The Cone Penetration Test (CPT) is an in-situ soil testing technique used to determine subsurface soil properties. This information can help assess bearing capacity, liquefaction, and soil stratigraphy. CPT measures geophysical properties relevant to infrastructure projects by utilizing tip resistance (qc) and sleeve resistance (fs).
Performing field tests can be costly and time-consuming, which is why various numerical techniques, such as finite element analysis and feature correlation, can provide a more efficient determination of CPT results. To fully leverage the benefits of these numerical techniques, customization may be essential. Different projects often have unique requirements based on geological conditions, environmental factors, and regulatory standards. Tailoring analytical models to address industry-specific needs can enhance their accuracy and applicability.
Despite the advantages of numerical methods, conventional approaches still face limitations, particularly when dealing with complex geotechnical challenges. A reliance on generalized models may not adequately account for local conditions, leading to inaccuracies in predictions and potential oversights in design considerations.
Recent advancements in artificial intelligence techniques have significantly enhanced their popularity and established them as valuable alternatives for addressing non-linear complexities without the assumptions often associated with traditional numerical methods, analytical techniques, laboratory tests, and field evaluations.
A study conducted by Vinicius Luiz Pacheco et al., at the University of Passo Fundo, Brazil, focused on predicting cone penetration test (CPT) tip resistance (qc) and sleeve resistance (fs) utilizing Random Forest (RF) and Deep Neural Networks (DNNs). The research was based on a dataset comprising 1339 samples, which included various input parameters: the cone resistance adjusted for water effects (qt), cone area ratio (a), current in-situ total vertical stress (σv0), current in-situ effective vertical stress (σ’v0), penetration pore pressure (u2), current in-situ equilibrium water pressure (u0), and excess penetration pore pressure (Δu).
Before proceeding with the modelling phase, it is essential to establish optimal hyperparameter settings to ensure that the model produces accurate and reliable results. To achieve this, the author employed a grid search technique for hyperparameter optimization. For the Random Forest (RF) model, the optimal hyperparameters identified are as follows: Number of estimators: 400; Maximum depth: 300; Minimum samples split: 15; Minimum samples leaf: 2. For the Deep Neural Network (DNN), the optimal settings include: Learning rate: 0.001; Optimizer: Nadam; Activation function: ReLU; Initializer: He Normal; Number of hidden layers: 6; Number of neurons per layer: 21; Epochs: 500; Batch size: 64; Validation split: 0.25.
The study aims to predict two outputs: tip resistance (qc) and sleeve resistance (fs). The Random Forest regressor algorithm demonstrated a coefficient of determination of 0.94 for the prediction of tip resistance (qc) and 0.82 for sleeve resistance (fs), thereby outperforming the deep neural networks in this context.
By implementing the Random Forest (RF) algorithm on previously unseen data, this study provided valuable insights into the algorithm’s ability to generalize across different geographical regions and datasets. This approach not only tested the robustness of the algorithm but also highlighted its potential applicability to soil data collected from diverse locations worldwide. Future research could enhance the model’s accuracy by incorporating Cone Penetration Test (CPT) data into the original dataset. This integration has the potential to refine the algorithm further, improving its performance and adaptability to various soil types and environmental conditions. Such enhancements could lead to more reliable predictions and analyses, catering to a broader array of applications in geotechnical engineering and environmental assessments.
Reference
Pacheco, V.L., Bragagnolo, L., Dalla Rosa, F. et al. Cone Penetration Test Prediction Based on Random Forest Models and Deep Neural Networks. Geotech Geol Eng 41, 4595–4628 (2023). https://doi.org/10.1007/s10706-023-02535-0