Cone Penetration Test Prediction Based on Random Forest Models and Deep Neural Networks

• Published in: Geotechnical and geological engineering Vol. 41; no. 8; pp. 4595 - 4628
• Main Authors: Pacheco, Vinicius Luiz, Bragagnolo, Lucimara, Dalla Rosa, Francisco, Thomé, Antonio
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.11.2023; Springer Nature B.V
• Subjects: Algorithms; Artificial neural networks; Civil Engineering; Coefficients; Cone penetration tests; Decision trees; Deep learning; Design parameters; Earth and Environmental Science; Earth Sciences; Geotechnical engineering; Geotechnical Engineering & Applied Earth Sciences; Hydrogeology; Learning algorithms; Machine learning; Mathematical models; Neural networks; Original Paper; Parameters; Physical properties; Prediction models; Quality control; Sleeves; Soil classification; Soil conditions; Soil properties; Terrestrial Pollution; Waste Management/Waste Technology; Geotechnical properties; Deep learning neural network; CPT; Machine learning; Tip resistance; Sleeve resistance
• ISSN: 0960-3182; 1573-1529
• DOI: 10.1007/s10706-023-02535-0

The cone penetration test (CPT) is widely used in soil characterization and the determination of physical parameters. The traditional interpretation of the results of CPTs relies on the experience and expertise of geotechnical engineers. However, recent innovations in machine learning have led to the development of predictive models that can accurately predict soil properties based on CPT data. The application of these techniques can provide more accurate and consistent predictions, even for complex soil conditions. Therefore, this article sought to evaluate the performance of two different machine learning algorithms, random forest and deep learning, with CPT test data to predict tip resistance (q c ) and sleeve resistance (f s ) based on soil classification inputs. The work was conducted with a database of tests in the regions of Germany and Austria, initially consisting of more than two million related observations. This allowed for an assessment of model generalization across different regions. The random forest regressor algorithm presented a coefficient of determination of 0.94 for tip resistance (q c ) and 0.82 for sleeve resistance (f s ) prediction, thus outperforming deep neural networks. The study applied the model to obtain coefficients of determination between 0.65 and 0.68 for tip resistance (q c ) and 0.14 to 0.75 for sleeve resistance (f s ) for different regions of testing. Practical implications include the possibility of obtaining the design parameters q c and f s from inputs obtained from simpler tests, which would reduce project costs, improve the quality and efficiency of CPTs, and assist in making decisions about geotechnical projects.