Benchmarking Machine Learning Algorithms for Microbial Electromethanogenesis: A Comprehensive Assessment with SHapley...

Area of Science:

• Bioelectrochemistry and Microbial Physiology
• Sustainable Energy and Bioprocessing
• Machine Learning in Environmental Science

Background:

• Microbial electromethanogenesis (EM) is a promising technology for sustainable biogas upgrading.
• Predicting EM performance is difficult due to complex, nonlinear process dynamics.
• Machine learning (ML) offers potential for improving predictive accuracy and process understanding.

Purpose of the Study:

• To systematically compare seven supervised ML algorithms for predicting biomethane production in EM.
• To identify the most influential operational parameters governing EM performance using feature importance analysis.
• To gain mechanistic insights into bioelectrochemical methanogenesis through ML-driven analysis.

Main Methods:

• Experimental data from EM bioelectrochemical systems (EM-BESs) were used, including optical density (OD600), pH, electrical conductivity (EC), average applied current, and CO2 availability.
• Seven ML algorithms were evaluated: 1D-CNN, MLP, GBR, AdaBoost, stacking regressors, and kNN.
• Hyperparameter optimization and SHapley Additive exPlanations (SHAP) were employed for model tuning and feature importance analysis.

Main Results:

• The 1D-CNN model achieved superior predictive performance with an R² of 0.934, outperforming traditional ML methods.
• SHAP analysis identified average current, OD600, and pH as the most critical factors influencing biomethane production.
• Complex, nonmonotonic effects of other variables were revealed, enhancing process understanding.

Conclusions:

• Deep learning architectures, like 1D-CNN, show significant potential for optimizing EM processes.
• ML models can uncover mechanistic insights into bioelectrochemical methanogenesis, grounded in fundamental principles.
• Findings are applicable to other bioelectrochemical systems (BESs), such as microbial electrosynthesis and microbial electrolysis cells, for data-driven operational control.