Leveraging machine learning for acid catalyzed steam explosion pretreatment: Towards supporting fermentation by the trade-off between glucose and inhibitors

Abstract

Pretreatment is essential for enhancing the sugar release from lignocellulose. Acid-catalyzed steam explosion (ACSE), a widely-used pretreatment method, still faces challenges, including inhibitors accumulation, which can be overcome by modeling. Here, artificial neural network models were constructed for sulfuric acid-based ACSE to predict sugars and inhibitors, from 12 variables regarding lignocellulose, acid, and steam explosion. Two expanding applications were demonstrated. Firstly, a constraint-based optimization strategy can provide the optimal ACSE condition for fermentation by considering glucose and the synergistic effect of inhibitors on microbial growth simultaneously. Compared to published works, the strategy led to 94% glucose with 22% inhibitors for corn stover, and 100% glucose with 13% inhibitors for wheat straw. Secondly, transfer learning was employed to model phosphoric acid-based ACSE with high accuracy (MSE 0.004) and low data requirement (∼30% of sulfuric acid-based ACSE). The proposed models and applications offer an effective optimization strategy for ACSE and other pretreatment methods for the following fermentation.