fields/astronomy/model families
Gradient-boosted trees in astronomy and astrophysics
How gradient-boosted trees, a method that builds many small decision rules in sequence, are used in astronomy and astrophysics, from sorting signals to reading physical properties out of observations.
How it is used
Astronomy works almost entirely from light and other signals that arrive from far away. A telescope or detector records a stream of numbers, and the task is to decide what produced them: a planet crossing in front of its star, a star collapsing, a disc of gas and dust around a young star. Gradient-boosted trees are fitted to tables of such numbers. A decision tree asks a series of yes-or-no questions about measured quantities and arrives at an answer; boosting fits one tree after another, each correcting the errors left by those before it, and adds them together.
In the examples the input is always a short list of measured or simulated quantities rather than a raw image. One study turns each moment of a star's brightness record into a handful of summary figures (the brightness itself, the local average, the spread, the skew of nearby readings) and asks the model for the probability that a planet is passing in front. Another feeds in brightnesses at particular wavelengths from a disc of gas and dust and asks for the physical properties, such as mass, that would produce them. A third sorts simulated gravitational wave signals from collapsing stars.
Why researchers use it
Much astronomical data arrives as a modest number of carefully chosen measurements per object: a few line strengths, a continuum brightness, a size. Gradient-boosted trees suit exactly this shape. They need no assumption that the relationship between measurement and answer is smooth or straight, they cope with quantities on wildly different scales, and they train quickly enough that several alternatives can be compared on the same data, which is what the exoplanet study does when it fits six different kinds of model to one feature set.
They also stand in for a slow calculation. Physical models of stars and discs run forwards: assume a mass and a composition, then compute what a telescope would see. Going backwards, from the observation to the physics, usually means running the forward model many times over. Instead, researchers run it on a grid of cases, then fit trees to that grid so the reverse step becomes a lookup. The DiskMINT work does this, using only its own model output as training data.
In the record
Limits
Trees interpolate within what they were shown and little beyond it. A model trained on a grid of simulated discs or waveforms inherits that grid's assumptions, so a real object outside them may be read confidently and wrongly. The summary figures must be chosen by hand, and anything the chosen figures do not capture is invisible to the method. Astronomical samples are also very unbalanced: genuine events are rare among candidates, so a model can score well while still flooding astronomers with false alarms, and a threshold set on one set of stars may not transfer.
Written by AIxSci: the background from general knowledge, the examples from checked records.
Related
Every article
aix-00271
Machine learning predicts galaxy 'relicness' from survey measurements, cutting spectroscopy demand
aix-00238
Classifiers label individual TESS brightness readings as planet transits or not
aix-00090
Machine learning reads planet-forming disk masses from archival ALMA observations
aix-00049