Applied machine learning multi-model benchmarking to plastic waste pyrolysis for optimizing crude oil yield production

Optimizing crude fuel yield in catalytic pyrolysis of non-recyclable plastic waste remains a major challenge due to high-dimensional process variables and costly experiments. Machine Learning (ML) promises a data-driven solution, but published studies often report inflated performance from data leakage or improper Machine Learning applications. This study implements a leakage-free ML pipeline via a seven-stage Kanban workflow process applied to 754 pyrolysis experiments. 18 supervised ML algorithms, spanning neural networks, tree-based boosters, and ensemble methods, were evaluated. Tailored preprocessing imputation, dimensionality evaluation via PCA/KPCA, and synthetic augmentation were confined to training folds before stratified splitting. Model performance was assessed using MAE, MSE, RMSE, R 2 , residuals, distribution analyses, and Q–Q plots under five-fold cross-validation. EvoTree Regressor was implemented in Julia for computational efficiency, while all other stages remained in Python. Also, real experimental reactor conditions were optimized via Particle Swarm Optimization (PSO), benchmarked against Random Search, yielding a 56.95 % liquid fraction (baseline 20 %). Results of crude oil show a calorific value of 17.94 kJ/g and density of 0.768 g/cm 3 in 2 real experiment trials. FTIR analysis confirmed structural differences under optimized conditions. All code and data are openly available, establishing a reproducible, leakage-free framework to accelerate sustainable waste-to-fuel research. • Leakage-free ML pipeline ensures reproducible plastic pyrolysis yield prediction. • Eighteen ML models benchmarked under leakage and clean conditions. • Random Forest achieved best leakage-free performance (R 2 = 0.74). • PSO optimization reached 56.95 wt% yield, doubling reactor baseline. • FTIR confirmed hydrocarbon-rich oil with 17.94 kJ/g calorific value.