Improved estimation of dark fermentation biohydrogen production utilizing a robust categorical boosting machine-learning algorithm

Abstract

Dark fermentation is attracting great attention due to some capabilities such as wastewater treatment and biofuel production. In this study, the performance of different machine learning algorithms, including Bagging-Extreme Gradient Boosting (BagXGBoost), Categorical Boosting (CatBoost), K Nearest Neighborhood (KNN), Light Gradient Boosting Machine (LightGBM), and Decision Trees (DT) in the prediction of biohydrogen production from wastewater during dark fermentation is investigated. Effect of parameters Ni, Fe, biomass, pH, chemical oxygen demand (COD), HRT, acetate, ethanol, butyrate, and acetate/butyrate ratio as inputs were considered for the model development. A databank with 208 data points was culled from the literature. For the validation of the proposed models, various statistical criteria were determined. CatBoost and BagXGBoost exhibited superior performance in predicting fermentative hydrogen (H2) production. Shapley additive explanations were used for relative analysis, revealing the highest absolute SHAP values of 1.1, 1, and 0.07 for COD, nickel, and butyrate, respectively, highlighting their significance in H2 production. The insights gained from this study can contribute to further optimization of the dark biohydrogen fermentation process.