Augmented machine learning with limited data for hydrogen yield prediction in wastewater dark fermentation

Abstract Current machine learning (ML) efforts for predicting hydrogen yield in dark fermentation are constrained by limited sample sizes and distributional skewness, yielding unstable models. These data characteristics fundamentally restrict generalization and hinder the optimization of process conditions. In this study, a generative adversarial network (GAN)-inspired strategy was developed to augment an initial dataset of 210 dark fermentation samples to 1050 synthetic instances, significantly enhancing data distribution normality and coverage. Across nine ML algorithms, the Histogram-based Gradient Boosting (HGB) model performed best on the test dataset ( R 2 ≈ 0.95; RMSE < 0.06; MAE < 0.05). SHAP and accumulated local effects (ALE) analyses indicated that butyrate, biomass, and Ni positively influenced hydrogen yield, whereas elevated COD, ethanol, and longer hydraulic retention time (HRT) reduced it. Two-dimensional ALE plots further identified the optimal operating conditions for dark fermentation (Fe/Ni ratio ≈ 1:3; HRT of 4–5 h; pH ≈ 4.9; and COD < 25 g L − 1 ). A Python-based graphical user interface (GUI) integrating the HGB model was developed for practical hydrogen yield prediction and process diagnostics. This study demonstrates that combining GAN-inspired data with gradient boosting models can enhance both prediction accuracy and process control in biohydrogen production from wastewater.