Multi-Product Modeling of Consolidated Bioprocessing Using a Literature-Derived Dataset: A Multi-Output Learning Framework for Ethanol and Co-Products

Consolidated bioprocessing (CBP) has been widely studied as an integrated route for converting biomass into biofuels and bioproducts, yet most quantitative modeling work has focused on ethanol as a single response. Because CBP systems can generate multiple products and co-products, this study develops a literature-derived benchmark for multi-product CBP modeling using a standardized dataset assembled from published endpoint experiments. Product prediction is formulated as both an observed-only product-wise problem and a joint multi-output problem, allowing direct comparison under study-aware grouped validation. The modeling space integrates biomass composition, pretreatment descriptors, microbial and consortium characteristics, reactor information, operating conditions, and engineered categorical descriptors of feedstock, pretreatment family, and process configuration. Predictive performance was strongly product-dependent and was shaped by target support and missing-label structure. The observed-only product-wise formulation consistently outperformed the joint missing-as-zero multi-output strategy, indicating that naive zero-filling of unreported products is not well suited to sparse literature-derived CBP data. Among the evaluated products, butanol showed the clearest predictive signal, ethanol was only moderately learnable, and the sparsest co-products remained too weakly supported for strong quantitative inference. Overall, this study provides a benchmark for multi-product CBP modeling and clarifies both the potential and the current limitations of literature-derived data for broader data-driven biorefinery analysis.