Future AIs from NASA and IBM want to take advantage of the petabytes of data the US Space Agency produces each year. A real treasure, which could make it possible to prevent certain disasters linked to climate change, and also to create a sort of ChatGPT for researchers in climatology.
While AI models like ChatGPT and Dall-E are tipped to “change the world,” other binary intelligences have more serious missions than generating images of alien cats. Like those jointly developed by NASA and IBM which today formalize a partnership around a huge challenge: to help us study climate change more effectively. And try to find suitable answers. At the heart of the process, two treasures: data and a model. The data is that of the American space agency. According to Rahul Ramachandran, researcher at NASA’s Marshall Space Flight Center (Huntsville, Alabama): NASA currently manages a volume of 70 petabytes and our predictions expect about 250 PB to be managed around 2025 “. A huge volume of data, potentially rich in information… if only we can take advantage of it!
Read also : Space nuclear propulsion to the rescue of trips to Mars… and beyond! (January 2023)
This is where an AI model that you know (perhaps without knowing) comes into its own: the so-called “foundation” model. If you played with ChatGPT, you used a specific AI that derives from GPT-3. It is this same kind of high-level model that IBM wants to develop. Because they are the only ones able to encompass such a gigantic volume of data. And above all, it is these models which, as their name suggests, can serve as a “foundation” for more specialized AIs. In the same way that ChatGPT generates text and Dall-E images, IBM’s model, co-developed with NASA data, will serve as the basis for scientific AIs. In particular, they will be able to help researchers, activists and other emergency services to fight against climate change. And their consequences.
A foundation, uses
While AI training models are abstract, AI applications are much easier to understand. When asked ” how to explain to someone who has zero idea of what an AI is what this model can achieve? “, Rahul Ramachandran answers straight away with a concrete example. “ Imagine rescue teams analyzing the image of a country hit by major floods. With a perfectly trained AI, the AI will be able to determine very quickly which areas are the most affected thanks to the corpus of past and current satellite images. And help teams deploy their efforts smarter and more efficiently “. All this, even if communications on the ground are cut off, the analysis of the images sufficient to make an inventory.
Read also : Why NASA and ESA are betting on RISC V for their future space chips (September 2022)
The mention of satellite images was no coincidence. Initially, it is on the data and images of Harmonized Landsat Sentinel-2 (HLS) that the IBM model will work. Taking advantage of the enormous strength of this kind of model – the fact of not needing pre-labeled data – IBM’s intelligence will learn to know our planet. For (according to the press release) “ identify changes in the geographic footprint of phenomena such as natural disasters, cyclical agricultural yields, and wildlife habitats “. Basically, to become a kind of big brain that will have seen and scanned the entire surface of the Earth, and will be able to support researchers in the interpretation of certain evolutions.
A ChatGPT for earth science researchers
The other intelligence that IBM and NASA will develop is an AI trained on ” 300,000 scientific articles specializing in Earth sciences “. An AI that will work a bit like ChatGPT: become an easily searchable body of knowledge. Partly based on IBM PrimeQA’s multilingual question-and-answer system, this module is intended to be hard-wired into NASA’s data systems. And to be able, at the exit, to be questioned by the researchers.
Enough to develop tools where researchers can query the AI, easily find information and its sources. To move forward on their files, more quickly than with simple classic search engines based on keywords. And above all, with already linking elements of knowledge that only an AI can achieve. The question is obviously whether the relevance of the results will meet the expectations of a specialized audience. In any case, the abuse of language and other elements of disinformation have at least the merit of being absent. The working materials being limited to scientific articles.
Read also : “There is nothing revolutionary about it”: French AI pioneer Yann LeCun is not impressed with ChatGPT (January 2023)
If these projects bear fruit, other initiatives are already being considered, which “include building a foundation model for weather and climate prediction using MERRA2, an atmospheric observation dataset. » It is logical that this project comes after the previous ones, because in addition to a large volume of data, the complexity of climate models requires additional computing power of a completely different magnitude.
Read also : IBM took 20 years to protect us from a quantum threat that does not yet exist (July 2022)
Last point for fans of public funds, unlike the models of OpenAI and others, where many key technologies are closed, here the whole chain of knowledge is open. From NASA data (images, articles, etc.) which are in the public domain, to IBM models which will be licensed open-source, the entire project is under the umbrella of NASA’s Open Source Science initiative. The question being whether IBM and NASA plan to design a “Climate ChatGPT” accessible to everyone!
