• Biomass pretreatment experiments and model development for process design purposes. • Hydrothermal pretreatment of wheat straw not competitive. • Dilute acid pretreatment of wheat straw has monomer yields of. Y = 98 % w / w . • Machine learning and mechanistic models have good validation metrics. • The latter performs best in optimization of operation conditions under uncertainty. The underlying study presents models for the optimization of operational conditions of lignocellulosic biomass pretreatment to facilitate the conceptual process design of biorefineries. Experiments for hydrothermal and dilute acid pretreatment are performed and analyzed. The highest xylose monomer yield obtained for dilute acid pretreatment is Y Xyl = 98 % at a temperature of T = 195 °C , a reaction time of t = 18 m i n and a dilute acid concentration of C acid = 1.25 w t % . The data is used to fit a response surface model (RSM), a Gaussian process regression model (GPR), and a mechanistic model based on thermodynamic principles and first-order rate equations. All models are validated respectively with a coefficient of determination of R RSM 2 = 0.914 , R GPR 2 = 0.999 and R mech 2 = 0.988 . Each model is used in an optimization problem to predict the optimal operational conditions that maximize the xylose yield. The conditions found by the mechanistic model T = 191 . 6 ° C , t = 18 m i n , C ac = 1.13 w t % with C Xyl , m e c h = 23.47 w t % and the GPR T = 195 ° C , t = 18 m i n , C ac = 1.25 w t % with C Xyl , G P R = 23.23 w t % are in agreement and stand out compared to the RSM metamodeling approach T = 182 . 4 ° C , t = 26.2 m i n , C ac = 1.25 w t % , which yields C Xyl , R S M = 25.72 w t % . Considering the scenario of uncertainty in the feedstock composition, the optimization under this uncertainty with the mechanistic model yields slightly different conditions T = 182 . 6 ° C , t = 18 m i n , C ac = 0.84 w t % and C Xyl , m e c h , u c = 20.88 w t % . Given the underlying phenomena in the biomass pretreatment, all models have shortcomings; however, the mechanistic model is validated best overall and is thus recommended for further engineering purposes as, e.g., the conceptual process design of biorefineries.
Read full abstract