Lawrence Berkeley National Laboratory

Feedstock blending as a strategy to mitigate risks in the supply of lignocellulosic biomass to commercial scale biorefineries across various geographical areas in the United States. Machine learning predictive models estimate sugar yields from feedstock blends expected to be available in Florida and Kansas. Performance of each model was assessed based on feature selection along with two-stack ensembles applied in both linear and nonlinear algorithms. Linear-weighted ensemble and nonlinear stochastic gradient boosting model ensembled with four base learners exhibited similar predictions as previously developed linear regression models when predicting glucose yields in the higher range. The ensemble models achieved a 10–50 % improvement in the root mean squared error with feature selection compared to models with full features from validation. Machine learning has the potential to predict sugar yields at high confidence for a given feedstock blend ratio and pretreatment conditions.