Prediction of the High-Temperature Performance of a Geopolymer Modified Asphalt Binder using Artificial Neural Networks

摘要

Complexity in the behaviour of an asphalt binder is further escalated with geopolymer (fly ash and alkali liquid) modification, thus making it difficult to accurately predict the performance of the binder. This study employs artificial neural network modelling to predict the complex shear modulus, storage modulus, loss modulus and phase angle outcomes of experimental results from dynamic shear rheometer (DSR) oscillation tests under four separate scenarios. The proposed artificial neural network models received test conditions (temperature and frequency) and three different geopolymer concentrations (3%, 5% and 7% by the weight of bitumen) as the predictor parameters.? The variants of the optimal algorithms were Levenberg-Marquardt (LM), Scaled conjugate gradient and Polak-Ribiere conjugate gradient (CPG) training algorithms with different combinations of network structures and tan-sig and log-sig as activation functions. The coefficient of determination, covariance and root mean square error (RMSE) were used as statistical measures of model prediction performance. Based on the statistical performance indicators, the LM algorithm with a 3-5-1 network architecture and tan-sig as the activation function was the best performing model for predicting the complex modulus with R2 values of 0.996 for the training dataset and 0.971 for the testing dataset and RMSE values of 0.118 and 0.139 for the training and testing datasets, respectively. Furthermore, it was observed that the least efficient model was the phase angle prediction model developed with the CPG training algorithm, which had a 3-8-1 network architecture and log-sig as the activation function.? The model yielded R2 values of 0.909 and 0.829 for the training and testing datasets, respectively. Poor prediction performance for the testing dataset indicated that the model was unable to learn complexity in the data and would perform below a significance level of 0.90 in predicting using untrained data.