Validating the predicted uncertainties of the specific dust mass on 18″ scales. The ordinate shows the difference between the predicted log(sM_d) and the ground truth, also known as the prediction error. The abscissa shows the uncertainty predicted by the xgboost model. The colour is not a KDE estimate as in the previous figures, but instead shows the percentile, with the brightest values being used for the median. The bright green line shows a binned RMS of the prediction error. Both the percentile and RMS are binned in the abscissa, i.e. always considering a constant predicted uncertainty. The contours show a KDE estimate and these contours contain 68.3%, 95.5% and 99.7% of the pixels. As done before, all pixels are weighted inversely proportional to the number of pixels in that galaxy.