Developing a prediction model for low-temperature fracture energy of asphalt mixtures using machine learning approach

This paper presents an augmented full quadratic model (AFQM), artificial neural network (ANN) and an innovative machine learning technique called self-validated ensemble modelling (SVEM) approaches to predict low-temperature fracture energy of asphalt mixtures. An experimental database including 852 fracture energy values obtained from low temperature disk-shaped compact tension (DCT) testing was utilised to develop the prediction models. The fracture energy was predicted in terms of several variables that are available during the mix design process. The collected data were categorised into three groups based on the availability of the data at different points during the mix design process. A sensitivity analysis was conducted to assess the impact of the design variables on fracture energy. Based on the model development results, both ANN and SVEM methods showed higher prediction accuracy than AFQM. Prediction models based on the ANN were time-consuming and computationally expensive due to the optimum model architecture. The SVEM technique was found to be a reliable prediction method with high prediction reliability even with a limited amount of data. Based on the sensitivity analysis, design traffic level, PG low temperature (PGLT) binder grade, amount of aggregate passing 9.5 mm sieve, and the voids in mineral aggregate (VMA) are the most effective factors impacting low-temperature asphalt mixture fracture energy. A web-based prediction model platform was developed using prediction models based on the SVEM technique which can be utilised as a predesign tool to evaluate low-temperature fracture energy of asphalt mixtures when laboratory testing is not feasible.