Machine Learning-Based Classification of Lignocellulosic Biomass from Pyrolysis-Molecular Beam Mass Spectrometry Data

Area of Science:

• Biomass analysis
• Analytical chemistry
• Computational biology

Background:

• High-throughput biomass analysis is crucial for agricultural and bioenergy feedstocks.
• Pyrolysis-molecular beam mass spectrometry (py-MBMS) is a popular technique for rapid biomass composition analysis.
• Diverse data analysis tools are often required for py-MBMS data, depending on the application.

Purpose of the Study:

• To analyze py-MBMS spectra of lignocellulosic biomass to understand spectral patterns and their relation to biomass composition.
• To apply machine learning (ML) approaches for classifying, differentiating, and predicting biomass types based on py-MBMS spectra.
• To develop a data analysis workflow applicable to various biomass samples and experimental conditions.

Main Methods:

• Lignocellulosic biomass samples were analyzed using py-MBMS.
• Generalized linear modeling (GLM) was used to correct spectra for instrumental variance using spike-in controls.
• Various ML classifiers, including random forest, k-nearest neighbor (k-NN), decision tree, Gaussian Naïve Bayes (GNB), gradient boosting, and multilayer perceptron, were employed.
• Performance of ML models was evaluated for classification and prediction of biomass types and mixtures.

Main Results:

• The k-NN classifier performed best on raw spectral data for classification, while the decision tree performed worst.
• After spectral normalization for instrumental variance, all classifiers showed comparable performance for biomass type prediction.
• GNB and extreme gradient boosting (XGB) classifiers demonstrated superior performance in predicting biomass mixtures compared to k-NN and decision tree.
• The study identified spectral patterns and variations associated with biomass composition.

Conclusions:

• Machine learning, particularly GNB and XGB, offers a robust approach for analyzing py-MBMS data to classify and predict biomass types and mixtures.
• Spectral normalization using GLM improves the accuracy of ML models for biomass analysis.
• The developed data analysis workflow can be extended to diverse biomass studies, aiding in understanding compositional variations and inferring feedstock identity without prior knowledge.