Random forests in astronomy and astrophysics

How it is used

Astronomy works almost entirely by collecting light. Telescopes record how bright an object is in different colour bands, how that brightness changes over time, and how its light splits into a spectrum. Astrophysics then asks what kind of object or explosion produced those measurements. A random forest takes a fixed list of measured numbers for one object and returns a label or a probability. In the examples it decides whether a new alert is a type Ia supernova, the exploding remains of a white dwarf star; whether a faint object is a cool T or Y dwarf; what type a white dwarf is; and whether a simulated stellar core explodes at all.

The numbers fed in are usually prepared by hand first. One example fits a smooth rising curve to the changing brightness and passes the fit's parameters, its goodness of fit, the average signal strength and the count of observations. Another turns three infrared brightnesses into colour differences between the bands. Another feeds the coefficients that describe a spacecraft's spectrum directly, without converting them to physical flux. Simulation-based work supplies labels: suites of collapse and explosion models, or synthetic light curves and gravitational waveforms, give known answers to train on.

Why researchers use it

Surveys now produce far more candidate objects each night than anyone can inspect or follow up with a spectrograph, the instrument that spreads light into its component wavelengths and gives the firmest identification. A random forest is cheap to train, runs quickly on each new alert and returns a probability, so it can rank candidates for scarce telescope time. The examples use it this way inside an alert stream, and for picking rare cool dwarfs out of a wide survey.

It also tolerates the conditions astronomical data usually arrive in: a modest number of hand-built features, small or awkward training sets, and labels borrowed from simulations or catalogues. One example starts from a few dozen labelled light curves. Because each tree splits on one measurement at a time, the method handles features on different scales without much preparation, and it stands in for a human expert reading a light curve or a colour diagram.

In the record

astronomyTelescope targets chosen nightly by a classifier learning from its own spectraA thousand-tree forest judged each alert as early type Ia supernova or not, from six features per band, and returned a probability for follow-up ranking.astronomyRandom Forest sorts 9,446 white dwarfs by type from Gaia spectraClassified white dwarf spectral types directly from the Hermite coefficients of Gaia spectra, trained on catalogue labels for nearby white dwarfs.astronomyMachine learning sifts three-colour infrared survey data for cold brown dwarfsDecided from three near-infrared colour indices whether an object is a T or Y dwarf, trained on simulated spectra and catalogue sources.astronomyMachine learning predicts whether a dying massive star explodes, from its density profilePredicted whether a stellar core explodes, from normalised density profiles of progenitor models labelled by two-dimensional simulation outcomes.

Limits

A random forest only sees the features it is given, so the hand-built summaries carry most of the astrophysics, and anything they omit is invisible to it. Labels often come from simulations or existing catalogues, which inherit their assumptions and their gaps: rare or unexpected objects may be absent from training and so get sorted into a familiar class with a confident probability. Early alerts carry few observations, making features noisy. The output is a label or probability, not an explanation, and it cannot replace a spectrum for identification.

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