Lightning-Fast Convective Outlooks: Predicting Severe Convective Environments with Global AI-based Weather Models

TL;DR

This work addresses the challenge of predicting severe convective environments by evaluating three global AI-based weather models (GraphCast, Pangu-Weather, FourCastNet) for convective parameters up to 10 days ahead against ERA-5 reanalysis and ECMWF IFS. It derives CAPE, DLS, and a combined WMS metric from model outputs and uses RMSE, bias, FSS, and SAL to assess forecast performance, including regional analyses across North America, Europe, Argentina, and Australia. The results show GraphCast and Pangu-Weather delivering CAPE and DLS forecasts with skill comparable to or better than IFS, while FourCastNet generally lags, largely due to moisture predictions and a problematic RH-to-Q conversion. These findings suggest that fast, inexpensive AI-based outlooks can support hazard-driven forecasting, with further gains possible from direct CAPE training and higher vertical resolution to more faithfully represent convective initiation and severity metrics.

Abstract

Severe convective storms are among the most dangerous weather phenomena and accurate forecasts mitigate their impacts. The recently released suite of AI-based weather models produces medium-range forecasts within seconds, with a skill similar to state-of-the-art operational forecasts for variables on single levels. However, predicting severe thunderstorm environments requires accurate combinations of dynamic and thermodynamic variables and the vertical structure of the atmosphere. Advancing the assessment of AI-models towards process-based evaluations lays the foundation for hazard-driven applications. We assess the forecast skill of three top-performing AI-models for convective parameters at lead-times of up to 10 days against reanalysis and ECMWF's operational numerical weather prediction model IFS. In a case study and seasonal analyses, we see the best performance by GraphCast and Pangu-Weather: these models match or even exceed the performance of IFS for instability and shear. This opens opportunities for fast and inexpensive predictions of severe weather environments.

Figures (4)

• Figure 1: Synoptic situation in ERA-5 on April 12 and 13 2020, showing a deep trough moving across the United States, with anomalously warm temperatures and high CAPE values in the warm sector; a) Climatological geopotential height anomaly and surface temperature anomaly, contours indicate positive (dashed) and negative (solid) anomalies of geopotential height in 20 m increments b) pressure-level-derived CAPE and DLS, hatched area indicates where DLS > 20 m s$^{-1}$.
• Figure 2: Forecast comparison of CAPE and DLS at 12h and 6.5 days lead-time; A) ERA-5 data of CAPE and DLS on April 12, 2020 at 12 UTC; a) CAPE and DLS from pressure levels, b) CAPE from model levels, c) density diagram of CAPE from pressure and model levels, d) density diagram of CAPE from pressure and model levels against observed soundings. A) 12h forecast of CAPE and DLS in comparison to ERA5 (e-h) B) 6.5-day day forecast of CAPE and DLS in comparison to ERA5 (i-l); contours indicated positive (gray) and negative (black) areas of $\Delta$ DLS in 5 m s$^{-1}$ increments
• Figure 3: Seasonal evaluation of CAPE for North America at 00 UTC initialization shows Pangu-Weather and GraphCast outperforming IFS, with a moisture bias impacting the performance of FourCastNet. Top row: a) RMSE, b) BIAS, c) FSS$_{300}$; Second row: SAL components for CAPE > 300 J kg$^{-1}$; color legend in panel c); Third row: RMSE of T, Q and RH at 925 hPa (g-i); Fourth row: BIAS (j-l); model legend in panel g). Solid line depicts the median and shading the interquartile range.
• Figure 4: Seasonal evaluation of CAPE in North America (Apr-Aug, a-d), Europe (Apr-Sep, e-h), Argentina (Sep-Feb, i-l) and Australia (Sep-Feb, m-p) shows good performance of GraphCast and Pangu-Weather; solid line depicts median of score, the IQR is shown in shading