Public Abstract

DE-SC0026756: GENESIS-AR: GENErative models and New observations for Improved S2S prediction of Atmospheric Rivers

Award Status: Active

Institution: Board of Trustees of the Leland Stanford Junior University, Redwood City, CA
UEI: HJD6G4D6TJY5
PM: Hnilo, Justin

Most Recent Award Date: 08/25/2026
Number of Support Periods: 1
PI: YANG, DA

Current Budget Period: 09/01/2026 - 05/31/2027
Current Project Period: 09/01/2026 - 05/31/2027

Public Abstract

Title: GENESIS-AR: Generative Models and New Observations for Improved Subseasonal-to-Seasonal Prediction of Atmospheric Rivers Stanford University: Da Yang (PI), Greg Beroza, Ching-Yao Lai, Noah Diffenbaugh Pacific Northwest National Laboratory: Ruby Leung (co-PI) National Center for Atmospheric Research: Christine Shields (co-PI) Abstract: Atmospheric rivers (ARs) are long, narrow corridors of concentrated water vapor that produce much of the extreme precipitation and water-supply variability along the U.S. West Coast. They are essential to regional water resources, but the strongest events can cause flooding, damage energy infrastructure, disrupt hydropower operations, and contribute to wind-related power outages. Although current forecast models can often predict ARs several days in advance, reliable forecasts two to six weeks ahead remain difficult. Better probabilistic forecasts at these lead times could help energy and water managers prepare for hydropower inflow variability and flood hazards. GENESIS-AR will develop and evaluate a physics-guided AI framework to improve subseasonal AR prediction. The project combines three complementary advances: generative AI to correct forecast errors and represent uncertainty; an efficient AI weather model informed by the physical energetics of moisture transport; and a feasibility study of seismic noise as an additional observational source. AI is central to the proposed workflow because it can learn complex forecast-error distributions, generate large forecast ensembles at low computational cost, and identify patterns in large atmospheric and seismic datasets that are difficult to extract using conventional methods alone. First, diffusion models will learn the full range of forecast errors, rather than only the average bias, to test whether they improve estimates of the probability of rare and extreme AR events and subseasonal AR forecasts. Second, the project will retrain the ACE2 AI model to incorporate integrated vapor kinetic energy (IVKE), a measure of the strength and energetics of atmospheric water-vapor transport. This will test whether incorporating physical knowledge improves forecast skill, maintains reliable performance over extended forecasts, and clarifies the processes controlling AR predictability. Third, the team will test whether microseisms—weak, continuous ground vibrations generated through interactions among the atmosphere, ocean, and solid Earth—contain detectable signals of ARs. Phase I will use overlapping seismic observations and satellite-era AR records to evaluate this possibility. If the feasibility is established, future work could apply the method to earlier seismic records to provide a longer observational context for rare events and long-term variabilities in Earth system. Phase I will develop and evaluate prototype methods, compare them with established forecasting and observational baselines, and quantify where AI provides improvement. Successful capabilities could support DOE's mission by improving information relevant to hydropower planning and the resilience of energy infrastructure along the U.S. West Coast.

Title: GENESIS-AR: Generative Models and New Observations for Improved Subseasonal-to-Seasonal Prediction of Atmospheric Rivers

Stanford University: Da Yang (PI), Greg Beroza, Ching-Yao Lai, Noah Diffenbaugh

Pacific Northwest National Laboratory: Ruby Leung (co-PI)

National Center for Atmospheric Research: Christine Shields (co-PI)

Abstract:

Atmospheric rivers (ARs) are long, narrow corridors of concentrated water vapor that produce much of the extreme precipitation and water-supply variability along the U.S. West Coast. They are essential to regional water resources, but the strongest events can cause flooding, damage energy infrastructure, disrupt hydropower operations, and contribute to wind-related power outages. Although current forecast models can often predict ARs several days in advance, reliable forecasts two to six weeks ahead remain difficult. Better probabilistic forecasts at these lead times could help energy and water managers prepare for hydropower inflow variability and flood hazards.

GENESIS-AR will develop and evaluate a physics-guided AI framework to improve subseasonal AR prediction. The project combines three complementary advances: generative AI to correct forecast errors and represent uncertainty; an efficient AI weather model informed by the physical energetics of moisture transport; and a feasibility study of seismic noise as an additional observational source. AI is central to the proposed workflow because it can learn complex forecast-error distributions, generate large forecast ensembles at low computational cost, and identify patterns in large atmospheric and seismic datasets that are difficult to extract using conventional methods alone.

First, diffusion models will learn the full range of forecast errors, rather than only the average bias, to test whether they improve estimates of the probability of rare and extreme AR events and subseasonal AR forecasts. Second, the project will retrain the ACE2 AI model to incorporate integrated vapor kinetic energy (IVKE), a measure of the strength and energetics of atmospheric water-vapor transport. This will test whether incorporating physical knowledge improves forecast skill, maintains reliable performance over extended forecasts, and clarifies the processes controlling AR predictability.

Third, the team will test whether microseisms—weak, continuous ground vibrations generated through interactions among the atmosphere, ocean, and solid Earth—contain detectable signals of ARs. Phase I will use overlapping seismic observations and satellite-era AR records to evaluate this possibility. If the feasibility is established, future work could apply the method to earlier seismic records to provide a longer observational context for rare events and long-term variabilities in Earth system.

Phase I will develop and evaluate prototype methods, compare them with established forecasting and observational baselines, and quantify where AI provides improvement. Successful capabilities could support DOE's mission by improving information relevant to hydropower planning and the resilience of energy infrastructure along the U.S. West Coast.