ATLAS, the Asteroid Terrestrial-impact Last Alert System is a NASA-commissioned, University of Hawaii developed system that scans the entire sky each night to give us early warnings on dangerous solar system bodies that might impact the Earth. The reason you’re not a Saurian reading this is because the dinosaurs did not have ATLAS.
This scanning seeks to detect new moving objects (such as ATLAS’ 2025 discovery of the 3rd ever discovered interstellar object ATLAS/3I aka 2025 N1).
To automate these processes across the sky and thousands of possible detections, it’s vitally important to be able to reject bogus objects.
Fugazi is a 2-step bogus object detection machine learning model, rejecting bogus images (via an image classification network) and then taking a second step of classifying tracklets (4 images of same thing which show whether there is movement or not).
Fugazi builds on an initial successful model by moving to a more accurate 2-step model using a CNN for object detection in images using ConvNeXt v2 pico and then an unbalanced categorical CatBoost model to analyze tracklets (time-based quads of images to detect movement) for new-found moving objects.
As of May 2026, we’re now testing this side-by-side with the original produciton model but, in testing, the new model outperformed production on existing datasets in both precisions and recall, as well as being vastly faster due to improvements in the machine learning engineering and better GPU usage.
Much like the COMA work, I’m really enjoying working on this with all its technical and scientific challenges. Was nice to do production ML again after a bit of a break. Very keen to do more work like this and even moar ambitious projects. So, if you found this via search and have interesting Astro/physics problems and projects, are interested in chatting with me about your astro/physics PhD program (I’m actively looking), or happen to be NASA, ESA, JPL, RocketLabs, VG or some other place doing cool space work and throwing objects into the void, mail me. We should definitely chat.