Large Universe Models for astronomy

What it ingests

Survey alert streams, gravitational-wave triggers, neutrino alerts, archival catalogues across multiple wavelengths, prior light curves for every object, and the observing schedules of the follow-up instruments themselves.

Continuous by necessity

A modern survey telescope emits millions of alerts a night. Each is a detection that something in a patch of sky changed brightness or position. Most are artefacts, known variables, or asteroids. A handful are the reason the telescope exists.

There is no version of this problem that batches. A transient that is interesting for six hours cannot wait for a weekly run, and by the time a frozen model has been retrained on the discovery the object has faded.

The field's answer is the alert broker: a system that ingests the full stream, cross-matches each detection against archival catalogues and the object's own history, classifies it, revises the classification as more photometry arrives, and escalates the few worth pointing another instrument at — within minutes.

That is observe, revise, act. It is the Large Universe Model loop, running in production, at scale, for years.

What makes it a Large Universe Model

No cutoff. The broker's picture of the sky is exactly as old as the last alert packet.

Belief, not classification. A transient is not labelled once. It carries a distribution over types that shifts as the light curve develops — supernova rising, then a specific subtype as the decay rate resolves. Early classifications are explicitly provisional.

Revision as the output. The valuable event is often not the classification but the change in it. An object that was confidently one thing and is now confidently another is worth a telescope; an object that has been stable for a week is not.

Provenance throughout. Every classification unwinds to the detections behind it. Astronomy requires this for the same reason regulated industries do: the conclusion is only usable if the evidence is inspectable.

What the field proves about the architecture

The astronomy case is important to the general argument for a reason that has nothing to do with astronomy: it demonstrates that continuous ingestion and belief revision are not speculative. They were built, they work, and they have been operating at extreme scale in a discipline that could not function otherwise.

What astronomy had to do, and what the Large Universe Model removes, is build the instrument by hand. Every broker is bespoke — purpose-built classifiers, hand-specified features, an observation model written by people who understood both the physics and the statistics. The technique was available to everyone. The engineering was affordable only to fields with no alternative.

Where a general Large Universe Model adds something

The literature as a stream. A broker sees photometry. It does not read the paper published this morning describing a new subtype with a distinctive early-time signature. A Large Universe Model that ingests both the alert stream and the literature can revise its classification scheme on evidence, not on the next software release.

Cross-messenger reconciliation. Gravitational-wave, neutrino and optical alerts arrive on separate systems with separate conventions. Holding one belief per astrophysical event, fed by all three, is a reconciliation problem of exactly the kind a general model handles better than three specialised ones.

Follow-up as an action. Deciding which of tonight's candidates deserves scarce telescope time is a decision under uncertainty with a real budget constraint — the natural closing of the loop from belief to act.