Machine-learning-assisted process interpretation of composting time across composting systems involving multiple...

Composting time across organic waste systems is associated with feedstock composition, operating conditions, and stage-specific temperature evolution. Conventional approaches have difficulty characterizing nonlinear variation among multiple process variables. This study developed a machine learning-assisted framework to identify process features associated with composting time in heterogeneous, multi-source composting systems and to characterize their structural associations. This study compared six machine learning models. XGBoost showed the strongest ability to characterize nonlinear associations between process variables and composting time. This study used repeated GroupShuffleSplit grouped by study source for outer evaluation and fivefold GroupKFold within each training set for hyperparameter optimization. XGBoost achieved a mean test R² of 0.77. The Bootstrap 95% confidence interval for test R² was 0.74-0.80. SHAP identified cooling-maturation-phase cumulative temperature (AT3) as the feature with the greatest contribution to the model output. AT3 remained the most important feature in both turning-only and forced-aeration subsets. Structural equation modeling (SEM) showed the strongest structural association between AT3 and composting time (β = 0.80, p < 0.001), and model fit and the principal paths remained stable in leave-one-study-out analyses. These findings provide data-driven information for process monitoring and subsequent controlled experiments in composting practice.