AI-powered TERRAHydro could help hydrologists better understand the water cycle

 Ama Dablam in the Himalayas. TERRAHydro will help researchers model complex hydrological systems, such as the amount of freshwater stored in Himalayan icesheets. (Image Credit: Kerensa Pickett)

Ama Dablam in the Himalayas. TERRAHydro will help researchers model complex hydrological systems, such as the amount of freshwater stored in Himalayan icesheets. (Image Credit: Kerensa Pickett)

8/23/23 – A new NASA software could become one of the first AI-based tools for creating composite hydrological models, making it easier for scientists to adapt AI tools for forecasting floods, freshwater availability, and droughts.

The “Terrestrial Environmental Rapid-Replication and Assimilation Hydrometeorological” (TERRAHydro) software is an open-source information system that leverages tensor-based modeling to generate hydrological models using artificial intelligence. With this tool, researchers will be better able to incorporate existing data sets and AI-based modeling components into new, original models describing dynamic components of the water cycle.

TERRAHydro will also be an ideal software foundation for future Earth System Digital Twins, which bring together disparate data sets gathered by airborne, spaceborne, and in-situ sensors to update models and simulations in real time, make predictions and conduct ‘what-if’ investigations.

Craig Pelissier, Principal Investigator for TERRAHydro and leader of the Advanced Software Technology Group at NASA’s Goddard Space Flight Center, said that while hydrologists already have many AI-based hydrological models and model components at their disposal, they don’t have an easy way to integrate them.

Different institutions, he said, may use different modeling techniques and data formats, which makes combining all of those elements into comprehensive modeling systems immensely challenging.

TERRAHydro would provide researchers with a software framework for bringing those different models and data sets together more efficiently in an AI context.
“The research part of that is really in how to couple them together, not on developing the individual components,” said Pelissier.

Grey Nearing, the Co-Principal Investigator for TERRAHydro, agrees. Nearing is a Senior Research Scientist at Google, where he helps develop hydrology models for applications like flood prediction.

“There’s a lot of AI-based hydrology work and land surface work going on, but what we’re doing is trying to bring that work together and design a platform that allows that work to be integrated into a single cohesive model,” said Nearing.

Integrating models is a necessary step for developing large scale hydrological models, but this is especially challenging when integrating AI model components developed by different teams, because integrating AI models requires coupling tensor networks, which is not supported by existing Earth systems modeling tools.

Tensor networks, a powerful mathematical construct for expressing AI-based models, provides the backbone of most existing AI software. They allow Pelissier and his team to produce an AI-powered hydrology modeling platform capable of training, validating, and coupling different models that can leverage existing AI software as the platform’s backend.

The TERRAHydro team is working to develop a general tensor coupling framework designed specifically for Earth systems modeling called the “Coupled Reusable Earth System Tensor” (CREST).

CREST is a graph modeling framework that Pelissier and his team are using to construct TERRAHydro. It serves a similar role to existing Earth systems modeling frameworks, but has the ability to work with AI-based modeling systems.

“The aim of CREST is to provide a framework that allows you to build large open-science coupled Earth systems models along with the required operational infrastructure,” said Pelissier.

Pelissier explained that he and his team also want CREST and TERRAHydro to be intuitive and user-friendly. They plan on developing a graphic user interface to simplify model construction, training, and operations..

“We’re really trying to set the stage for people to use these AI technologies in fully comprehensive ways for weather and climate modeling, specifically the land component of weather and climate modeling,” said Pelissier.

Pelissier and his team have completed the core components of TERRAHydro. Now, they’re moving towards a complete prototype of TERRAHydro’s digital twin capabilities, which they will use to perform ‘what-if’ analyses and case studies.

Specifically, they plan to use historic data to test TERRAHydro’s ability to both forecast the 2006-2010 Syrian drought and calculate the amount of freshwater stored in the Himalayan mountains.

NASA’s Advanced Information Systems Technology (AIST) program, a part of NASA’s Earth Science Technology Office (ESTO), funds this research.

Gage TaylorNASA Earth Science Technology Office