Prediction of standard penetration test value on cohesive soil using artificial neural networks

Abstract

Soil investigation is the main key in starting construction. Standard Penetration Test (SPT) and Cone Penetration Test (CPT) are field tests often used to estimate soil parameters for foundation design purposes. The SPT value (N-SPT) shows a correlation between the CPT value and other soil parameters. At present, there have been many conventional correlations examining these correlations, but the nonlinear nature of the soil due to very complex soil formations means that this correlation cannot be used in all situations. This research aimed to predict the value of SPT on cohesive soil based on CPT test data and soil physical properties using artificial neural network capabilities using the Backpropagation algorithm, and the activation function was bipolar sigmoid. This study used 284 data from several places in Sumatra Island, Indonesia, with data input were tip resistance, shaft resistance, effective overburden pressure, percentage of liquid limit, plastic limit, sand, silt, and clay. The results showed that the training data of RMSE was 3.441, MAE and R2 were 0.9451 and 2.318, respectively while test data showed RMSE, MAE, R2 were 2.785, 2.085, and 0.9792, respectively. It means that the proposed artificial neural network NN_Nspt(C) was promising to predict the N-SPT value with a minimum error value and a strong regression equation.