Unravelling Systematic Surface Solar Irradiance Biases in NWP Models Using Detailed Cloud Observations: an Application to AROME

Evaluations of medium-term radiation forecasts from Numerical Weather Prediction (NWP) models, despite being sparse, have gained increasing attention for model verification (Ahlgrimm & Forbes, 2012 ; Ahlgrimm et al. 2016 ; Tuononel et al. 2018). Indeed, surface radiative fluxes are a direct proxy of cloud representation which remains a large source of uncertainty in NWP forecasts. Recent studies have shown that the French operational model AROME (Seity et al. 2011) exhibits a positive SSI bias in cloudy conditions across all France, suggesting optically too thin clouds and/or underestimated cloud clover on average (Magnaldo et al. 2024). The objective of our study is to better constrain the origin of this positive bias and to attribute it to specific cloud conditions.

To this end, we use one year of observed geometrical and macrophysical properties derived from the L3 products of the network of cloud observations ACTRIS-Cloudnet (Illinworth et al., 2007). These properties are retrieved at Palaiseau (France), Cabauw (Germany), and Julich (Germany) observatories and consist in Total Cloud Cover (TCC), Cloud Base Height (CBH), Cloud Top Height (CTH), Liquid Water Path (LWP), Ice Water Path (IWP), Integrated Cloud Thickness (ICT) and the number of cloudy layers. These observed features are use to train and label each cloud profile using a Kmeans clustering approach. Once the clusters are constructed, each hour is labelled both from the observations and from the model outputs in order to create a contingency table and identify the dominant pair in terms of bias.

Our results show that low level stratiform clouds in the observations, dominate the overall positive bias. While these clouds typically correspond to overcast conditions (TCC > 90%), AROME seems to almost systematically underestimate their TCC. This underestimation is strongly correlated to a lack of total condensed mass (underestimated LWP), itself correlated to an insufficient vertical extension of the cloud. While the mechanisms at play are numerous, the presentation will investigate some deficiencies using different diagnostic variables. The proposed methodology could be used to systematically evaluate NWP models and directly point to the dominant situations in terms of errors. It could also be extended spatially thanks to the EarthCARE satellite mission to increase its robustness and impact.