Simulation surrogate in astronomy and astrophysics

How it is used

Astronomy and astrophysics study objects that cannot be brought into a laboratory. Instead, researchers write computer programs that follow the physics forward: given a star's mass, or a planet's interior layers, or the chemistry of an atmosphere, the program predicts what a telescope would see. These programs are slow, and fitting one to real data means running it thousands of times with different settings. A simulation surrogate is a stand-in. The slow program is run over a spread of settings to make a library of examples, and a statistical model is trained on that library until it can guess the program's output directly. In the examples here, surrogates stand in for binary star light curves, exoplanet interiors, disc chemistry, black hole images and galaxy cluster maps.

The stand-in can also run the other way. Rather than predicting what the telescope sees, some models learn the inverse mapping, from the observed quantities back to the physical parameters that produced them. One example trains only on that inverse direction, using a grid of disc models post-processed into a short list of observable brightnesses and radii. Another learns both directions at once, parameters to image and image to parameters, from a library of black hole pictures made by conventional ray tracing. A related trick replaces part of the physics rather than the whole program: a search is used to cut a large chemical reaction network down to a smaller one that still behaves like the original.

Why researchers use it

The attraction is arithmetic. Fitting a model to data, especially in the Bayesian way that gives honest error bars on each parameter, requires an enormous number of trial runs. If each trial takes minutes or hours of supercomputer time, the fit is impossible; if a trained stand-in answers in a fraction of a second, it becomes routine. The expensive program still does the work, but only once, in advance, to make the training library. Several examples show the care that goes into building that library: parameters are sampled so as to cover the space evenly, ranges are set from what the observations already suggest, and outputs are rescaled or put on a log scale before training.

Surrogates also let a library be reused. One example builds a large master database once, then narrows it to a particular planet by keeping only the samples close to that planet's measured mass and radius and resampling around the measurement. Another uses stand-ins to bridge between kinds of simulation, learning to produce observable maps of galaxy clusters from cheap simulations that follow only dark matter, so that the expensive runs including gas need not be repeated.

In the record

Limits

A stand-in is only as good as the library it learned from, and only inside it. Asked about a star, planet or cluster unlike anything in the training grid, it will still answer, confidently and possibly wrongly. Grids are coarse, so fine structure can be smoothed away, and any simplification in the original program is inherited silently. Errors in the stand-in are hard to separate from errors in the physics, so a fit can look precise while being wrong. Reduced chemical networks and cheap-simulation shortcuts drop effects deliberately, and what was dropped may matter later.

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

Related

• Simulation surrogate: the idea, across every field
• AI in astronomy and astrophysics

Every article