Reactor-Aware Machine Learning Coupled with Differential Evolution for Predicting and Optimizing Cumulative Methane Production from Agro-Industrial Waste Co-Digestion.

Anaerobic digestion of agro-industrial residues supports waste valorization and renewable-energy production, but reliable prediction requires validation that accounts for repeated measurements within reactors. This study compared 16 regression models for predicting cumulative methane production from digestion time and banana peel-sugarcane molasses composition using 5007 observations from seven batch reactors. Models were evaluated by leave-one-reactor-out cross-validation (LORO-CV). Radial-basis-function support vector regression (SVR-RBF; C = 10, gamma = "scale", epsilon = 0.1) achieved the lowest pooled RMSE (118.09 NmL CH4), with R2 = 0.9482 and MAE = 75.75 NmL CH4, and was selected as the surrogate model. However, reactor-level Wilcoxon tests with Holm correction showed no significant differences between SVR-RBF and the other algorithms. Held-out-reactor R2 values ranged from -1.366 to 0.928, indicating heterogeneous generalization. Differential Evolution consistently identified approximately 100% banana peel and 0% molasses as the optimal composition. Across 70 runs, the median optimum was 310.10 h and 1433.25 NmL CH4. Bootstrap analysis placed 99% of composition optima at ≥99% banana peel, although uncertainty in optimal time was substantial. Kinetic benchmarking supported the slower, higher-volume methane production observed in complete banana-peel reactors. This boundary solution is therefore a model-supported candidate requiring experimental confirmation, not a universal co-digestion optimum.