Artificial intelligence helps to verify processes in NWP and climate models

Physical processes on the weather time-scale importantly modulate the large-scale midlatitude circulation. In particular, the most rapidly ascending air streams in extratropical cyclones, so-called warm conveyor belts (WCBs), have a major impact on the dynamics and are sources and magnifiers of forecast uncertainty. Thus, an adequate representation of WCBs is desirable in numerical weather prediction and climate models. Most often, WCBs are defined on the base of Lagrangian trajectories which are computationally expensive to calculate and require data at high spatio-temporal resolution. Thus, systematic evaluations of the representation of WCBs in large numerical weather prediction (NWP) and climate model data sets are missing.

A major goal of the Helmholtz Young Investigator Group “Sub-seasonal Atmospheric Predictability: Understanding the Role of Diabatic Outflow” (SPREADOUT) at IMK-TRO is to develop statistical techniques that allow the identification of WCBs without performing expensive trajectory computations [1]. Most recently, convolutional neural network (CNN) models were trained to predict the occurrence of WCBs using a combination of meteorological predictors which are physically meaningful to describe the WCB and which are routinely available in NWP forecast and climate model projections. The validation of the CNN models against a trajectory-based data set confirms that the deep learning based approach reliably replicates the climatological frequency of WCBs as well as their footprints at instantaneous time steps [2].

With its comparably low computational costs the new deep learning based diagnostic is ideally suited to systematically verify WCBs in large datasets such as ensemble reforecast or climate model projections. Further, the diagnostic can be easily applied to different modeling systems such as the ICON NWP model (Fig. 1). Accordingly, model intercomparisons concerning the representation of WCBs will be conducted in future studies. A first systematic verification of WCBs in ECMWF‘s subseasonal to seasonal reforecasts highlights that significant systematic biases in the occurrence frequency of WCBs exist (Fig. 2) and that reliable predictions of WCBs are not possible beyond 10 days forecast lead time [3].

 

Fig. 1: Exemplary application of CNN-based WCB diagnostic to ICON forecast. Colored dots indicate ascending air parcels identified with the trajectory-based approach. Green contour denotes ascent region identified with the CNN-based WCB diagnostic. Further are shown mean sea level pressure (gray contours), WCB inflow region (red contour) and WCB outflow region (blue contour) as identified with the CNN-based WCB diagnostic. Figure provided by Annika Oertel.

 

Fig. 2: Frequency bias of WCB ascent for DJF 1997-2017 (shading) at 7 days forecast lead time. Robustness is indicated by the point hatching (first, weaker level) and the line hatching (second, stronger level). The black contours indicate a climatological WCB ascent frequency of 1, 5, 10, 15%. Figure provided by Jan Wandel.

 

Overall, the novel diagnostic demonstrates how deep learning methods may be used to advance our fundamental understanding of processes on the weather time-scale that are involved in forecast uncertainty and systematic biases in NWP and climate models.

References:
[1] Quinting, J. F., and C. M. Grams (2021). Towards a systematic evaluation of warm conveyor belts in numerical weather prediction and climate models. Part I: Predictor selection and logistic regression model. Journal of Atmospheric Sciences, in revision.

[2] Quinting, J. F., and C. M. Grams (2021). Deep Learning for the Verification of Warm Conveyor Belts in NWP and Climate Models. Part I: Model development. In preparation.

[3] Wandel, J., J. F. Quinting, C. M. Grams (2021). Towards a systematic evaluation of warm conveyor belts in numerical weather prediction and climate models. Part II: Verification of operational reforecasts. Journal of Atmospheric Sciences, submitted.

Julian Quinting, Working Group: “Large-scale Dynamics and Predictability” LINK: https://www.imk-tro.kit.edu/english/7425.php