ChatGPT’s cousin was just hired by NASA. On February 1, NASA and IBM announced a new partnership between the two major organizations, aimed at applying artificial intelligence (AI) tools to climate science, scanning research literature for quick answers and identifying features in Earth science data.
This is far from NASA’s first foray into artificial intelligence, or even the agency’s first collaboration with IBM. In 2014, NASA collaborated with the tech giant to infer measurements of the sun’s extreme radiation when a sensor failed on the Solar Dynamics Observatory. A year later, NASA started a summer bootcamp to bring scientists together with Silicon Valley engineers, known as the Frontier Development Lab.
Plus, since the dawn of machine-learning techniques, scientists across NASA’s domains have been using these tools in their own projects, from looking at the sun to designing autonomous data-gathering robots. As AI has grown in power and complexity, though, it has become harder for individual researchers to harness the full potential of these tools. Each time they start a new project, many NASA engineers and scientists build a bespoke model for each dataset. To solve that problem, in 2020, NASA hosted a workshop on AI. It sought answers to large-scale, extra-challenging problems, dreaming bigger than one-off models for each problem—and IBM’s tech seemed like a perfect match for their needs.
“We have all heard and seen the magic” of widely-applicable machine learning models, especially language models like ChatGPT, said IBM lead developer Priya Nagpurkar in a press conference. “We are at this unique point where it’s time to take those advances and apply them to different domains…as well as advancing science.”
This collaboration is the first time a particular kind of AI—a flexible, broadly-applicable technique known as a foundation model, which IBM is at the forefront of developing—has been applied to Earth sciences. “While NASA and IBM have discussed using AI to solve various problems for the past few years, IBM’s foundation model research was the catalyst for the current collaboration,” says IBM representative Danielle Cerasani.
As described in a recent press release, the collaboration plans to tackle two main projects: answering questions based on scientific literature, and analyzing large datasets of Earth to identify patterns and trends. NASA is providing access to its vast collection of Earth-observing data and its scientists, while IBM is adding AI development expertise and their existing research into this tech.
The literature search is based on technology similar to ChatGPT, and NASA hopes it will serve as a sort of ultra-advanced search engine for scientists.One of its key selling points is that its answers will come with citations—direct links to the research papers it’s pulling information from—unlike other tools that act more like a mysterious black box. Rahul Ramachandran, senior research scientist at NASA’s Marshall Space Flight Center, said in a press conference this technology could be ready as early as mid-2023.
Still, some scientists are skeptical. “The ability of the model to summarize information and answer questions, which is the most innovative aspect especially for the broader community, is also at higher risk of bias,” says Viviana Acquaviva, physicist and AI specialist at the City University of New York. “We have seen how state-of-the-art models like ChatGPT can easily produce biased or incorrect answers, while sounding plausible and confident.” In an advertisement for Google’s new Bard chatbot, for instance, the AI incorrectly stated that the James Webb Space Telescope imaged the first exoplanet, when the European Southern Observatory’s Very Large Telescope had done so years prior.
Meanwhile, applying AI to Earth observations is the more scientifically interesting half of the collaboration, at least to Acquaviva. NASA hosts the world’s largest archives of data on our planet—enough to fill around a million typical iPhones—and they hope to sort it more effectively with IBM’s models.
“Our archive is currently at 70 petabytes and it’s projected to grow within a few years to 250 petabytes…We support 7 billion users worldwide who access our data for research and applications,” Ramachandran told reporters. “Clearly, given the scale of the data that we have, we have a big data problem.”
With the new AI tech, they hope to easily track weather and natural disasters across the planet—as diverse as tornado tracks to dust clouds. Ramachandran imagined a scenario where a disaster response team could quickly analyze the extent of flooding after a hurricane, enabling faster and more effective emergency aid. The team plans to first analyze a data set known as Harmonized Landsat Sentinel-2, a combination of observations from two powerful NASA satellites. This work has just started, however, with Ramachandran describing it as an “open area” of research.
The collaboration also intends to publicly release the code and other tools they develop through these projects, making them available to anyone interested in their use. “It is exciting to witness progress toward the creation of an inclusive and interdisciplinary community,” Acquaviva says, “that can make climate data and AI tools more accessible to scientists and the public.”