Fig. 4: Uncertainty estimation and its reflection to surface Pourbaix diagram.
a Machine learning (ML) (BE-CGCNN with Dropout Neural Network in this case) and DFT prediction results of adsorption energy difference. Each point represents average value of 1000 sampled case of predicted adsorption energy difference and the corresponding error bar represents ± one standard deviation value. b Surface Pourbaix diagram of Pt55 (Coh) with uncertainty estimations. The phase boundary line is represented by the dashed line (the average of 1000 sampled cases), and its uncertainty is represented by the gradient at each line. For a clearer understanding of the uncertainty values, the probability density distributions of each phase boundary lines at pH 14 are shown on the right. The white, blue, orange, and red shaded area represent bare Pt NPs, OH-covered, O-covered, and Pt dissolution phases, respectively. As the color became darker, more adsorbates are adsorbed.
