EXPLAINABLE MACHINE LEARNING FRAMEWORK FOR CATALYTIC PYROLYSIS PARAMETER OPTIMIZATION IN SUSTAINABLE HYDROGEN PRODUCTION

Abstract

This study proposes an explainable machine learning framework for optimizing catalytic pyrolysis parameters for sustainable hydrogen production from biomass. The framework is designed to model the highly nonlinear relationship between feedstock composition, catalyst properties, and operating conditions that collectively govern hydrogen yield in catalytic pyrolysis systems. A structured experimental dataset is constructed from reported catalytic pyrolysis studies and includes biomass-related variables, catalyst descriptors, and reaction parameters as model inputs, with hydrogen yield as the continuous prediction target. Multiple supervised regression algorithms are evaluated to identify the most suitable predictive model, and comparative analysis shows that ensemble-based methods outperform conventional baseline models in capturing the multivariable and interaction-driven nature of the process. The selected model is further examined using SHAP-based explainable artificial intelligence to quantify the contribution of individual features and to reveal the dominant influence of reaction temperature, nickel loading, catalyst support, and biomass hydrogen content on hydrogen-production behavior. To move beyond forward prediction, the trained model is integrated into an optimization framework that explores the process-variable space and identifies parameter combinations associated with improved hydrogen yield under practically meaningful operating bounds. The results demonstrate strong agreement between predicted and observed hydrogen-yield values, stable residual behavior, and a well-defined high-performance operating region in the optimization landscape. The explainability analysis confirms that the learned variable importance is physically consistent with known catalytic and thermochemical mechanisms, thereby strengthening the scientific credibility of the framework. Overall, the proposed approach provides a predictive, interpretable, and optimization-oriented methodology for catalytic pyrolysis analysis, offering a practical pathway to reduce experimental trial-and-error, improve process understanding, and support the rational design of more efficient and sustainable biomass-based hydrogen-production systems.