Most of the existing soil-water characteristic curve (SWCC) prediction models do not have a high level of prediction accuracy. The R2 values of these model predictions range from 0.1 to 0.6 when applying them to a large data set. The inaccurate prediction of SWCC diminishes the prediction accuracy of engineering properties of unbound material. To overcome this issue, the goal of this study was to improve the prediction accuracy of SWCC using an artificial neural network (ANN) approach. Two three-layer ANN models were constructed for plastic and nonplastic soils separately, which consisted of one input layer, one hidden layer, and one output layer. The input variables included soil gradation indicators, particle diameter indicators, Atterberg limits, saturated volumetric water content, and climatic factors. The hidden layer, including a total of 20 neurons, used a log-sigmoidal function as a transfer function and the Levenberg-Marquardt back propagation method as the training algorithm. The output layer variables were the fitting parameters of the Fredlund-Xing equation. The SWCC database from the NCHRP 9-23A project was used to develop ANN models with 80% of the data set for training and 20% of the data set for validation. The developed ANN models had R2 values between 0.91 and 0.95 for predicting the SWCCs of unbound material, which are significantly higher than other regression models. Finally, the developed ANN models were validated by comparing a new data set collected from both the NCHRP 9-23A project and other literature sources to the model predictions.
|Journal of Materials in Civil Engineering
|Early online date
|19 Feb 2018
|Published - 1 May 2018