Diagnosing Forecast Error Propagation and Large-Scale Dynamics of Weather Extremes With an AI Weather Model

September 17, 2026

A paper titled “Diagnosing Forecast Error Propagation and Large-Scale Dynamics of Weather Extremes With an AI Weather Model” was recently published in Geophysical Research Letters. This study was led by CW3E’s Yanbo Nie and co-authored by Agniv Sengupta, Jorge Baño-Medina [Instituto de Fisica de Cantabria (IFCA), CSIC-Universidad de Cantabria], Benjamin Moore, and Luca Delle Monache. The study introduces fast, interpretable methods for investigating why artificial intelligence (AI) weather forecasts miss specific extreme events and how remote atmospheric processes shape them. This research was funded by the Office of Naval Research and the California Department of Water Resources Atmospheric River Program Phase V.

The work supports the “Advanced Precipitation and Streamflow Prediction” priority in CW3E’s 2025–2029 Strategic Plan, particularly its objective to advance machine learning–based prediction of extreme weather. Using GraphCast, the researchers developed a true-state constraint (TSC) method that repeatedly replaces predicted atmospheric states in a selected region with reanalysis data—an observation-informed reconstruction of the atmosphere—while allowing forecasts elsewhere to evolve freely (Figure 1). This reveals how regional errors affect distant forecasts. A complementary climatology constraint method replaces regional conditions with their seasonally varying average to isolate the influence of anomalous atmospheric conditions.

Applied to the June 2021 western North American heatwave and a March 2016 Scandinavian blocking event, the methods identified remote processes consistent with earlier nudging experiments using physics-based models. For the heatwave, correcting conditions over East Asia and the northwestern Pacific improved the representation of an atmospheric river and associated heating over the North Pacific, strengthening the predicted heatwave. TSC-based global sensitivity scans can produce maps of candidate error sources within minutes (Figure 2), highlighting additional Pacific error sources for the heatwave and confirming the Gulf Stream region for Scandinavian blocking. These retrospective experiments demonstrate how AI models can efficiently diagnose forecast failures and investigate extreme-weather drivers, providing a foundation for future forecast improvements.

Figure 1. Schematic diagram of the true-state constraint (TSC) method. After each forecast step, predicted conditions inside the solid box are replaced with concurrent ERA5 reanalysis data before the next step. The band between the solid and dashed boxes blends the forecast and reanalysis, while the model evolves freely elsewhere. Figure S1 from Nie et al. (2026, Supporting Information).

Figure 2. Global sensitivity maps for (a) the June 2021 western North American heatwave and (b) the March 2016 Scandinavian blocking event. Darker blue indicates greater reduction in downstream forecast error when conditions in that grid box are constrained. Shading shows the fractional error reduction, averaged across four initialization times, for temperature at 850 hPa in (a) and geopotential at 500 hPa in (b). These variables describe lower-atmospheric temperature and mid-atmospheric circulation, respectively. Red boxes mark the regions tested using prior research; pink boxes mark additional candidate regions tested for the heatwave. The panels use different color scales. Figure 2 from Nie et al. (2026).

Citation:

Nie, Y., Sengupta, A., Baño-Medina, J., Moore, B., & Delle Monache, L. (2026). Diagnosing forecast error propagation and large-scale dynamics of weather extremes with an AI weather model. Geophysical Research Letters, 53(18), e2026GL123505. https://doi.org/10.1029/2026GL123505