Using Machine Learning and GPT Models To Enhance Electrochemical Pretreatment of Anaerobic Cofermentation: Prediction, Early Warning, and Biomarker Identification

Electrochemical enhancing anaerobic cofermentation of waste activated sludge and food waste to produce volatile fatty acids (VFAs) represents an innovative and promising approach. Despite its potential, optimizing system performance, providing early warnings, and identifying biomarkers remain challenging tasks due to the intricate interplay of numerous environmental variables and unclear dynamics of microbial interactions. This study first employed machine learning (ML) models including XGBoost, random forest (RF), support vector regression (SVR), and CatBoost to forecast VFA production by integrating initial feedstock properties, electrochemical pretreatment conditions, and fermentation parameters. CatBoost demonstrated the highest R² of 0.977 and the lowest root-mean-square error (RMSE) at 95.69 mg COD/L. Key environmental factors, including fermentation days (VFA production reaching 90% by day 5), salinity (0.5-1.0 g/L), and the carbon-to-nitrogen (C/N) ratio (16.53-22), were identified as optimal for VFA production. To enhance long-term monitoring and facilitate early warning systems, process indicators (pH, ORP, PNs, SCOD, and PSs) from the last day were used to predict VFA production on the following day by fine-tuning the generative pretrain transformer (GPT), with the gpt-3.5-turbo-0125 model exhibiting the highest R² of 0.837 ± 0.004 and lowest RMSE of 296.98 ± 3.65 mg COD/L. Local sensitivity analysis revealed that SCOD was the most important process factor affecting VFA production. Moreover, this study employed ML models to uncover microbial biomarkers at the genus levels, including Prevotella_7, Veillonella, Megasphaera, and Lactobacillus, thereby elucidating the nexus among environmental factors, microbial communities, and VFA production. This study offered a novel modeling workflow for anaerobic cofermentation, enabling process optimization and mechanism exploration with the assistance of ML and large language models.