By Katie Elyce Jones, PillarQ
NASA is known for out-of-this-world images, rocket launches, and microgravity living, but the agency also plays a major role in observing our home planet, particularly its complex climate. Increasingly, NASA is advancing this mission using edge computing devices and artificial intelligence (AI) to bridge data gaps and speed solutions.
“Like many organizations, our use of AI and machine learning has really accelerated in the last 3 to 5 years,” said Daniel Duffy, chief of the Computational and Information Sciences and Technology Office and the NASA Center for Climate Simulation (NCCS) at the NASA Goddard Space Flight Center. The NCCS is a high-performance computing (HPC) center for NASA-sponsored scientific research with a focus on Earth science.
Duffy moderated the panel “Computing at the Edge: A Discussion about Supporting Recent US Space Missions” at the SC22 supercomputing conference in November. The session brought together experts from NASA, Axiom Space, Microsoft, and HPE to discuss computing at the edge for space- and terrestrial-based missions.
From its orbit around the Sun, the James Webb Space Telescope’s onboard Command and Data Handling System controls the telescope’s instruments and communications. Meanwhile, the International Space Station is home to HPE’s Spaceborne Computer-2, an edge computing system that is reducing data processing times for space missions over 250 miles from Earth.
However, for the NCCS, edge devices and AI are solving problems closer to home. “One of NASA’s primary jobs is to observe the Earth,” Duffy said.
“There’s a huge amount of machine learning applications for climate and emergencies such as fires, landslides, flooding, heat waves, and cold spells that can improve our ability to predict events in time frames by which we can help people.”
Predicting climate and extreme weather starts with complex Earth system models that integrate detailed models of the physical and chemical changes in the atmosphere, land, ocean, ice, and other systems.
The Earth itself is notoriously complex—with its 4.5–billion-year history, 320 billion cubic miles of ocean, 200 million square miles of land, millions of plant and animal species, and hundreds of ecoregions. It’s no surprise that Earth system studies necessitate collecting large amounts of data over many places and across many time scales.
Inevitably, there are gaps. That’s where edge devices and AI can help.
Terrestrial data from the edge
NASA collects observational data from a host of edge devices—such as uncrewed vehicles or drones from the atmosphere to underwater, weather balloons, ocean buoys, and space-based satellites including smaller CubeSats.
Often, edge devices are gathering data where people or larger instrument installations cannot—whether due to environmental conditions, cost, time, or practicality. In this way, they provide data that Earth system models otherwise wouldn’t have.
“Where are the gaps in our understanding of the Earth model? [Answering this question], we can develop new edge devices to fill that gap,” Duffy said.
However, edge devices face challenges when it comes to processing data. For example, they can be limited by power consumption and the environment. To reach remote places or to enable many devices to provide coverage over vast spaces, edge devices typically need long-lasting, low-power sources. Devices in extreme environments, like orbiting satellites, may also need to be hardened against radiation, temperature, or other factors.
Although processing power on edge devices is increasing, many of these devices simply collect and transmit data back to the cloud and data centers for analysis. This presents a central, big data challenge because not all data may be useful, and scientists may be wasting time and HPC power on compute-intensive data analysis. In the case of severe weather emergencies, such as wildfires or storms, time is precious.
What if edge devices were equipped to prioritize data on the spot?
“We would like to get to the point where we’re putting inference at the edge,” Duffy said. “We could train models at the HPC level where all the data is, and once you’ve trained models, push the inference to the edge. Let’s make sure we’re making the discoveries, not missing anything critical.”
AI insights for climate and planetary science
In addition to training models for edge devices, AI can expand insights scientists are gathering from existing data—such as extending the time scales for climate or weather predictions. “We can do physics-based weather modeling fairly accurately to 7 to 9 days. We’re really trying to make a push for 1 to 3 months, or seasonal predictions,” Duffy said.
One NASA project is using machine learning to extrapolate information about biomass from satellite maps and estimate carbon sequestration in trees over millions of square miles.
AI will also be key to creating digital twin models. Duffy described current Earth system models as the “engines” to drive digital twins that could examine what-if scenarios and answer questions for different users. For example, a stakeholder may be interested in the best place to build a wind farm or want to understand the risk of flying an aircraft through a storm.
“Not all researchers or stakeholders can run an Earth system model. We can build AI and machine learning tools into the model to help those users downscale, to help them with decision-making tools,” Duffy said.
Researchers are already using models derived from Earth system models to project the climate of other planets, and AI may enable future predictions of the weather and climate on destinations like Mars and the moon.
Duffy said while the applications of AI are being realized, NASA is also focusing on the Year of Open Science in 2023 with its Transform to Open Science initiative. “Trustworthy AI is extremely important for research to be done ethically,” he said. Researchers need to understand where data collection and model training may be bias. For example, urban areas tend to be heavily covered with edge devices while rural areas are not.
“AI can be used to help span that gap, but we’ve got to be careful how we train models. As an HPC center, we’re asking ‘What can we do to support trustworthy AI research?’” he said.
You can read about other NASA computing initiatives at their virtual NASA@SC22 booth.