Figure 9.

SHAP/Shapley value statistics per input feature for cloud cover predictions on vertical layer 32 (at ≈ 7 km) of the column‐based models with a focus on q _v and q _i in (b–e). Input features the models have not in common are neglected. As in Figure 8, the Shapley values for both models are computed on the same sets of 10⁴ random NARVAL R2B5 samples (using 10 different seeds). (a): The sum of average SHAP values over all vertical layers. The black lines show the range of values (min/max). The absolute Quasi‐Biennial Oscillation in a Changing Climate R2B5 model bias (of 0.95%) on layer 32 (cf. Figure 6a) can approximately be recovered by summing over all orange values (which yields 0.81%). (b, c): The vertical profiles of SHAP values for q _v and q _i for all 10 seeds. In the SHAP dependence plots (d, e) we zoom in on the features with the largest SHAP values (q _i and q _v of layer 32). (d, e): Each dot corresponds to one NARVAL R2B5 sample. The lines show smoothed conditional expectations computed over all seeds. The dashed lines show the average SHAP value of the input features q _v and q _i on layer 32 whose values can also be found in (b and c).