Hybrid application of response surface methodology modelling and machine learning exploratory prediction for reducing sugar recovery from potato peel waste for bioethanol production

Global interest in bioethanol production from waste lignocellulosic biomasses (LCB) has heighten, however, current efforts are often constrained by lack of predictability of outcome in terms of sugar recovery towards bioethanol produced. Thus, focusing on reducing sugar (RS) recovery, this study models reduced sugar release from nano-based PPW pretreatment. RSM was initially employed to design 15 experimental runs for each nanoparticle treatment and to optimize RS recovery. Subsequently, four ML algorithms: Random Forest Regression (RF), Gradient Boosting Regression (GBR), Support Vector Regression (SVR), and Linear Regression (LR) were trained using the RSM-generated experimental datasets. The RSM-RS recovery models showed high R 2 > 0.91 with Fe 2 O 3, NP regime achieving the highest RS recovery (0.280 g/g) followed by 0.142 g/g (Fe 3 O 4 ) and 0.177 g/g (Fe 3 O 4 /Ag). GBR demonstrated the highest in-sample predictive performance, with R 2 values of 0.959–0.981 and RMSE values < 0.022 g/g across the three nanoparticle-based responses. RF also showed strong performance, achieving R 2 values of 0.926–0.957 with RMSE values < 0.029 g/g. In contrast, LR produced substantially lower R 2 values (0.147–0.242), indicating that the input-output relationship is fundamentally nonlinear. However, cross-validation (5-fold CV R 2 = 0.280–0.747 for Gradient Boosting; 0.280–0.747 for Fe 3 O 4 /Ag RS) confirmed acceptable generalisation for the Fe 3 O 4 and Fe 3 O 4 /Ag responses while revealing overfitting limitations for Fe 2 O 3 RS due to small dataset (n = 15). Heatmap visualization further demonstrated appropriate coverage of the experimental design space and highlighted the importance of reactant concentration in determining RS recovery. Following pretreatment and saccharification, fermentation of the PPW hydrolysates produced maximum bioethanol concentrations of 31.30, 34.20, and 29.70 g/L for Fe 2 O 3 , Fe 3 O 4 , and Fe 3 O 4 /Ag nanoparticle treatments, respectively. The findings therefore provide a basis for integrating RSM and ML to improve the prediction and optimization of RS recovery from starch-based LCBs for bioethanol production.