NSF grant ~$2.21M: GAIA, forecasting earthquakes, landslides, and floods with near-real-time digital twins (University of Washington)

The problem this award attacks is that the observations exist but do not talk to each other. Earthquakes, landslides, and floods are linked: heavy rain saturates a hillslope, an earthquake weakens that same slope, and the slope fails in a later storm.

Yet the measurements that could warn of such cascades — ground vibrations, deformation, satellite imagery, meteorological data, groundwater levels — are gathered by separate systems that rarely interoperate. GAIA answers by fusing these streams with numerical simulations of ground and water dynamics into a continuously updating digital twin of a region, estimating hazards in near real time.

1Surrogate models plus data assimilation

The most interesting technical choice is the pairing of surrogate models with data assimilation. Full physics simulations are precise but computationally expensive — too slow for real-time vigilance. GAIA trains fast AI surrogates that approximate those simulations and couples them to incoming observations through data assimilation, buying speed without abandoning physics.

It reads as a transplant of the model-plus-assimilation methodology that weather forecasting spent decades refining, into solid-earth hazards such as landslides.

2A six-year investment in software infrastructure

Just as notable: this is a six-year investment in software infrastructure, not a single finding.

A data aggregator that returns data from many national archives in consistent cloud-native formats; fusion methods for observations of differing resolution; and prototype research agents that retrieve and reason over literature and code to recommend methods, flag knowledge gaps, and generate reproducible workflows — every line of work aims at tools anyone can use.

The proving ground is landslides in Alaska and the Pacific Northwest, connecting to real disaster resilience, while the record states the tools are designed as a template for any field rich in data and physical models. Open tutorials, workshops, and graduate courses fold in workforce training.