Machine learning based prediction of TOPSIS scores for internal combustion engine performance and emissions

Abstract

Internal combustion engines operating on alternative fuels face an inherent challenge: optimizing power output and fuel efficiency while simultaneously minimizing exhaust emissions requires the simultaneous management of multiple, often conflicting objectives. Although multi-criteria decision-making (MCDM) and machine learning (ML) methods have each been applied extensively in engine research, they are typically used in isolation, and the direct prediction of a composite MCDM decision score from engine control parameters has not previously been demonstrated. This study addresses that gap by proposing a hybrid TOPSIS-Stacking framework in which a TOPSIS-based composite score, aggregating eight performance and emission indicators into a single figure of merit, is itself treated as a continuous regression target. Experimental data from a single-cylinder Otto-cycle engine operating across a range of ethanol-water blend ratios (0 to 50% water by volume), two compression ratios (7.44:1 and 9.44:1), and five rotational speeds (2000 to 4000 rpm) were used. Entropy-based weighting was adopted to assign objective criterion importance, and three regression models (RSM, XGBoost, and a stacking ensemble combining both) were trained to predict the TOPSIS score from four engine control parameters. A 5-fold Group Cross-Validation strategy was implemented to prevent repetition leakage between experimental replicates and ensure conservative generalization estimates. The stacking ensemble achieved the highest predictive accuracy (R² = 0.814, RMSE = 0.061, MAPE = 6.36%), outperforming both individual models. Feature importance analysis identified water fraction as the dominant predictor, followed by compression ratio, consistent with their thermodynamic influence on NO x formation and thermal efficiency. These results demonstrate that the proposed framework provides a reliable and computationally efficient tool for virtual engine calibration and scenario-based performance evaluation, significantly reducing the need for exhaustive multi-criteria experimental campaigns.