Analysis of Marshall's Stability of Hot Mixtures with Machine Learning

Road transport is the most widely used transport network globally and in Turkey. Every year, asphalt mix designs are prepared for the deterioration or reconstruction of road pavements. Since asphalt mixture designs affect the mechanical and physical properties of roads, they also directly affect their performance. For this purpose, Marshall design, one of the experimental methods, is used. The calculation of design parameters used in mixture design is complex, time consuming and costly. To address this, developing models that predict experimental results using machine learning methods offers an economical and practical approach. A data set consisting of 13 variables including specific gravities, aggregate proportions and bitumen content was created for 453 hot mix asphalt specimens subjected to the Marshall test. Marshall Stability (MS) was determined as the target variable in the model. Random Forest, XGBoost, GBoost and Extra Trees algorithms were used to predict MS; models were compared with performance metrics such as MAPE, R² and RMSE. The XGBoost model was the most successful model for MS prediction; it provided high accuracy and low error rate with R² (coefficient of determination) of 0.96, RMSE (root mean square error) of 0.37 and MAPE (mean absolute percentage error) of 0.02% on the test data. SHAP analysis was performed to determine the variables affecting MS prediction, and the most effective variables were found to be Gef (effective specific gravity of aggregate) and Gsb (bulk specific gravity of aggregate). Such explainability analyses facilitate decision making in the early design stages by supporting the optimization of design parameters.