Interpretable machine learning reveals physicochemical drivers of bio-crude yield in hydrothermal liquefaction of water hyacinth
Abstract
Hydrothermal liquefaction (HTL) provides a promising route for converting wet biomass into energy-rich bio-crude without energy-intensive drying. Water hyacinth is well suited to HTL because of its high moisture content, rapid growth, and widespread occurrence as an invasive aquatic biomass. However, bio-crude formation is governed by coupled process and feedstock variables, making prediction and optimization challenging. This study developed an interpretable machine-learning framework using 88 experimental observations from seven independent publications on water-hyacinth HTL. Reaction temperature, residence time, hyacinth-to-water ratio, atomic H/C ratio, and catalyst presence were used to predict bio-crude yield using five machine-learning models. Model reliability was evaluated using leakage-controlled grouped validation strategies. Gradient boosting (GB) showed the best performance, with a pooled out-of-fold R2 of 0.923, MAE of 3.046 wt%, and RMSE of 4.213 wt%. Its performance remained robust across 100 repeated grouped partitions (R2 = 0.895 ± 0.073) and nested grouped cross-validation (R2 = 0.895 ± 0.085). SHAP analysis identified the hyacinth-to-water ratio and reaction temperature as the most influential predictors, while partial dependence analysis revealed nonlinear responses to the operating variables. O/C sensitivity analysis showed negligible improvement in predictive performance after its inclusion, and its feature importance was not stable under robustness testing. Model-based optimization identified 311 °C and 46 min as a favorable operating condition under the selected reference conditions, corresponding to a predicted bio-crude yield of 56.39 wt%. This condition requires independent experimental validation. The findings provide insight into the process variables associated with bio-crude formation and can guide targeted experiments, selection of favorable operating conditions, and data-driven optimization of water-hyacinth hydrothermal biorefineries.