4 minute read
Goddard Scientists Teach Algorithms to Detect Patterns in Data for Future Ocean World Studies
Credit: NASA/ Rebecca Roth
Probes sent to investigate ocean worlds like Saturn’s moon Enceladus and Jupiter’s Europa will collect vast amounts of data, and may need to make autonomous decisions in real time about what warrants a closer look.
This is where machine learning becomes useful. Theiling and MacKinnon are currently teaching machine learning algorithms how to decipher
Probes sent to investigate ocean worlds like Saturn’s moon Enceladus and Jupiter’s Europa will collect vast amounts of data, and may need to make autonomous decisions in real time about what warrants a closer look.
Dr. Bethany Theiling, an Early Career geoscientist at Goddard, hopes to solve the problem of big data as well as communications constraints that come with the territory of investigating planets millions of miles away. These probes will not be able to depend too much on the decision-making or data analysis of human controllers here on Earth.
To start, she simulates oceans of data in the lab. “When I go in the lab, I actually make other worlds there,” Theiling said. “I make oceans that we’ve never even been to. Then I try to figure out if we could determine what those oceans are actually made of and how hard that might be.”
These oceans that Theiling and her team build in their lab are vital to the construction of algorithms that can detect, analyze, and define data. The real worlds they hope to visit are much farther away than a Goddard laboratory, which creates the problem of communicating to the spacecraft with limited bandwidth.
“The problem is the spacecraft are so far away that the actual data rates and latency are a big limiting factor in what the spacecraft can do.” said James MacKinnon, Goddard artificial intelligence and machine learning researcher.
An AI-powered probe, however, could conduct preliminary onboard analysis in order to prioritize transmission of the most important data back to Earth first. The spacecraft would decipher the collected data in real time, determining what specific data points should be sent back for further research, what aspects of that data can remain cached on the spacecraft, and where the spacecraft should turn its focus to learn more.
This is where machine learning becomes useful. Theiling and MacKinnon are currently teaching machine learning algorithms how to decipher and categorize traits using data from the worlds constructed in their lab, along with data Theiling brought with her from her time at the University of Tulsa and other large Earth science datasets. Once these data are prepared, the machine learning algorithm forms a neural network, making connections between the data and prioritizing certain traits, effectively training the machine.
“We almost immediately got really interesting results,” MacKinnon said. “We saw these distinct clusters of points, and usually that’s a good indicator that there are patterns in that data.”
These clusters, found on a visual map of possible data after processing, clearly visualize the unknowns as well. If the algorithm designates that data to an already defined cluster, the composition of the ocean world being studied is assumed to be similar to that of known oceans. However, if the algorithm places data in a spot that isn’t defined, it indicates those characteristics have never been discovered, an equally exciting possibility.
Theiling and MacKinnon aim to use their algorithm for ocean worlds like Europa and Enceladus, which have liquid oceans below their surface ice layers. The tidal forces of their host planets Jupiter and Saturn effect both of these ice moons, keeping their cores from freezing and causing liquid to vent above the surface from below. A spacecraft flying through the plume could collect samples and run them through a mass spectrometer for the algorithm to study. The machine learning process could decipher this data, informing about the ocean below it, ideally characterizing the possibility for life in that landscape.
Closer to Home
This process of machine learning also applies to problems closer to home.
“Climate change is incredibly important to me. And so, I’ve been hoping for a way that I can contribute to that science with the skill set that I have,” Theiling said.
Theiling and MacKinnon ran preliminary tests for machine learning related to Earth studies using a database from the National Ecological Observatory Network, held by the National Science Foundation. This data represents levels of carbon dioxide from varying heights away from Earth’s surface, and from ecosystems all around the United States. “Understanding how things like carbon emissions can change is an important part of knowing the problem exists, and then knowing how to fix it,” MacKinnon said.
The research team aims to identify seasonal patterns, locational patterns, yearly patterns, and even the effects of feedback loops with the machine learning process. Streamlined processing of this data allows it to be used in a multifaceted manner, and it also gives insight to the adaptability of the algorithms.
“How transferable is what we learned from the labcreated ‘worlds’ to a very complicated system like Earth that has plants, water, a hydrological system, and a biological system that’s changing everything?” Theiling asks.
Her current Internal Research and Development (IRAD) grant enabled the purchase of the mass spectrometer and peripheral suites. With this equipment, and an additional grant through Fundamental Laboratory Research (FLaRe), they can continue their work in perfecting the algorithm to analyze and make decisions regarding data. (See profile: CuttingEdge, Summer 2021, Page 6)
CONTACT: Bethany.P.Theiling@nasa.gov or 301-614-6909