2026/07/15 by Mark Korang Yeboah, Ahmad Addo, Nana Yaw Asiedu
Computer Science · Decision Sciences · Engineering · #Advanced Multi-Objective Optimization Algorithms #Bioprocess #Control theory (sociology) #Optimal Experimental Design Methods #Process Optimization and Integration #Robustness (evolution) #Work (physics)
paper · pdf · doi:10.1016/j.ces.2026.124663
published in Chemical Engineering Science 338, 124663 (Elsevier BV)
openalex publication_date 2026/07/15 · openalex created_date 2026/07/16 · openalex updated_date 2026/07/25
Consolidated bioprocessing (CBP) is highly sensitive to operating conditions, yet systematic optimization remains difficult because batch experiments are lengthy, costly, and affected by process uncertainty. This study proposes a surrogate-assisted batch-to-batch optimization framework for CBP under measurement noise and plant–model mismatch. The approach combines a bootstrapped Extra Trees endpoint surrogate with acquisition-based sequential decision-making and a physics-informed hybrid CBP model for in silico campaign evaluation. Temperature, pH, substrate loading, and pretreatment are optimized sequentially from batch-end ethanol measurements, while predictive uncertainty is estimated empirically from the dispersion of bootstrap-ensemble predictions. The framework is assessed in paired 30-run Monte Carlo campaigns against random search, balanced design-of-experiments (DoE), and Gaussian-process Bayesian optimization (GP-BO) policies under measurement noise and kinetic perturbations representing plant–model mismatch. Under ideal calibration, the proposed ET-BO strategy achieved a mean final best scaled ethanol endpoint of 0.1029, compared with 0.0981 for GP-BO, 0.0844 for random search, and 0.0842 for balanced DoE. These values correspond to improvements of approximately 22.0 % and 22.3 % over random search and balanced DoE, respectively. The proposed ET-BO strategy significantly outperformed both non-adaptive baselines across all tested mismatch levels while remaining statistically comparable to GP-BO. Outcome variability increased under stronger mismatch, and uncertainty diagnostics showed that raw bootstrap dispersion was useful for acquisition ranking but was not fully calibrated. Mechanistic trajectory analysis showed that improved operating policies were associated with stronger biomass and enzyme accumulation, greater substrate depletion, a larger transient sugar pool, and enhanced late-stage ethanol formation. Overall, these results support bootstrapped tree-ensemble acquisition-based optimization as a data-efficient strategy for endpoint-only CBP operating-condition search and provide a foundation for future constrained, multi-objective, and experimental implementations.