When viewed from Earth, objects in space are seen at a specific brightness, called apparent magnitude. Over time, ground-based telescopes can track a specific object’s change in brightness. This time-dependent magnitude variation is known as an object’s light curve, and can allow astronomers to infer the object’s size, shape, material, location, and more. Monitoring the light curve of satellites or debris orbiting the earth can help identify changes or anomalies in these bodies.

The problem, however, is that light curves are missing a lot of data points. The weather, the season, dust accumulation, time of day, eclipses—these all affect not only the quality of the data, but whether or not it can be taken at all.

A black background with vertical stripes that change color from blue at the top and bottom, to purple, to red, to yellow at the center. From left to right, the vertical lines follow a wavy pattern, showing the seasonality of the data.
Figure: The light curve of a single object over the course of four years, where the color represents the brightness. The black stripes and hole shapes are regions of missing data due to weather and the eclipse, respectively. Data courtesy of Dave Monet and the public catalog www.space-track.org.

Two important hurdles must be overcome in order to effectively monitor satellites’ health: Filling in this data gap—known as the light curve completion problem—and determining an efficient way to find and predict patterns within the data—known as the forecasting problem.

As part of a Laboratory Directed Research and Development program project, a group of researchers spanning LLNL’s Computing, Engineering, and Physical and Life Sciences directorates utilizes a Livermore-developed machine learning (ML) process for light curve modeling and prediction. Called MuyGPs, the process drastically reduces the size of a conventional Gaussian process problem—a type of statistical process that often does not scale well—by limiting the correlation of predictions to their nearest neighboring data points, reducing a large linear algebra problem to many smaller, parallelizable problems. This type of ML enables training on more sensitive parameters, optimizing the efficient prediction of the missing data. Furthermore, while most similar scaling methods for Gaussian processes are local, MuyGPs creates a global model using the time of day and day of the year to predict the magnitude of a given object in space.

“We want to be able to establish patterns of light so that we can in near-real time project where we expect things to be and how they should look, and detect when they deviate from that,” says Min Priest, one of the computer scientists on the project. MuyGPs are better than other ML models at determining their own uncertainty, making it clear when data falls far beyond its expected range. They continue, “So if actual observations are outside of that interval for more than a few observations in a row, something is happening, and it’s something a person should look at.”

Priest and their team tested the predictive performance of MuyGPs using a publicly available catalog of 43 satellites, with data ranging in time periods from hours to weeks. The team measured how accurately their model predicted data either within the set or at its end, representing the light curve completion problem and the forecasting problem, respectively.

For small gaps in observations, MuyGPs behaves comparably to other ML techniques at predicting the missing data. For larger gaps in observations, however, MuyGPs does a much better job. Not only does it achieve better accuracy, it does so faster and with less usage of memory and computing resources. Furthermore, because MuyGPs provides error bars on its predictions, it allows operators to determine whether a future measurement falls out of the likely range and qualifies as unexpected.

Building a catalog of different kinds of objects and their paths is like taking fingerprints of satellites. With a complete set of observations on hand, scientists who see something they don’t immediately recognize can use the patterns of light to figure out what it may be.

“If we have this historical catalog of light curves that is labeled with those features—shape, size, and so on—then we can compare this new observation to the old ones and perform a classification task to make a guess about those sorts of features,” Priest explains.

The group is continuing to make their models more sophisticated to capture more complex dynamics, such as the relative sensitivity of changes along different features of the objects tracked.

MuyGPs is a general-use ML technique, not limited to applications in astronomy. A video explaining how MuyGPs works is available here. An open-source Python implementation of MuyGPs is publicly available here.

The top image is courtesy of RubinObs/NSF/AURA/H. Stockebrand.

—Anashe Bandari