From biomass structure to hydrogen-rich syngas: Interpretable machine learning for hydrothermal gasification

Most machine-learning studies of hydrothermal gasification (HTG) rely on small, single-feedstock datasets and represent biomass using elemental analysis alone, which captures bulk stoichiometry but not the macromolecular structure that controls hydrothermal decomposition. In this study, a semi-automated large language model-assisted extraction workflow was used to convert published tables, figures, and text from independent HTG/SCWG studies into structured experimental records. The resulting dataset was used to train target-specific stacked-ensemble models for predicting H 2 , CO 2 , CO, and CH 4 under elemental and structural input configurations. The ensemble combined XGBoost, gradient boosting, and random forest base learners with a ridge-regression meta-learner, together with physically motivated feature engineering and target-specific feature selection. On the full elemental dataset, the stacked ensemble achieved test R 2 values of 0.94–0.96 for H 2 , CO 2 , and CO, with RMSE values of 2.40, 4.60, and 1.76 percentage points, respectively. CH 4 remained more difficult to predict, but the model achieved a test R 2 of approximately 0.84 with an RMSE of 4.45 percentage points. On the matched structural subset, cellulose, hemicellulose, and lignin descriptors achieved R 2 = 0.88 for H 2 and R 2 values of 0.75–0.82 for the remaining gases, outperforming elemental-only models evaluated on the same records. Cross-conformal prediction intervals achieved empirical coverage of 90–100% at the nominal 90% level, providing calibrated uncertainty estimates for model-informed experimental planning. Interpretation results showed that temperature and solids loading dominate elemental models, whereas hemicellulose and lignin become key structural drivers in matched models. These findings indicate that structural biomass descriptors provide a chemically meaningful bridge between feedstock architecture, hydrothermal decomposition pathways, and syngas selectivity, enabling uncertainty-aware support for feedstock screening, operating-window selection, and HTG/SCWG process optimization.