Assessing the Importance of Factors Affecting Steam Gasification Efficiency from Eucalyptus Wood: Uncertainty-Aware Machine Learning and Sobol Indices
This work presented an experimentally validated machine learning framework to predict and interpret the properties of producer gas obtained through steam-aided Eucalyptus wood gasification in a downdraft fixed-bed gasifier. The systematic experimental data were collected by varying the equivalence ratio (ER) and biomass consumption rate (BCR). Under optimum conditions, the process recorded a cold gas efficiency (CGE) of up to 82%, a lower heating value (LHV) of 6.8 MJ/kg, and a peak hydrogen concentration of 22.29%, proving that steam integration was effective in improving the quality of the producer gas. Random forest (RF)- and extreme gradient boosting (XGBoost)-based machine learning models were developed as surrogate predictors across the three targets (CGE, LHV, and hydrogen yield). As a result, RF attained training R2 values of 0.990–0.994, while XGBoost achieved ranges of 0.953–0.973. In addition, SHAP-based explainable artificial intelligence was employed to provide interpretable insights into feature significance, indicating that ER is a strong determinant of hydrogen production and CGE, whereas BCR is the primary determinant of LHV. More importantly, the combined surrogate-based Sobol sensitivity analysis and Monte Carlo uncertainty quantification demonstrated rapid convergence within 200–400 samples, confirming the predominant dominance of BCR on LHV (Sobol index of 0.95–0.98) and ER on the hydrogen yield (0.98). Overall, the presented framework was more progressive than traditional modeling approaches because it embedded the crucial elements of interpretability and uncertainty quantification, allowing operators to optimize steam-assisted gasification units with high confidence rates.