Opening Black-Box Machine Learning Models for Greener Solvent Pretreatment of Lignocellulosic Biomass

The fractionation of lignocellulosic biomass and the utilization of its components remain a bottleneck in unlocking the potential of this source for energy and chemicals. Explainable Artificial Intelligence (XAI) approach has emerged as a way to increase the reliability of models that are considered black boxes. This review brings together the use of post hoc interpretability techniques applied to black-box models in lignocellulosic biomass pretreatment with green solvents. Machine learning (ML) models are able to learn relationships between delignification/cellulose enrichment and a set of features. In post hoc analyses, variables related to biomass composition, solvent descriptors, and operating conditions repeatedly emerged as influential predictors. Three aspects are highlighted: (1) combining interpretable models by design with more powerful models explained by post hoc techniques; (2) applying best practices in ML workflows; and (3) investing in standardized and open databases. It can be concluded that XAI combined with ML can accelerate the design of solvents and processes optimization, increase the reliability of results/predictions, and strengthen the competitiveness of biorefineries toward sustainable chemistry and a circular bioeconomy.