Fig. 1. Schematic of the overall three-stage workflow: training a generalized ML model and interpretation (1); discovery of the type and optimal amount of mixed oxide catalysts using ML-BO (2); and screening for OXZEO based on the outputs of the first and second stages (3).

Fig. 2. Specifications of data collection.

Fig. 3. Alluvial diagram of the STO data set, showing metal types (columns 1-3), synthesis method (column 4), and different zeolites (column 5). Each node shows the frequency. Composite catalyst combinations are shown from left to right.

Fig. 4. Box normal plots representing the mean and quartile values, along with the data distribution of the reaction conditions: WHSV (a), reaction temperature (b), H2/CO ratio in the syngas (c), and reaction pressure (d). The x-axis represents reaction parameters, and the y-axis represents the frequency of data distribution.

Fig. 5. Pearson correlation of features represented as a heatmap. The color of the boxes represents the significance of feature correlations, with ρ-values displayed in each box. Larger absolute ρ-values indicate stronger correlations between feature pairs, while smaller absolute values correspond to weaker correlations.

Fig. 6. Regression plots of the space-time yields (STYs) of light olefins predicted by extreme gradient boosting (XGB) (a), random forest (RF) (b), and gradient boosting decision tree (GBDT) (c), showing R2, root-mean-square error (RMSE), and mean absolute error (MAE) for the train and test data sets and the distribution on gridlines. (d) Feature importance of the best model (GBDT) represented based on normalized Shapley additive explanation (SHAP) values.

Fig. 7. Two-way partial dependence plots (PDP) of the reaction parameters on the light olefin STY: Reaction temperature and Ox/Zeo ratio (a), H2/CO ratio and Ox/Zeo ratio (b), reaction temperature and reaction pressure (c), and reaction temperature and WHSV (d). The surfaces are colored according to their relative light olefin STY values (dark green: low; dark red: high).

Fig. 8. STOx data set analysis: (a) Dependence of CO conversion on CO2 selectivity and oxygenate selectivity. Filled stars, triangle, and circles represent Zn-Al oxide, Cu-Zn-Al oxide, Zn-Zr oxide, respectively. (b) Predictions (blue hollow stars), and experimental data (black hollow circles) are represented as functions of CO conversion vs. light olefin selectivity. The color map indicates light olefin yields. Oxygenates include methanol, ketene, and DME.

Fig. 9. The STY values for catalysts derived from experimental data are compared with those of novel catalysts predicted within a defined range of reaction conditions. The experimental data are depicted as black hollow circles, while the prediction data are illustrated as blue hollow stars. The prediction data are further specified by purple hollow stars, which denote catalysts sampled during the ML-BO process.

Fig. 10. (a) XRD patterns of fresh and spent Cu-Mg and Zn-Zr mixed oxides. CO2-TPD profiles (b) and BET isotherm results (c) for both Cu-Mg and Zn-Zr mixed oxides.

Fig. 11. The first CO hydrogenation step on ZnZrO(101) and CuMgO(100) surface models.