Predicting current and hydrogen productions from microbial electrolysis cells using random forest model
Jinyoung Yoon, Dae-Yeol Cheong, Gahyun Baek
Abstract
The current- and H 2 -producing performances of microbial electrolysis cells (MECs) were predicted by constructing machine learning models based on the previous 76 MEC datasets, making it the largest dataset to date. All models showed high correlation efficiency (R 2 > 0.92) in predicting MEC performances. When the models were constructed separately based on the organic substrate type used in the anode of MECs, the models based solely on acetate-fed MEC data exhibited higher prediction accuracies compared to those on all kinds of substrate or complex substrate-based data. As a results of the feature importance analysis, the applied voltage and cathode surface area were identified as the two most critical factors in the acetate-fed MEC data models. Still low prediction accuracies in the models here seem to be due to several important features which could not be numerically presented and thus not be considered as input variables such as electrode material types.