Abstract
<title>Abstract</title> <p>Plant microbial fuel cells (PMFCs) operating in unsaturated soil are affected by dynamic interactions among water status, environmental conditions and biochar amendment, yet the relative predictive contributions of these factors remain unclear. This study analysed a time-resolved dataset generated from a controlled PMFC experiment comprising unamended control soil and soils amended with apple wood, corn straw or reed straw biochar. Relative humidity, temperature, volumetric water content, log-transformed matric suction and treatment indicators were used as predictors, while biochar properties and 16S rRNA sequencing data were retained for treatment-level interpretation. Seven regression approaches, comprising a mean baseline and six machine-learning models, were evaluated using within-treatment blocked five-fold cross-validation to limit temporal leakage, with all preprocessing and model selection conducted within the training folds. Among four electrical outputs, electrical potential and power density were the most predictable. Random forest achieved R² values of 0.605 and 0.577 for these targets, respectively. Feature-set ablation showed that hydraulic–environmental variables and treatment indicators provided complementary predictive information. Permutation importance identified the reed straw biochar treatment and matric suction as the strongest treatment-level and continuous predictors, respectively. Microbial community patterns provided independent biological context for treatment-specific responses but were not used as row-level predictors. These findings show that treatment-aware, leakage-controlled machine learning can clarify biochar-specific hydraulic predictors and support material selection and operational assessment within comparable unsaturated-soil PMFC systems.</p>