Accelerating integrated prediction, analysis and targeted optimization for anaerobic digestion of biomass after hydrothermal pretreatment using automated machine learning

Exploring the complex mechanism of anaerobic digestion with hydrothermal pretreatment (HTAD) for biomass efficiently and optimising the reaction conditions are critical to improving the performance of methane production. This study used H2O automated machine learning (AutoML) for comprehensive prediction, analysis, and targeted optimization of the HTAD system. An IterativeImputer system for data filling was constructed. The comparison of three basic regressors showed that random forest performed optimally for filling (R2 > 0.95). The gradient boosting machine (GBM) model was searched by H2O AutoML to show optimal performance in prediction (R2 > 0.96). The software was developed based on the GBM model, and two prediction schemes were devised. The generalization error of the software was less than 10%. The Shapley Additive exPlanations value showed that solid to liquid ratio, hydrothermal pretreatment (HT) temperature, and particle size have greater potential for improving cumulative methane production (CMP). A Bayesian-HTAD optimization strategy was devised, using the Bayesian optimization to directionally optimize the reaction conditions, and performing experiments to validate the results. The experimental results showed that the CMP was significantly improved by 51.63%. Compared to the response surface methodology, the Bayesian optimization relatively achieved a 2.21–2.50 times greater effect. Mechanism analyses targeting the experiments showed that HT was conducive to improving the relative abundance of Sphaerochaeta, Methanosaeta, and Methanosarcina. This research achieved accurate prediction and targeted optimization for the HTAD system and proposed multiple filling, prediction, and optimization strategies, which are expected to provide an AutoML optimization paradigm for anaerobic digestion in the future. [Display omitted] •Built a machine learning framework for data filling and automated search prediction.•Developed software based on gradient boosting machine with prediction errors of 6.33%.•The solid to liquid ratio influences cumulative methane production most critically.•Bayesian optimization enhanced methane production by 51.63% over the control group.•The Methanosarcina and high energy production pathways increased after optimization.