Ray Bursts, typically powered by black holes. Current gamma-ray telescopes have limitations in coverage and sensitivity. The project aims to improve the capabilities of Imaging Atmospheric Cherenkov Telescopes using deep learning to analyze these events more effectively.

The real-time analysis is crucial for an efficient observation of transient phenomena, which can unlock a better understanding of cosmology and a sign of new physics. The LAPP is currently in charge of producing the Real-Time Analysis of CTAO.

GammaLearn and the real-time analysis

GammaLearn already provides the means to analyse LST-1 data with increased performances compared to the standard offline analysis pipeline. Applying in real-time would bring the best of both worlds: the speed of the real-time analysis and the performances of GammaLearn.

However, applying it in real-time is a challenge, as the current implementation is not optimized for speed. The goal of this project is to adapt GammaLearn to the real-time analysis pipeline, and to improve its performances to be able to process the data in real-time.


The objectives of this project are:

  • Test the network gamma-PhysNet on real-time.
  • Optimise the network to real-time constraints.
  • The integration into the collaboration real-time pipeline.
  • Study the physics performances in real-time.
  • Analyze sources observed by the LST-1.

Open-source and FAIR Software

The project is committed to Open Science and will make all developed code and models publicly available. It aims to follow FAIR principles to ensure the work is Findable, Accessible, Interoperable, and Reusable. Proper documentation will be provided, and special attention will be given to publishing fully reproducible workflows and results. The goal is not only to ensure the reproducibility of the project’s results but also to contribute to the broader movement for open and transparent scientific research.

The team

Thomas Vuillaume, PI, LAPP

Sami Caroff, Co-PI, LAPP

Alexandre Benoit, Professor, LISTIC

Michaël Dell'aiera, doctorant, LAPP

Cyann Plard, doctorante, LAPP