An eXtreme Gradient Boosting model for predicting dynamic modulus of asphalt concrete mixtures

Abstract Dynamic modulus (DM)—an important stiffness property and a basic input parameter in the design of flexible pavements—is often obtained through expensive, laborious, and time-taking laboratory tests, which can be subjected to human error as well. Predictive models are, therefore, developed in lieu of laboratory testing, enabling the estimation of DM values with some degree of error. Focussing on minimising this error in predicting the DM of asphalt concrete (AC) mixtures, this study proposes an eXtreme Gradient Boosting (XGBoost) approach for modelling DM. Sixteen AC mixtures are prepared in laboratory and tested for DM at different testing temperatures and loading frequencies. The data obtained from the laboratory test are divided into three categories: testing conditions, mix volumetric properties, and gradation, which served as the input to the XGBoost model. More specifically, testing conditions include four testing temperatures (4.4, 21.1, 37.8, and 54.4 °C) and six loading frequencies (25, 10, 5, 1, 0.5, and 1 Hz), mix volumetric properties entail optimum bitumen content, air voids, voids filled with aggregate and asphalt, stability and flow of a mix, binder type, and type of layer, gradation parameters consist of per cent passing of different sieve sizes. Several goodness-of-fit measures are employed combined with k-fold validation technique to robustly evaluate the performance of the developed model. Furthermore, the XGBoost model is compared with some well-known regression-based models as well as other machine learning approaches. The comparison analysis reveals that the XGBoost model outperforms its competing models from the literature by showing a higher degree of accuracy in predicting the observed DM values. Moreover, to further demonstrate the efficacy of the proposed XGBoost model, its transferability is tested for a completely new dataset, and the findings suggest that the model is able to capture the difference in mix properties and preparation type of the new dataset with a reasonable accuracy. In addition, the XGBoost model exhibits a superior performance in transferability analysis compared with its competing models. The findings of this study advocate the use of XGBoost approach for predicting DM values as well as in the design of flexible pavements in future studies.