On February 25, 2026, the Vera C. Rubin Observatory in Chile sent its first batch of real-time alerts — 800,000 of them in a single night. Each one flagged something that changed in the sky: a new point of light, a brightening star, a moving dot that might be an asteroid. The Legacy Survey of Space and Time (LSST) had officially started talking.

That was the gentle version. At full survey depth, Rubin will generate up to 10 million difference-image alerts every night. No research group, no observatory control room, no grad student with a caffeine problem can review that by hand. The only reason the alerts are useful at all is a network of nine community software platforms — called alert brokers — that run machine-learning classifiers on every packet before the Sun comes up.

What an alert packet contains

Rubin detects changes through difference imaging. The telescope takes a new exposure, the pipeline subtracts a deep reference image of the same sky patch, and anything left over — any residual flux above the noise — triggers an alert. Each alert packet ships with three small postage-stamp cutouts: the new image, the reference, and the difference. It also includes time-series photometry (the object’s brightness history across prior visits), positional data, and cross-matches against known astronomical catalogues.

The latency target is 60 seconds from shutter close to alert distribution. That constraint shapes every ML design decision downstream. You can’t run a model that takes 10 minutes per stamp when millions of stamps are queued behind it.

The broker network

Rubin doesn’t classify alerts itself. It distributes the raw alert stream to nine officially selected community brokers, each built by a different team with a different ML stack and a different scientific emphasis:

  • ALeRCE (Chile) runs a stamp classifier CNN for real/bogus filtering, plus a light-curve classifier using balanced random forests. It maintains a web explorer where you can browse recent transient candidates by class, brightness, and sky position.
  • Fink (France) uses a modular architecture with pluggable science modules. Feature engineering feeds gradient-boosted trees and neural nets for transient typing. Open-source and designed so anyone can write a custom classifier module.
  • Lasair (UK) focuses on cross-matching and user-defined SQL filters. You write a query — “show me events brighter than magnitude 19, detected in the last 48 hours, within 5 arcseconds of a known galaxy” — and Lasair streams matches to your inbox.
  • ANTARES (US, NOIRLab) runs a feature extraction pipeline feeding an anomaly detection engine. It’s particularly strong at surfacing rare, unclassified events that don’t fit standard transient categories.
  • AMPEL (Germany) is more a toolkit than a single classifier: a pipeline framework that lets researchers define multi-stage filtering and analysis channels for their specific science case.
  • Pitt-Google (US) is cloud-native, streaming alerts into Google Cloud Pub/Sub for researchers who want to plug Rubin data into their own cloud ML workflows.
  • Babamul, SNAPS, and POI Broker round out the network with their own classification schemes and community focus areas.

The diversity is deliberate. Different science questions need different classifiers. A supernova cosmology team wants Type Ia candidates ranked by redshift and host-galaxy properties. An asteroid survey team wants fast-moving objects flagged before they leave the field. A gravitational-lensing group wants anomalous brightening events near known lens systems. No single ML pipeline optimises for all of these at once, so the broker ecosystem lets each team subscribe to the filter that fits.

The two-stage ML cascade

Most brokers run a similar two-stage process, even though the specific architectures differ.

Stage 1: real or bogus. The first filter separates genuine astrophysical sources from imaging artifacts — cosmic rays, satellite trails, bad pixel columns, optical ghosts, diffraction spikes from bright stars. Convolutional neural networks trained on labelled alert stamps handle this. Heritage models from the Zwicky Transient Facility (ZTF), which ran a similar alert system at lower volume, hit ~97% accuracy on ZTF data. But Rubin sees about 2 magnitudes fainter and covers a wider field, so the artifact population looks different. Transfer learning — fine-tuning ZTF-trained models on early Rubin data — is the most common approach right now. Domain adaptation and self-supervised pretraining on unlabelled Rubin stamps are catching up.

Stage 2: transient classification. Once an alert passes the real/bogus cut, the brokers try to label it. Is this a Type Ia supernova? A core-collapse supernova? A tidal disruption event? A cataclysmic variable? An AGN flare? An asteroid? The classifiers are more varied here: recurrent neural networks that consume the light curve as a time series, transformer architectures like the Tiny Time-series Transformer optimised for low-latency inference, and feature-based methods that combine light-curve shape parameters with host-galaxy photometric redshifts.

The practical goal: assign a confident label fast enough to trigger follow-up observations — spectroscopy on a 4-metre or 8-metre telescope, for instance — before the transient fades below detection. For a supernova, that window can be days. For a tidal disruption event, it might be hours.

What amateurs can do with this

Rubin’s alerts are public. You don’t need a university affiliation or a data-access agreement to subscribe through the brokers. ALeRCE’s web explorer lets you browse classified transients with a few clicks. Fink exposes a public API. Lasair lets registered users write SQL filters and receive email notifications when something matches.

Rubin also partners with Zooniverse for citizen science — human classifiers reviewing edge cases where the ML isn’t confident. If you’ve used Galaxy Zoo, the workflow is familiar.

There’s a more direct path too: follow-up photometry. When a broker flags a bright transient — say, magnitude 15 or 16 — amateur telescopes can contribute useful data by monitoring the light curve over subsequent nights. Smart scopes like the Seestar S50 won’t resolve spectra, but they can track brightness changes in 30-second stacks. I’ve been setting up ALeRCE filters for bright candidates visible from Nicosia, and the lag between a broker classification and a useful amateur observation is now measured in hours. It used to be weeks, back when discoveries came through IAU circulars and amateur email lists.

The ramp-up ahead

The survey is still early. Alert volume will climb as the camera covers more sky and the reference images deepen. A January 2026 white paper from the LSST Dark Energy Science Collaboration mapped AI/ML needs across the entire DESC workflow — photometric redshifts, weak-lensing inference, the alert brokers. The paper’s core argument: ML for Rubin isn’t optional infrastructure. It’s load-bearing.

The brokers are already retraining on real Rubin data. Models tuned on simulations and ZTF transfers are being replaced by classifiers that have seen actual Rubin noise, actual Rubin PSFs, actual Rubin artifacts. A Rubin Alerts & AI Hackathon ran earlier this year, focused on improving classifier performance on the real stream.

A different kind of telescope

Rubin gets described by its hardware: the 8.4-metre primary mirror, the 3.2-gigapixel camera, the 9.6-square-degree field of view. But the telescope that matters runs after the shutter closes — the broker network that converts raw photon counts into “probable Type Ia supernova at redshift 0.04, three nights before peak.” The mirror collects light. The ML makes it legible.

For anyone who’s watched transient astronomy from the outside, this is the real shift. Discovery used to mean a human noticed something on a photographic plate. Now it means a random forest scored a light curve above threshold, a CNN confirmed the stamp wasn’t a cosmic ray, and a Slack bot pinged a graduate student in Santiago who pointed a spectrograph at it 40 minutes later. The human is still in the loop, but the loop got faster by orders of magnitude.