Machine-learning-aided predicting anaerobic digestion of the aqueous phase by-product from biomass hydrothermal conversion.

A significant amount of aqueous phase (AP) by-products is retained after hydrothermal treatment (HTT) of biomass feedstock, and it can be converted into methane using anaerobic digestion (AD). However, experimentally investigating the effects of biomass characteristics, HTT conditions, AP properties, and AD parameters on AD performance is challenging. Therefore, this study employed a series of machine learning (ML) models (Models 1-4) to explore the effect of these variables on methane yield (MY) and methane production rate (MPR). Model 1, incorporating all the four groups of factors, shows that both Random Forest (RF, test R2 of 0.82) and Gradient Boosting Regression (GBR, test R2 of 0.79) demonstrated strong predictive accuracy and generalization capability. Feature analysis revealed that AD parameters were the most significant factor for MY, followed by the biomass characteristics, HTT conditions, and AP properties. Interestingly, Model 2 maintained robust generalization ability despite excluding AP properties that were commonly considered key feed indicators for AD (indicated by Model 4). This is because AP properties could be properly described by the biomass characteristics and HTT conditions (indicated by Model 3). Finally, Models 1 and 2, after out-of-plate samples validation (R2 of 0.65 and 0.62), were developed into software tools for predicting and optimizing AD performance. Under optimal conditions, different biomass feedstocks exhibited distinct MY potential and biodegradability. These models provided valuable insight for understanding how biomass characteristics and HTT conditions impact the AP properties, which in turn affect AD performance.