@article{RoueguiBellabairElAslietal.2025, author = {Rouegui, Meryem and Bellabair, Hind and El Asli, Abdelghani and Amar, Amine and Z{\"o}rner, Wilfried and Rachidi, Fouad and Lghoul, Rachid}, title = {Bio-Methanization of Sheep Manure and Beet Waste in the Meknes-F{\`e}s Region, Morocco: Effects of Pretreatment and Machine Learning Applications for Biochemical Methane Potential Prediction}, volume = {10}, pages = {213}, journal = {Recycling}, number = {6}, publisher = {MDPI}, address = {Basel}, issn = {2313-4321}, doi = {https://doi.org/10.3390/recycling10060213}, year = {2025}, abstract = {Sheep manure and beet waste (the uneatable leaf part of the beet) are promising feedstock for biogas production due to their abundance and organic richness. However, their high lignocellulosic content reduces anaerobic digestibility and controls methane yield. This study investigates the effect of various pretreatment strategies, namely physical, thermal, and combined physical-thermal methods, on the Biochemical Methane Potential (BMP) of sheep manure and beet waste. Batch anaerobic digestion experiments were conducted under mesophilic conditions, with BMP values recorded for each treatment. The highest BMP for sheep manure, 125 Nml CH4/g VS, was achieved using combined physical and thermal pretreatment. This approach enhanced methane production by 16\%, 25\%, and 60\% compared to physical pretreatment (PP) alone, thermal pretreatment (TP) alone, and no pretreatment, respectively, while the one BMP for beet waste is 80 Nml CH4/g VS and obtained with thermal pretreatment. To predict BMP outcomes, three machine learning approaches are applied, namely Linear Regression (LM), Random Forest Regression (RFR), and Gradient Boosting Machine (GBM), using digestion time (N days), total solids (Ts), volatile solids (Vs), pretreatment type, and biomass type. The variance analysis confirmed that the interaction between pretreatment and biomass type significantly improved model performance. While diagnostic checks revealed non-linear patterns limiting the linear model, ensemble methods achieved stronger results. The RFR model explained 79.5\% of the variance with a Root Mean Square Error (RMSE) of about 15.7, whereas the GBM model achieved the lowest RMSE of 5.05. GBM captures complex non-linear interactions. In addition, variable importance analyses identified digestion time, solid content, and pretreatment as the most influential factors for methane yield, with the combined chemical and physical pretreatment producing the highest biogas outputs. These findings underscore the potential of advanced machine learning models, particularly GBM (Gradient Boosting Machine), for optimizing anaerobic digestion strategies and maximizing biogas recovery from sheep manure and beet waste.}, language = {en} }