Understanding and optimizing xylose conversion to furfural for biomass valorization using experimental-machine learning methods

Furfural is an important biobased platform chemical produced via the acid-catalyzed dehydration of xylose; however, its industrial selectivity remains constrained by competing degradation pathways. This study evaluates the catalytic performance of ZSM-5 zeolites, SO3H-functionalized biochar, and sulfonic acid-functionalized Material Institute Lavoisier (MIL)-101 (Chromium) in various solvent environments (water, toluene, and gamma-valerolactone (GVL)) to identify conditions that maximize xylose conversion and furfural selectivity. Pearson correlation coefficients were calculated to assess how the input variables (catalysts and solvents) correlated with the target variables (xylose conversion, furfural selectivity, and furfural yield). Further trained and tested tree-based machine learning (ML) models, Random Forest (RF) and eXtreme Gradient Boosting (XGBoost), to assess their predictive performance. Five-fold cross-validation was performed, using 80% of the data for training and 20% for validation in each fold. Finally, to interpret the ML models, the SHAP values were computed, which quantify each input variable’s contribution to the prediction. Among the tested systems, ZSM-5(80) in a 1:9 water-toluene biphasic medium at 160 °C and 3 h was identified as the optimum condition for pure xylose conversion, achieving 96.9% xylose conversion, 73.2% furfural selectivity, and a maximum furfural yield of 71.1%. The results revealed that solvent composition predominantly influences conversion, while zeolite acidity and pore architecture govern selectivity. Overall, the XGBoost model predicted the target variable more accurately than RF, and catalyst type and reaction medium influenced it. Application of the experimentally optimized conditions to lignocellulosic hydrolysate further confirmed their robustness under realistic, impurity-containing feed compositions.