Astronomy has entered the era of Big Data. Observatories like the LSST (Vera C. Rubin Observatory) will generate nearly 20 terabytes of data per night. Manually analyzing these mountains of information is impossible. AI, and more specifically Machine Learning and Deep Learning, have become essential tools for sorting, classifying, and discovering hidden phenomena in these data streams.
Historically, discoveries relied on patient observation and manual calculation. Today, algorithms trained on millions of images can identify galaxies, supernovae, or asteroids in a fraction of a second, often with greater accuracy than the human eye.
Supervised machine learning techniques allow models to be trained to recognize specific features in light spectra or light curves. For example, the detection of Type Ia supernovae, crucial for measuring the expansion of the Universe, is now largely automated thanks to these algorithms.
The transit method, which detects drops in a star's brightness as a planet passes in front of it, has discovered thousands of exoplanets. However, the signals are tiny and often drowned out by instrumental noise or stellar variations.
Neural networks, such as those used by the Kepler and TESS missions, filter this noise much more effectively than traditional statistical methods. They learn to distinguish a real transit from a natural variation in the star's brightness. The TRAPPIST-1 star system, with its seven rocky planets, has particularly benefited from these advanced analyses to confirm complex orbital periods.
N.B.:
The AstroNet algorithm, developed by Google, discovered two new exoplanets (Kepler-90i and Kepler-80g) by reanalyzing old Kepler data, demonstrating the power of AI to find what humans had missed.
The morphology of galaxies reveals crucial information about their history and evolution. The Galaxy Zoo project, launched in 2007, initially mobilized hundreds of thousands of volunteers to manually classify galaxy images. Today, convolutional neural networks accomplish this task with comparable or even superior accuracy in seconds.
Algorithms can distinguish between spiral, elliptical, and irregular galaxies, and even identify fine structures such as galactic bars or spiral arms. This automation allows the processing of catalogs containing billions of galaxies, paving the way for statistical studies of unprecedented scale on the formation and evolution of cosmic structures.
| Application Area | AI Technique | Typical Accuracy | Time Saved |
|---|---|---|---|
| Exoplanet detection (transits) | Convolutional neural networks | 95-98% | × 1000 |
| Galaxy classification | Deep Learning (CNN) | > 95% | × 10,000 |
| Supernova detection | Supervised learning | 90-95% | × 100 |
| Automatic spectral analysis | Recurrent networks (RNN) | 92-96% | × 500 |
| Search for gravitational lenses | Computer vision | 88-93% | × 5000 |
| Reconstruction of astronomical images | Autoencoders / GANs | 90-97% | × 2000 |
| Space weather forecasting (solar winds, CME) | Recurrent networks / LSTM | 85-92% | × 500 |
| Identification of comets and asteroids | CNN + supervised learning | 90-96% | × 1000 |
| Detection of gamma-ray bursts | Deep learning on time series | 93-97% | × 300 |
| Mapping of interstellar dust clouds | Autoencoders + segmentation | 88-94% | × 2000 |
| Optimization of observation scheduling | Reinforcement learning | 80-90% | × 50 |
| Analysis of stellar brightness variations (variability) | RNN / Transformer | 91-95% | × 400 |
Source: Compiled from Baron (2019), arXiv:1904.07248 and Fluke & Jacobs (2019), WIREs Data Mining and Knowledge Discovery.
AI does more than analyze: it predicts and simulates. Cosmological models simulating the formation of large-scale structures in the Universe (filaments, galaxy clusters) are extremely computationally intensive. GANs can now generate realistic simulations in record time, allowing researchers to test thousands of cosmological parameters \( \Omega_m, \sigma_8 \) and compare them with observations.
It also helps predict the behavior of variable objects such as fast radio bursts (FRB) or stellar flares, by identifying precursor patterns in historical data.
N.B.:
Convolutional neural networks (CNN) are particularly suited to the analysis of astronomical images because they can automatically learn relevant features at different spatial scales, from individual pixels to large-scale structures, without requiring manual programming of detection filters.
The main challenge remains the interpretability of AI models. A neural network can identify an anomaly, but does not always provide a clear physical explanation. XAI is a crucial research field for astronomy.
In the future, AI will be integrated directly into telescopes for real-time processing, deciding for itself to point towards rare transient events, such as gravitational waves or kilonovae. Projects like the ZTF (Zwicky Transient Facility) and soon the LSST at the Vera C. Rubin Observatory already use algorithms to analyze data on the fly and trigger automatic alerts within minutes.
N.B.:
A kilonova is a transient explosion resulting from the merger of two compact objects (neutron stars or a black hole and a neutron star). First detected in 2017 during the GW170817 event, its brightness fades rapidly over a few days, requiring almost instantaneous observation.
AI, the Astronomer's New Eye: From Analysis to Cosmic Prediction 
What the First Images from Perseverance Show Us: The Surface of Mars, As If We Were There 
GRAIL Twin Probes: Mapping the Moon's Hidden Interior 
Earth Under Surveillance: Observation Satellites 
March 2010: The Ring of Fire Captured by the SDO Observatory 
Positions of Space Probes in 2025 
Giants of Observation: The Largest Ground-Based Telescopes 
Low Earth Orbits and their uses 
Pioneer, first message to extraterrestrials ! 
How to see infrared images from JWST? 
Sputnik 1 and 2: The Dawn of the Space Age 
ENVISAT: Ten Years of Uninterrupted Observation of Earth's Ecosystems 
Lagrange Points: Gravitational Gateways of the Solar System 
Mars Reconnaissance Orbiter: The Lynx Eye Revealing Martian Secrets 
Kepler: 4,000 Worlds and More, A New Map of the Sky 
Why Measure Space at the Nanometer Level? 
High-risk landing for Curiosity in 2012 
Cheops Space Telescope: A New Vision on Exoplanets 
The world of Planck 
Rosetta Space Probe: Comet Churyumov-Gerasimenko 
The Hubble Space Telescope: Three Decades of Revelations 
Première photo du télescope Hubble 
Satellites that measure underwater relief 
MESSENGER: The First Probe to Visit the Mysterious Planet 
How GPS Locates Your Position at Any Time? 
The ISS and Beyond? Towards the End of a Space Chapter 
Space Telescopes: Humanity's Eye Beyond the Atmosphere 
Space Probes: Interplanetary Travelers and Beyond
The GAIA satellite maps the Milky Way 
How to calculate the synchronous orbit? 
E-ELT: The World's Largest Optical Telescope 
Mercury probes 
Space Debris: The Nightmare of Modern Satellites 
Aquarius: A Mission to Map Ocean Salinity 
JWST: An Unprecedented Look at the First Lights of the Universe 
METEOSAT: A Key Satellite for Climate Monitoring 
Curiosity, the first shovel, sample of Martian soil 
From Mariner to Perseverance: Successes and Failures of Mars Probes 
Where is the geostationary orbit? 
MOM, the technological demonstration 
Venus under surveillance: an overview of space probes 
What Is an Astronomical Interferometer?
The Philae Robot and the Rosetta Comet 
Mauna Kea Under the Stars: The CFHT Telescope in Pursuit of the Universe's Mysteries