<p>The traditional approaches adopted for calculating the optimum blend of bitumen and aggregates resulting in desired Marshall stability (MS) and Marshall flow (MF) are time-consuming, costly, and complex. This study mapped the relationship between the input variables to predict the MS and MF based on 85 samples collected from four highway projects in Pakistan. The input parameters that are readily available and do not require separate testing were used to develop machine learning models. Three algorithms, namely, Ridge regression, Lasso regression, and Decision tree (DT), were employed to build these models. R-square, adjusted R-square, mean absolute error (MAE), and root mean square error (RMSE) were used to assess the prediction power of each model. Several methods, including k-value computation, score analysis, Taylor’s diagram, and statistical performance assessment approach comparison, ranked the DT regressor as the most effective method for both MS and MF predictions. R-square of 0.982 and 0.986 was recorded for the DT model for training and testing data, respectively. Furthermore, the absolute sensitivity analysis revealed that MS and MF relied more on the percentage of bitumen and aggregate. Based on the predictions made with the DT model, air voids’ effect on MS was more efficacious when the bitumen content was kept between 3 and 3.5%. However, the highest MS was observed for 4.5% air voids and bitumen content between 3% and 3.5%. On the other hand, the lowest MF was observed at 4.0% bitumen content and 4.5% air voids, while the maximum rate of change in MF was observed between 4% and 4.5% bitumen content irrespective of air voids. Increasing the bitumen content increased the MF and reduced the MS after hitting a flex point. Meanwhile, the air voids reduced the MF and increased the MS at a constant bitumen content. Overall, this study presents key insights to automate the prediction of MS and MF using machine learning approaches.</p>