Accurate and computationally efficient prediction of heavy-duty diesel-engine performance and emissions is important for engine calibration, real-time emissions estimation, and model-based control. This study compares five prediction methods: multiple linear regression (MLR), Kriging, an artificial neural network (ANN), a one-dimensional convolutional neural network (1D-CNN), and random forest (RF). A common steady-state experimental dataset was used for the comparison. Ten input variables representing engine operation, air-path conditions, fueling, and thermal behavior were used to predict brake thermal efficiency (BTE), diesel oxidation catalyst (DOC) inlet temperature, and engine-out nitrogen oxides (NOx), soot, carbon monoxide (CO), and carbon dioxide (CO₂). Model performance was evaluated using the coefficient of determination (R²) and root mean square error (RMSE) with a 70/15/15 training–validation–test split. Based on the reported test-set R² values, the ANN showed the strongest overall performance, exceeding 0.96 for DOC inlet temperature (0.9978), CO₂ (0.9952), BTE (0.9927), NOx (0.9872), and CO (0.9795). DOC inlet temperature was predicted accurately by all five models, with R² values ranging from 0.9832 to 0.9978. Soot was more difficult to predict, with R² values between 0.3608 and 0.7781, reflecting its complex dependence on fueling, mixture formation, temperature, and oxygen availability. MLR performed particularly poorly for CO and soot, whereas the nonlinear models captured their behavior more effectively. These findings illustrate how model choice affects predictive performance and provide useful guidance for selecting efficient data-driven models for diesel-engine calibration and emissions monitoring.