Optimizing asphalt mix design through predicting effective asphalt content and absorbed asphalt content using machine learning

Superpave mix design procedure is still empirical. Random and lengthy trials in a laboratory during the design procedure can consume immense manpower and materials resources. To simplify and accelerate the Superpave mixture design procedure, this study explored the use of machine learning models to achieve an automated and timesaving mix design method through predicting effective asphalt content (P-be) and absorbed asphalt content (P-ba). Five different machine learning (ML) algorithms including Support vector regression (SVR), bagging with ridge regression, Random Forest, Adaptive boosting (AdaBoost), and Gradient boosting were trained. The inputs related to mixture design include aggregates gradation, the bulk specific gravity of aggregates, blend absorption, air void, PG grade of the asphalt binder, and the number of gyrations (N-des). The kernel density estimation was used to detect outliers of the datasets. The effect of different methods used to encode PG grade, which is categorical variables, on the performance of ML models was analyzed. The performances of ML models were evaluated by calculating different performance scoring metrics, such as coefficient of determination (R-2), root means square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). The improved mix design procedure based on an ML model was proposed. The result showed that compared to other encoding methods, the ordinal encoding method used to encode the variables of PG grade can achieve the best performance of ML models. The Gradient boosting algorithm (R-2 = 0.9479, 0.9459) was found as the best estimator to predict the P-be, and P-ba, respectively, compared with those obtained by other machine learning models. Furthermore, determining P-b through predicting P-be and P-ba indirectly with ML models was more accurate than predicting P-b directly with ML models. Therefore, the Gradient boosting model was used in the improved mix design procedure. A case study was implemented to verify the feasibility of the proposed mix design method. The gradation and the optimal asphalt content obtained from the proposed method with the Gradient boosting model was consistent with that obtained by the traditional Superpave method with laboratory tests.

Automated summary

This study explores the use of machine learning models to simplify and accelerate the Superpave mixture design procedure. The model predicts effective and absorbed asphalt content to achieve an automated and timesaving mix design method. The study compares different machine learning algorithms and encoding methods, and finds that the Gradient boosting model performs the best in predicting asphalt content. The proposed method is verified to be consistent with the traditional Superpave method through a case study.