Energy-oriented automated machine learning framework for process optimization and product prediction in waste tire pyrolysis

The escalating volume of waste tires poses a severe environmental threat due to their non-degradable polymer structure. While pyrolysis offers a green pathway for resource recovery, its process optimization is hindered by complex multi-parameter coupled reactions. However, traditional machine learning approaches often suffer from poor generalization and overfitting when processing such heterogeneous feedstocks. To address this, this study establishes a ‘Robustness-Oriented’ framework by comparing single-algorithm models with Automated Machine Learning (AutoGluon & H2O). Comparative evaluation reveals that while the H2O framework achieves superior fitting on training data, it exhibits overfitting tendencies. In contrast, the AutoGluon model demonstrates robust generalization capability and stability, achieving test set R2 values of 0.937 and 0.899 for oil and gas yield predictions, respectively. SHAP (SHapley Additive exPlanations) analysis quantitatively disentangles complex feature interactions, identifying Temperature, oxygen, and sulfur as decisive factors; specifically, temperature regulates reaction kinetics, while intrinsic heteroatom content dictates the chemical composition. Multi-objective optimization pinpoints the FCC (fluid catalytic cracking catalyst) and a 540–550 °C temperature window as optimal, predicting a maximum oil yield of 52.25 wt % and an estimated energy recovery density of 21.9–23.0 MJ/kg. Crucially, preliminary energy balance estimates indicate that these optimized conditions achieve a significant net energy surplus, effectively converting solid waste into high-density liquid energy. These findings provide a quantifiable, data-driven basis for the intelligent prediction and green optimization of pyrolysis processes, supporting the advancement of “zero-waste city” development.