The persistent use of internal combustion engines in the transportation sector underscores a growing need for efficient methods to optimize their performance and emissions calibration. Conventional calibration methods, which rely heavily on experimentation, are often time-consuming, costly, and incapable of capturing the complex, non-linear interactions among critical parameters. This study presents a hybrid methodology integrating Design of Experiments (DOE) and Machine Learning (ML) to model Brake Specific Fuel Consumption (BSFC) and HC and NOx emissions in a turbocharged gasoline engine. Two advanced ML algorithms were trained using an optimized DOE based on the Fedorov algorithm. Their predictive performance was systematically evaluated and compared across varying data volumes, ranging from 12.5% to 50% of the design space. Under limited data conditions (25% of data), both models demonstrated stable performance for predicting HC and NOx emissions, attributable to their inherent robustness against overfitting. Nevertheless, the prediction errors for these pollutants remained significant compared to those for BSFC, highlighting the necessity for further model refinement and experimental validation. When the data volume was increased to 50%, both models achieved high predictive accuracy, with mean R² values of 99% for BSFC, 96% for NOx, and 90% for HC. This approach demonstrates that a desired calibration accuracy can be achieved using only 50% of the experimental data, thereby potentially reducing testing time and associated costs by half.