AstroLens – AI that watches the sky and finds what nobody catalogued

Name: AstroLens – AI that watches the sky and finds what nobody catalogued
Availability: InStock
Author: samantaba

by samantaba·Feb 20, 2026·1 point·0 comments

Visit Project View on HN

AI Analysis

●●●BangerZero to OneWizardryBig Brain

AI autonomously rediscovered known supernovae and gravitational lenses across 21k sky survey images.

Strengths

•Proven validation: independently recovered SN 2014J, NGC 3690, gravitational lenses—real astronomy, not toy results.
•Multi-source ensemble (Vision Transformer + OOD + YOLOv8) reduces false positives and scales to streaming discovery.
•True autonomy: 3-day runs with self-correction, cross-referencing SIMBAD/NED/VizieR, zero human intervention required.

Weaknesses

•Narrow audience—requires domain expertise in astronomy and infrastructure for multi-day compute jobs.
•Early-stage: single validation run shown; unknown how results generalize across different sky regions or survey depths.

Post Description

# Show HN: AstroLens -- AI that watches the sky and finds what nobody catalogued

*https://github.com/samantaba/astroLens** (MIT licensed, Python)

AstroLens is an open-source tool that downloads images from sky surveys (SDSS, ZTF, DECaLS, Pan-STARRS, Hubble, and others), runs them through a Vision Transformer + out-of-distribution ensemble + YOLOv8 pipeline, computes galaxy morphology, and cross-references everything against SIMBAD/NED/VizieR. It's designed to run autonomously for days.

*Results from a 3-day validation run* (zero human intervention):

published results in https://www.linkedin.com/pulse/astrolens-v110-teaching-ai-wa...

- 20,997 images from 7 sources analyzed - 3,458 anomaly candidates across 354 sky regions - Independently recovered SN 2014J (Type Ia supernova in M82), NGC 3690 (galaxy merger), and SDSS J0252+0039 (gravitational lens) - YOLO transient detection went from 51.5% to 99.5% mAP50 by training on data collected during the run itself - 140 self-correction cycles, zero errors

*What makes it interesting*: The pipeline is self-correcting — it adjusts OOD thresholds, rebalances survey sources based on anomaly yield, recalibrates its reference distributions, and handles errors autonomously. It's not a batch job; it's a continuous system that gets better as it runs.

*Honest limitations*: It found known objects, not new discoveries — this validates the pipeline but the real test is pointing it at less-explored regions. OOD detection on astronomical images is inherently noisy (the boundary between "unusual galaxy" and "imaging artifact" is fuzzy). The self-correcting system helps, but false positives remain a challenge.

Runs on a laptop (CPU/MPS/CUDA). Desktop app, web UI, CLI, Docker.

Built with Python 3.10+, FastAPI, PyTorch, Ultralytics, PyQt5.