Surrogate Modeling of Methane Yield in Anaerobic Digestion Under Heterogeneous Feedstock Conditions

Predicting methane and biogas yields from heterogeneous feedstocks requires distinguishing performance within an available dataset from transfer to unrepresented substrate labels. This study compares artificial neural network (ANN), Gaussian process regression (GP), and proper orthogonal decomposition combined with radial basis function (POD–RBF) using 83 laboratory observations and five measured descriptors. The two measured fresh-mass yields are the primary endpoints; four DM/ODM-normalized yields are secondary. Shared five-fold cross-validation (CV) and 30-label leave-one-substrate-out (LOSO) validation exclude substrate identity from the predictors and fit preprocessing and POD exclusively on training data. For the primary endpoints, POD–RBF has the lowest mean CV root mean square error (23.84 m3/Mg fresh mass), whereas GP has the lowest mean CV symmetric mean absolute percentage error (sMAPE; 29.71%). Mean group-wise LOSO sMAPE is 46.35%, 46.78%, and 45.09% for ANN, GP, and POD–RBF, respectively. Mean coverage of conditional nominal 95% GP observation intervals decreases from 93.37% in CV to 77.11% in LOSO. Three full-data POD modes retain 99.16% of centered output energy, partly reflecting deterministic relationships among mass-normalized yields. Model rankings depend on the metric and validation setting, and the single CV partition leaves split dependence unquantified. Large errors for some excluded labels and GP under-coverage limit claims of reliable prediction for unseen feedstocks or industrial application.