Prediction of hydrogen production via supercritical water gasification of biomass using interpretable machine learning models: Considering the element compositions to enhance model generalization and robustness

Supercritical water gasification (SCWG) is a highly promising thermochemical route for converting wet biomass into hydrogen-rich synthesis gas, but existing models are biomass-specific and lack interpretability. This paper collects 150 sets of experimental data and constructs five models: Extreme Gradient Boosting (XGB), Random Forest (RF), Gradient Boosted Decision Trees (GBDT), Support Vector Regression (SVR) and Lightweight Gradient Boosting (LGB), with XGB and GBDT showing top accuracy and LGB performing well. SHAP analysis identified temperature and feedstock concentration as key factors, with average contributions of 2.9611 and 3.2772, respectively. LGB demonstrated strong interpretability and pattern-capturing ability. Given LGB overall average metrics (MSE = 0.6379, R² = 0.9206, MAE = 0.4879) and its excellent interpretability, LGB was selected for multi-biomass analysis. Incorporating biomass elemental composition raised R² for four gases from 0.4476 to 0.7064 to 0.8832-0.9099, confirming that biomass diversity greatly improves prediction. This study provides a rapid and easily interpretable machine learning tool for optimizing SCWG process parameters and predicting syngas composition.