Fig. 3: Predictive ML models constructed from the DFT dataset of adsorption energies (EAds) of ground and excited states of ceria-based group 9–11 transition metal (TM) SACs.
From: Data-driven models for ground and excited states for Single Atoms on Ceria
Predictions by the best-performing Elastic Net (EN), Random Forest (RF), and Bayesian Machine Scientist (BMS) models are shown in the panels a, b, and d, with training (testing) data points indicated as circles (crosses). The gray areas mark a region with a deviation of up to 0.2 eV. Panel c shows the courses of RMSE values for the EN and RF models during sequential feature selection (SFS), evaluated via K-fold cross-validation (KFCV; opaque bars) and leave-one-group-out cross-validation (LOGOCV; transparent bars). The use of secondary features for Elastic Net (see main text) requires a minimum of two descriptors in the reduced primary space. The actual error values and selected primary features obtained during SFS, panel c, are listed in Supplementary Table 6.
