Leakage-controlled target-wise stacking for turbocharged spark-ignition engine response modeling and BSFC stability diagnosis

Accurate prediction of performance and thermal responses is important for calibration support, virtual sensing, and operating-domain assessment in turbocharged spark-ignition engines, where interacting boost, combustion-control, and thermal effects make exhaustive test-bench mapping costly. This study develops a leakage-controlled, target-wise stacking framework for six responses of a gasoline-operated EF7TC engine: turbine inlet pressure, brake power, brake torque, brake specific fuel consumption (BSFC), indicated mean effective pressure, and exhaust temperature. The main novelty is the combination of response-specific learning, group-aware out-of-fold stacking, and a physically informed BSFC stability diagnostic, allowing heterogeneous engine outputs to be evaluated without treating them as statistically equivalent. Using 2871 usable test-bench samples and seven measured operating and control variables, 10 heterogeneous regressors were combined through Ridge and Attention stacking. On the untouched hold-out set, Attention Stacking achieved the highest mean R 2 of 0.980, followed closely by Ridge Stacking at 0.979, while ExtraTrees was the strongest standalone model at 0.966. Formal nested group-aware cross-validation showed no statistically significant difference between the two stacking meta-learners. Target-wise R 2 values reached 0.995 for turbine inlet pressure, 0.999 for brake power, 0.998 for brake torque, 0.998 for IMEP, and 0.986 for exhaust temperature. BSFC remained the most difficult response, with an R 2 of 0.902, and its reduced predictability was associated mainly with low-power and non-positive-power operating regions. The results show that leakage-controlled target-wise stacking provides accurate in-domain prediction while preserving response-specific physical limitations that are important for reliable engineering interpretation.