Property prediction in astronomy and astrophysics

How it is used

Almost everything in astronomy is known only through light. A telescope records how bright an object is in a few colour bands, or spreads its light into a spectrum, a graph of brightness against wavelength. The properties astronomers actually want, temperature, surface gravity, chemical make-up, mass, distance, are hidden inside those measurements. Property prediction trains a model on examples where both the light and the answer are known, then applies it to objects where only the light is known. In the examples here, models read low-resolution stellar spectra to return temperature, gravity and element abundances; read near-infrared colours to sort cool dwarf stars by type; and read the flickering brightness of quasars over time to estimate black hole properties.

Training labels come from two sources. Sometimes they come from better observations: spectra from a survey that measures a modest amount of light per object are matched to the same stars observed at higher quality elsewhere, and the model learns to reproduce the better answer from the poorer data. More often, and in most of the examples here, the labels come from simulation. Researchers run a physical model forward, atmospheres of cool dwarfs, a thermochemical model of a dusty disc around a young star, ray-traced light echoes around a black hole, then add the instrument's filters, cadence and noise so the simulated data look like real data. The model learns the reverse mapping, from observation back to physical parameter.

Why researchers use it

The honest way to measure a property is to fit a physical model to each object, adjusting parameters until the predicted light matches what was seen. That works, and it is slow. Surveys now deliver millions of spectra and billions of light measurements, and fitting each one separately is not practical. A trained model does the fit once, during training, and afterwards answers in a fraction of the time. In the examples, this is what allows stellar parameters to be produced for a very large sample of giant stars from a single survey release.

The second reason is that some mappings have no tractable formula to invert. A quasar's brightness varies irregularly, and it is sampled whenever the telescope happens to look, leaving gaps. Relating that ragged record to the geometry of the disc and the mass of the black hole is a hard inverse problem. A model trained on simulations learns it by example. The same logic covers reconstructing the expansion history of the universe without committing to a particular functional form, and standing in for a slow piece of analysis software inside a larger simulation chain.

In the record

Limits

A trained model is only as good as its labels. Where those come from simulation, the model inherits every simplification in the physics and every mismatch between simulated and real noise, and it has no way to signal that an object lies outside what it was shown. Where labels come from a better survey, the model carries that survey's own errors and is reliable only over the range of stars that were matched. Models also tend to pull estimates towards the middle of the training distribution, flattening the rare extremes that are often the most interesting objects.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Related

• Property prediction: the idea, across every field
• AI in astronomy and astrophysics

Every article