Comparative investigation of microwave-assisted and conventional pyrolysis: a machine learning-based approach

Pyrolysis can convert biomass into renewable energy carriers and chemicals. While microwave-assisted pyrolysis (MAP) enables volumetric heating with higher energy efficiency than conventional pyrolysis (CONV) relying on conductive-convective heat transfer, the benefits of MAP with respect to product profiles have yet to be conclusively demonstrated. There is still a lack of coherent understanding of how feedstock properties and operating conditions differentially shape pyrolysis outcomes under MAP versus CONV. To address this gap, we examine literature data on sludge pyrolysis with multiple machine-learning algorithms and SHapley Additive exPlanations (SHAP) analysis, systematically comparing the key input features impacting product distributions in MAP and CONV processes. Model comparison indicates that the ridge regression model offers an appropriate balance between generalizability and predictive performance, and was selected for subsequent analyses. Both feature importance and SHAP analyses consistently reveal that, in CONV, product partitioning is predominantly explained by proximate composition (e.g., fixed carbon and ash), which collectively account for 36-68% of the total feature contribution across gas, liquid, and solid products. In contrast, in MAP, ultimate composition (elemental analysis, e.g., C, H, and N) emerges as a more powerful descriptor, contributing up to 54% of the explained variance in product distributions. Furthermore, the dielectric loss tangent of microwave-absorbing additives is identified as a pivotal factor that ranks among the top three influential variables in MAP, plausibly operating independently of the bulk temperature considered in this study. Our findings offer new insights into the fundamental similarities and differences between the control regimes of the two pyrolysis methods.