Enhanced bio-oil production from biomass catalytic pyrolysis using machine learning

Abstract

This study leverages machine learning technology, coupled with an evolutionary algorithm, to forecast and optimize the distribution and composition of products from in-situ biomass catalytic pyrolysis. Among the four machine learning models employed, the ensemble learning model emerged as the frontrunner, demonstrating superior prediction performance (R² > 0.89, RMSE <0.03, and MAE <0.01) compared to generalized additive, support vector regressor, and artificial neural network models. Multi-objective optimization results favored catalyst-to-biomass ratios near unity for bio-oil production, with optimal catalyst acid site content ranging from 0.04 to 2.49 mmol/g for various bio-oil applications. For energy applications, the optimal parameters yielded a bio-oil with 63.36 wt% hydrocarbon content and a bio-oil yield of 41.49 wt%. For chemical applications, the optimized parameters resulted in a bio-oil with 60.63 wt% phenolic content and a bio-oil yield of 48.93 wt%. For pharmaceutical applications, the bio-oil contained 10.42 wt% aldehydes and 21.49 wt% ketones, with a bio-oil yield of 36.56 wt%. Feature importance analysis revealed that biomass properties and catalyst characteristics could significantly influence process modeling, accounting for 61.3 % and 24.7 % of the impact on bio-oil yield, respectively, while operating conditions showed the slightest effect. These findings provide valuable insights for future experimental studies, enabling the optimization of in-situ biomass catalytic pyrolysis for energy, chemical, and pharmaceutical applications. Moreover, the feature importance analysis enhances understanding of the complex in-situ catalytic pyrolysis process, guiding the design of more efficient pyrolysis reactors and contributing to sustainable biofuel and biochemical production technologies.