Machine Learning–Driven Data Analytics and Multi-Objective Optimization of Waste Plastic Pyrolysis for High-Performance Lubricant Oils, Road Construction Materials, and High-Purity Chemical Feedstocks

The accumulation of unsegregated municipal and industrial waste plastics represents both an escalating environmental burden and an underexploited carbon feedstock. This study presents data-driven investigation of thermochemical (catalytic and non-catalytic) pyrolysis of mixed waste plastics - polyethylene (PE), polypropylene (PP), polystyrene (PS), polyethylene terephthalate (PET), and polyvinyl chloride (PVC) - integrated with machine learning (ML) surrogate modeling and multi-objective evolutionary optimization to simultaneously target three valorization pathways: high-performance lubricant base oils, bitumen-modifying road construction binders, and high-purity aromatic (BTX) chemical feedstocks. A synthetic dataset of 1,240 pyrolysis batch records spanning temperature (400–650 °C), vapor residence time (10–120 min), catalyst type and loading (0–15 wt.%), heating rate, reactor pressure, and feedstock blend ratio was generated to emulate realistic process–property relationships reported in the pyrolysis literature. Seven regression algorithms - multiple linear regression (MLR), support vector regression (SVR), random forest (RF), XGBoost, LightGBM, artificial neural networks (ANN), and Gaussian process regression (GPR) - were trained and benchmarked using 5-fold cross-validation, with XGBoost achieving the best generalization performance (R² = 0.951, RMSE = 2.18 wt.% on held-out test data). SHAP-based interpretability analysis identified pyrolysis temperature, feedstock PE fraction, and residence time as the dominant predictors of liquid oil yield. The trained surrogate models were embedded within a Non-dominated Sorting Genetic Algorithm II (NSGA-II) framework to resolve competing process objectives, generating Pareto-optimal operating windows that jointly maximize oil yield, lubricant viscosity index, and BTX aromatic purity while minimizing specific energy consumption. Downstream property simulations indicate that optimized pyrolysis oil fractions can achieve a viscosity index comparable to Group I mineral base oils (VI ≈ 92), that a 10% pyrolysis-oil-modified bitumen blend can satisfy ASTM/AASHTO penetration-grade specifications for road construction, and that Ni-modified HZSM-5 catalysis can elevate combined BTX yield to approximately 44 wt.% of the recovered oil fraction. The results, while synthetic, illustrate a coherent end-to-end analytics workflow that is directly transferable to real experimental pyrolysis datasets.