<![CDATA[The scarcity of experimental data for diesel engines fueled by waste plastic oil (WPO) is a critical obstacle to optimizing engine performance. In this study, only 42 experimental data points covering six blend ratios and seven load conditions were available. To overcome this limitation, 121 synthetic data points were generated by training a suite of machine‑learning models—Random Forest, Gradient Boosting, and AdaBoost—on the original dataset and then predicting outputs across a grid of WPO blend ratios (0–50% in 5% increments) and engine loads (0–100% in 10% increments). The synthetic data were rigorously validated using Kolmogorov–Smirnov tests, kernel density estimation, and principal component analysis to ensure statistical similarity with the original measurements. Subsequently, a Multi‑Input Multi‑Output (MIMO) deep neural network was trained on the combined real and synthetic dataset to predict four key performance metrics—power, torque, specific fuel consumption (SFC) and brake thermal efficiency (BTE)—and its hyperparameters were fine‑tuned using Bayesian optimization via Optuna, achieving coefficients of determination (R²) above 0.95. Optimization analysis indicated that a 17% WPO blend at 82% load delivers the best trade‑off between power, efficiency and fuel consumption for non‑road applications. This integrated framework demonstrates how synthetic data generation, rigorous validation and deep‑learning modelling can effectively mitigate data scarcity and provide actionable insights for performance optimization of plastic pyrolysis oil in diesel engines. ]]>
