Can biomass structural composition be predicted from a small dataset using a hybrid deep learning approach?

Abstract

The accurate determination of biomass composition can enhance the efficacy of biofuel yield evaluations and provide valuable estimates for process economic viability. This study proposes a generative adversarial algorithm (GAN)-based machine learning (ML) model (hybrid deep learning method) for predicting the composition of lignocellulosic biomass, which includes cellulose, hemicellulose, and lignin content. The model employs proximate and ultimate analyses of biomass as input features, while comparing the performance of three different ML models, with random forest (RF) outperforming all other models). The implementation of the GAN significantly enhanced the model's performance. For the test dataset, the R2 values experienced an increase across all biomass components, with cellulose demonstrating the most noteworthy improvement, elevating from 0.7233 to 0.854. To further enhance the predictive capabilities of the model, hyperparameter optimization was performed. Before optimization, the RF model displayed strong correlations between predicted and actual values. However, post-optimization improvements were relatively minimal, with Grid Search and Bayesian Optimization yielding only marginally better results than Randomized Search. Additionally, an interpretable analysis was conducted to provide a deeper understanding of the relationship between proximate and ultimate analysis, and the structural composition of biomass. The findings presented in this study could prove useful in assessing the potential of biomass for the production of bio-based chemicals and biofuels.