Improving deep neural network using hyper-parameters tuning in predicting the bearing capacity of shallow foundations

Abstract

ABSTRACT Ultimate bearing capacity is one of the most important parameters in designing shallow foundations. This study focused on developing a hybrid model using Random Search (RS) technique and Deep Neural Network (DNN) to predict the maximum bearing capacity of shallow foundations in sandy soil. The data included 97 load tests on the steps that were used to train and test the model. This data is divided into two parts of the training data set (7%) and the testing set (30%) to build and validate the corresponding models. The performance of the final DNN model is comprehensively assessed with a random hyper-parameters DNN model developed independently using the same data. The values of performance evaluation measures such as R-squared (R2), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and the Variance Accounted For (VAF) are used to determine to get the performance of the DNN model in predicting the ultimate bearing capacity of shallow foundations. In addition, a parallel coordinate plot is utilized to show and evaluate the effect of hyperparameters combination on the performance of DNN model. Besides, a global sensitivity analysis technique was deployed to detect the most important input variables in predicting the ultimate bearing capacity of shallow foundations. This study can provide an effective tool to identify the ultimate bearing capacity of shallow foundations.