CelestAI

Astronomy's rarest phenomena, including gravitational lensing, kilonovae, and fast radio bursts, occur infrequently, which severely limits observational data and hampers analysis. CelestAI tests whether generative models can produce synthetic yet realistic astronomical images to close that gap.

Galaxy10 DECals, an enhanced version of the original Galaxy10. 17,736 colored galaxy images at 256x256 pixels in g, r, z bands, sourced from the DESI Legacy Imaging Surveys and labeled by Galaxy Zoo volunteers.

Ten morphological classes: Disturbed, Merging, Round Smooth, In-between Round Smooth, Cigar Shaped Smooth, Barred Spiral, Unbarred Tight Spiral, Unbarred Loose Spiral, Edge-on without Bulge, Edge-on with Bulge. Class counts range from 334 to 2,645 images.

Generated synthetic images with each model, added 200 synthetic images to 1,000 real images, and assessed rare-event detection performance using Recall, F1 Score, and ROC-AUC.

Qualitative comparison

Analysis

• Quality vs. efficiency: VQGAN provided the best trade-off between detection performance and computational cost, achieving solid Recall and F1 gains at relatively low compute.
• Metric comparison: DCGAN achieved the highest anomaly detection performance. StyleGAN also showed robust metrics with elevated Recall and ROC-AUC, likely attributable to transfer learning from pre-trained weights.
• Detection improvement: DCGAN led with Recall 0.72 against a 0.57 baseline, a gain of 15 percentage points (26% relative).

CelestAI demonstrates the effectiveness of generative models in synthesizing high-quality astronomical images to address the scarcity of rare astronomical phenomena. Among the four evaluated models, DCGAN emerged as the optimal choice, offering the best balance between image quality and computational efficiency.

• E. E. O. Ishida et al., "Active anomaly detection for time-domain discoveries," Astronomy & Astrophysics, vol. 650, p. A195, Jun. 2021.
• A. Kunwar et al., "Ocular Disease Classification Using CNN with DCGAN," in Advances in Computer Science and Ubiquitous Computing, Springer, 2024.
• R. Gupta, D. Muthukrishna, M. Lochner, "A Classifier-Based Approach to Multi-Class Anomaly Detection for Astronomical Transients," arXiv [astro-ph.IM], 2024.