Multi-objective optimization of engine performance and emissions using a CuO–E10 nanofluid: A hybrid modeling framework integrating machine learning, response surface methodology, and NSGA-II/TOPSIS

While nanofluid additives alter internal combustion engine characteristics, integrating experimental data with interpretable hybrid modeling and multi-objective optimization requires further assessment relative to conventional modeling techniques. This study evaluates and optimizes spark-ignition engine parameters using CuO–E10 fuel blends. Experiments measured engine responses at CuO concentrations of 10–40 ppm and speeds of 1500–3000 rpm. Hybrid algorithms combining machine learning and response surface methodology (RSM) were developed, utilizing two-level SHAP analysis to quantify individual model contributions. Three evolutionary algorithms were then compared via TOPSIS decision-making across three operational scenarios: fuel economy, maximum performance, and urban emission control. Results indicate maximum brake torque (BT) increased by 8.27% (at 2500 rpm and 30 ppm), and minimum brake-specific fuel consumption (BSFC) decreased by 4.81%. CO and HC emissions were reduced by 18.18% and 7.69%, respectively, whereas CO 2 and NO x increased. The Gradient Boosting Machine (GBM)–RSM model produced the highest accuracy for BT, CO 2 , and NO x , while Extreme Gradient Boosting (XGB)–RSM was optimal for BSFC and CO. Quantitatively, XGB–RSM reduced the root mean square error (RMSE) for CO by 72.12% compared to standalone RSM, and Random Forest (RF)–RSM reduced the mean absolute error (MAE) for HC by 78.8% versus standalone RF. SHAP analysis identified engine speed as the dominant variable, showing that RSM preserves the physical baseline while ML corrects data nonlinearities. NSGA-II outperformed NSGA-III and MOEA/D, successfully identifying balanced Pareto optimal points for each designated operational scenario.