A new CPT virtual calibration chamber in sand based on Machine learning algorithms

Abstract

Interpretation of Cone Penetration Tests (CPTs) still relies greatly on empirical correlations that are mostly developed in resource-demanding and time-consuming calibration chambers. This paper presents a CPT virtual calibration chamber using machine learning approaches. For such purpose, the multilayer perceptron (MLP) and long short-term memory (LSTM) neural networks are implemented to predict the cone resistance (q_c) profiles under various soil states and testing conditions. The Bayesian optimization (BO) is first adopted to find the optimal neural network hyperparameters of MLP and LSTM. Thereafter, the BO-MLP and BO-LSTM networks are trained with the available data from published datasets. Further comparison and validation of the prediction results are carried out against numerical results obtained from a Coupled Eulerian-Lagrangian (CEL) model. The results show that BO reduces the prediction error of the neural networks by 73.1% (MLP) and 59.5% (LSTM) in the training set as well as 44.4% (MLP) and 40% (LSTM) in the testing set compared to that without BO. The established machine learning models are proven competent to reproduce the q_c profiles with the coefficient of determination (R²) of 98.65% (MLP) and 98.51% (LSTM) in the training set as well as 95.13% (MLP) and 94.65% (LSTM) in the testing set. Apart from matching the numerical model results in terms of accuracy, the proposed methods show a much greater computational efficiency. Eventually, to showcase the use of this new virtual calibration chamber, the predicted q_c are used to obtain a new relationship to predict the relative density, D_r, of the sand. The improved correlation has an R² of 92.7% compared to all data, including those generated by the machine learning method and experiments, and 88.3% compared to the pure experimental data. This is a better generalization than other previously suggested relationships.