Optimizing aromatics yield-selectivity trade-off in zeolite-catalyzed biomass fast pyrolysis using machine learning

Catalytic fast pyrolysis of biomass to aromatic hydrocarbons is an attractive valorization technology to produce high-value chemicals and biofuels. However, process optimization to produce desired products remains challenging due to reaction complexity. Herein, we present a data-driven approach to address this issue. A literature dataset of ZSM-5-based catalysts, including promoted and unpromoted systems is compiled, resulting in 254 datapoints and subjected to data-mining. A random forest model with uncertainty quantification is devised to predict the aromatic hydrocarbon yield, and benzene, toluene, and xylene (BTX) selectivity. The models achieved comparable performance ( R 2 > 0.80) and feature-importance analysis identified feedstock composition, catalyst total acidity, catalyst-to-biomass ratio, and space velocity as key drivers of performance metrics. Given the inherent trade-off between aromatic yield and BTX selectivities, an optimization framework was devised to screen over 20,000 promoted-zeolite and reaction condition combinations, uncovering Cu, Ga, Sn, and Zr promoted-ZSM-5 as Pareto-optimal catalysts. Lastly, motivated by challenges encountered during data collection and model limitations in out-of-box predictions, we urge the community to report biomass fast pyrolysis protocols in standardized manner, and provide effective guidelines for the cause. Our study provides a predictive toolkit to identify promising catalytic systems and optimize biomass fast pyrolysis performance while improving machine readability.