Distribution of the prediction errors, r = y−ŷ, of the final models for Fe (upper part) and H (lower part) in the different sets (from top to bottom): training (‘train’), validation (‘val’), test (‘test’), and holdout curves (‘hold’). The standard deviation of the prediction errors (σ), their mean value (r̄ i.e. the mean bias of the predictions), and mean absolute values (ē; i.e. the mean absolute error) are shown in the tails of the violin plots.