Materials Science

Lignin Biorefinery Optimization Through Machine Learning

Authors

Abstract

Lignin is an abundant biomaterial that currently emerges as a low value by-product in the pulp and paper industry but could be repurposed for high-value products as part of the ongoing global transition to a sustainable society. To increase lignins value, rational and efficient approaches to optimizing lignin biorefineries to produce high value bioproducts are required. Here, we report the optimization of the AquaSolv Omni (AqSO) Biorefinery, a newly introduced biorefinery concept based on hydrothermal pretreatment and solvent extraction. We employ a machine-learning framework based on Bayesian optimization, to provide sample-efficient and guided data collection as well as surrogate model building. The surrogate models allow us to map multiple experimental outputs, including the extracted lignin yield and main structural properties obtained by 2D NMR, as functions of the hydrothermal pretreatment reaction severity and temperature. Our results show that with Bayesian optimization, predictive models can be converged with only 21 data points to within a margin of error comparable to the underlying experimental error. By applying a Pareto point analysis, we demonstrate how the predictive models can be used in tandem to identify optimal extraction conditions for concrete applications in lignin valorization.

Content

Thumbnail image of lofgren21.pdf

Supplementary material

Thumbnail image of SI_lofgren21.pdf
Supporting Information: Lignin Biorefinery Optimization Through Machine Learning
The supporting information contains a more technical introduction to BO and the concept of Pareto optimality. A HSQC spectrum for the acetone-extractable lignin is also included, along with complementary results for the acquisition strategy comparisons and the model validation.