Predicting xylose yield from prehydrolysis of hardwoods: A machine learning approach

Hemicelluloses are amorphous polymers of sugar molecules that make up a major fraction of lignocellulosic biomasses. They have applications in the bioenergy, textile, mining, cosmetic, and pharmaceutical industries. Industrial use of hemicellulose often requires that the polymer be hydrolyzed into constituent oligomers and monomers. Traditional models of hemicellulose degradation are kinetic, and usually only appropriate for limited operating regimes and specific species. The study of hemicellulose hydrolysis has yielded substantial data in the literature, enabling a diverse data set to be collected for general and widely applicable machine learning models. In this paper, a dataset containing 1955 experimental data points on batch hemicellulose hydrolysis of hardwood was collected from 71 published papers dated from 1985 to 2019. Three machine learning models (ridge regression, support vector regression and artificial neural networks) are assessed on their ability to predict xylose yield and compared to a kinetic model. Although the performance of ridge regression was unsatisfactory, both support vector regression and artificial neural networks outperformed the simple kinetic model. The artificial neural network outperformed support vector regression, reducing the mean absolute error in predicting soluble xylose yield of test data to 6.18%. The results suggest that machine learning models trained on historical data may be used to supplement experimental data, reducing the number of experiments needed.