Hybridizing Deep Neural Network Models to Predict Undrained Shear Strength from CPT Data

Many geotechnical models have been developed to predict soil behavior; however, they often fail to establish accurate correlations between design parameters and geotechnical investigations, leading to reduced prediction accuracy. Such limitations can result in technical challenges, increased construction costs, and project delays. In cohesive soils, the undrained shear strength Su is a critical parameter influencing several geotechnical applications. This study employs a machine learning (ML) model based on the Bayesian method to predict Su from cone penetration test (CPT) data. Models with nine and four input variables and point-estimated predictions were developed using Bayesian Deep Neural Networks (BDNN). Their performance was assessed through multiple metrics, including mean absolute error (MAE), mean squared error (MSE), R-squared score, variance accounted for (VAF), and the a20-index, achieving an R2 of 0.814 for the training dataset and 0.869 for the testing dataset with nine inputs, and an R2 of 0.743 for the training dataset and 0.766 for the testing dataset with four inputs. These results highlight the potential of ML, particularly Bayesian-based approaches, to improve geotechnical predictions while ensuring cost efficiency.