Forecasting Climate-Compounded Geohazards with AI-accelerated Digital Twins
Please login to view abstract download link
Climate-compounded geohazards arise from tightly coupled, nonlinear interactions among the atmosphere, ocean, hydrosphere, soils, and the solid Earth, yet operational hazard systems typically treat these components in isolation. We present a physics-aware, AI-accelerated digital twin (DT) of the shallow Earth that explicitly represents soil hydromechanical memory—the evolving state of water content, pore pressure, permeability, and strength that governs cascading hazards such as landslides, floods, and earthquake-induced ground failure. Our proposed DT builds on an established framework of a multi-scale data-fusion architecture that integrates satellite remote sensing of surface states and fluxes (e.g., Destination Earth), with dense geophysical sensor networks (seismic, geodetic, hydrological), and model-generated geospatial data within a unified data-assimilation and reduced-order modeling framework. Foundation models for Earth data, specifically OlmoEarth, are leveraged for geospatial representation learning, feature extraction, and cross-modal alignment across heterogeneous observations, enabling efficient learning from sparse, irregular, and partially observed fields. Atmospheric forcing is provided by observations, ERA5, and AI-based weather foundation models, including ACE2, to deliver high-resolution, ensemble forecasts that drive short-term hazard evolution. Designed for near-real-time operation, the DT produces probabilistic nowcasts and forecasts of hazard susceptibility while supporting counterfactual and scenario-based analyses under non-stationary climate conditions. A modular architecture that combines curated FAIR datasets, containerized physics-aware AI components, and transparent evaluation workflows ensures extensibility and transferability. We evaluate and validate the framework using complementary Pacific Northwest use cases, notably the the 2001 Mw 6.8 Nisqually earthquake, assessing the evolving liquefaction and ground-failure susceptibility given present-day hydromechanical and climatic states; and the Pacific Northwest (PNW) atmospheric-river–driven landslides and floods in December 2025, testing the twin’s ability to capture storm sequencing, soil preconditioning, and basin-scale flood response. Together, these cases demonstrate how data fusion and foundation models enable digital twins to bridge process understanding and actionable decision support for climate-resilient infrastructure and hazard mitigation.
