Harnessing machine learning to decode and optimize bioelectrochemical systems: Principles, progress and future directions.
Bioelectrochemical systems (BES) represent an interdisciplinary convergence of biology, electrochemistry, materials science, environmental engineering and mechanical engineering, offering transformative potential for renewable energy generation, wastewater treatment and resource valorization. However, the inherent structural intricacy and mechanistic complexity of BES pose significant challenges to system understanding and optimization. With robust capabilities in pattern recognition and nonlinear system modeling, machine learning (ML) appears to be a good approach to decipher the complex mechanisms of BES. A systematic literature review reveals that ML applications in BES date back to 2006, with a marked surge around 2021, reflecting the growing research interest in this interdisciplinary field. The application domains primarily fall into four categories: (1) analysis and prediction of microbial communities, (2) intelligent design of system components, (3) performance prediction and system optimization, and (4) real-time monitoring and assisted intelligent control. Among these, performance prediction and system optimization constitute the dominant application area, and model interpretability and generalization are cross-cutting requirements for reliable and transferable ML deployment in BES. Among the BES subtypes that have employed ML, microbial fuel cells (MFC) account for the largest share (42.4%), followed by electrochemical biosensor (EB, 37.6%) and microbial electrolysis cells (MEC, 10.9%), with other types occupying smaller proportions. Regarding algorithmic choices, artificial neural networks are the most frequently used method (30.9%), followed by support vector machines or support vector regression (17.3%), principal component analysis (14.5%), and regression trees (13.1%). ML has exhibited significant potential in elevating BES design, manufacture, operation and application. However, constrained by data scarcity and heterogeneity, present models are with limited transferability across scales. Further efforts are warranted to promote the application of ML in BES by expanding data accumulation, diversifying datasets, and developing targeted models. This will ultimately enable a deeper understanding, enhanced optimization and broader deployment of BES.