Use of Artificial Neural Network Model to Predict Thermal Cracks in Asphalt Concrete Pavements

Improving riding quality, promoting sustainability and resiliency, and providing proper analysis procedures to maintain and predict the performance of existing structures have been profound goals of the transportation sector. Thermal cracks, also universally known as transverse cracks, are considered one of the most prevalent and critical forms of pavement distress in asphalt concrete pavements. Such cracks have been directly linked to various pavement failure modes that adversely impact pavement structures' performance and integrity and the riding quality experienced by typical road users. While recent studies have proven that using regression models to explain thermal cracks is not an accurate representation to quantify distress, linear models are still commonly used in engineering practices. Using Long-term Pavement Performance (LTPP) data from 15 different sections located in the Midwest region of the US, an Artificial Neural Network (ANN) model was developed using MATLAB to predict the count of thermal cracks given the extracted input parameters: average annual temperature, annual average freeze index, 18 Kip ESAL, thermal conductivity, heat capacity, surface shortwave absorption, and coefficient of thermal contraction. The proposed ANN model divides the temperature and distress data: 70% training, 15% testing, and 15% validation data. The predicted and actual outputs were compared by calculating Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). Comparably, the results for 7-9-9-1 ANN structure with TANSIG-LOGSIG transfer functions generated the closest thermal cracking estimate with RMSE of 0.089, MAPE of 0.10, and a regression coefficient (R) of 0.94, which confirmed that the model was adequate to predict thermal cracks in the pavement.