<p>Asphalt Mixture Performance (AMP) is critical to the service condition and lifespan of roads. Traditional AMP testing is often hindered by high experimental costs and long design cycles, which significantly reduces research efficiency. To address these limitations and enable rapid and accurate AMP prediction, this study proposes a data-driven approach incorporating model interpretation. Firstly, based on experimental data, three Machine Learning (ML) models were developed: Backpropagation (BP) Neural Network, Extreme Gradient Boosting (XGBoost), and Random Forest (RF). Subsequently, the models were trained and evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²). Additionally, 10 sets of unseen data were used to validate the models’ predictive performance. Finally, Shapley Additive exPlanations (SHAP) was employed for a comprehensive interpretability analysis of the model predictions, quantifying the importance of each material composition feature. Results indicate that the XGBoost model achieved R² values 0.03–0.40 higher than the BP Neural Network and 0.01–0.08 higher than the RF model on both training and testing sets, demonstrating its superior capability in explaining AMP trends. The ML-SHAP integrated method proposed in this study not only enables the accurate prediction of AMP and reduces experimental costs, but also provides clear scientific guidance for material performance regulation and reverse design, demonstrating significant potential for engineering applications in advancing the intelligent design of road materials.</p> Graphical Abstract <p></p>