Figure 40.

Summary plot of SHAP feature importance for a GBDT model, trained using pore properties descriptors (POV: pore occupiable volume, D_i: diameter of the largest included sphere, D_f: diameter of the largest free sphere, D_if: diameter of the largest included sphere along the free path) to predict N₂ uptake from the CO₂/N₂ mixture data from Boyd et al.¹³ Note that we chose the N₂ uptake as one expects that the pore geometry is more important than the chemistry, which simplifies the example. The violins in this plot show the distributions of the importance, i.e., the spread of the SHAP values (along the abscissa) and how many samples we have for different SHAP values (the thickness of the violin). The coloring encodes the value of the features, red meaning high feature values whereas blue represents low feature values (e.g., high vs low density). The SHAP value is shown on the abscissa and reflects how a particular feature (one feature per row) with a value represented by the color impacts the prediction. For example, a high density (red color in the second row) leads to lower predictions for N₂ uptake (indicated by negative SHAP values).