Recurrent neural networks in astronomy and astrophysics

How it is used

Much of astronomy arrives as a sequence. A telescope measures the brightness of a distant object again and again, a detector records a passing wave of distorted space, a radio dish logs a burst of energy sweeping across frequencies. A recurrent neural network is a model built for data of this shape: it takes one measurement at a time and keeps an internal memory, a running summary that it updates at each step, so that what it saw earlier can shape what it makes of what comes next. In the examples here, such networks read light curves from quasars, waveforms from collapsing stars and simulated radio bursts.

The networks are rarely used alone. One example pairs a recurrent encoder, which compresses a six-band light curve into a compact description, with a decoder built from a differential equation, so that the model both summarises the observations and proposes a physical process behind them, then estimates black hole properties from it. Elsewhere the sequence is reshaped before any network sees it: supernova light curves are smoothed with a Gaussian process and sampled onto a grid of brightness and uncertainty, and radio bursts are presented as frequency against time. Training data, across all these cases, is simulated rather than observed.

Why researchers use it

Astronomical time series are awkward. Observations come at uneven intervals set by weather, scheduling and the turning of the Earth, bands are measured at different moments, gaps are long, and noise varies from night to night. Classical fitting requires choosing a model of the variation in advance. A recurrent network instead learns the shape of the variation from examples, and variants of it can take the gaps and measurement errors as part of the input rather than as something to patch over first.

There is also a shortage of labelled real data. The physical parameters a researcher wants, a black hole's properties, a star's equation of state, the amount of material a radio burst crossed, are not written on the sky. So the networks are trained on simulations, where the answer is known because it was put in: waveforms from hydrodynamic codes, light curves passed through survey cadences and noise models, bursts given dispersion, scattering and scintillation by hand. The network learns the inverse of a simulation that can be run forwards cheaply but not backwards.

In the record

astronomyNeural network reads simulated quasar light curves to estimate black hole propertiesA recurrent encoder with a neural differential equation decoder, trained on simulated six-band quasar light curves, modelled the variability and estimated black hole parameters.astronomyEight machine learning methods sorted supernova gravitational wave signals by nuclear physics modelModels were compared on classifying gravitational wave signals from simulated stellar collapse, using waveforms for several equations of state and many rotation rates.astronomyNeural networks read a fast radio burst's dispersion straight from its spectrumTrained on synthetic frequency against time waterfalls with added dispersion, drift, scattering, scintillation and noise, to estimate dispersion measures for fast radio bursts.astronomyNeural networks sort supernovae and estimate their distances from brightness aloneSimulated supernova light curves were smoothed with a Gaussian process and sampled onto grids of brightness and uncertainty for type classification.

Limits

The chief risk is that these networks learn the simulator rather than the sky. Every example here trains on synthetic data, so any omission in the simulated cadence, noise or physics becomes a blind spot in the trained model, and the errors it reports can be confidently wrong on real observations. Rare or unexpected objects, the ones astronomers most want, are by definition thin in the training set. The running memory is also hard to inspect: the network can give a parameter without any account of which part of the signal persuaded it.

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