Application of Data-Driven Techniques in 3D Soil Characterization for Top-Hole Design

Abstract Present work investigated the application of 3D Sparse Bayesian Learning (SBL) and Gaussian Process Regression (GPR) for data-driven offshore site characterization focusing specifically on estimating critical soil parameters for top-hole design in oil and gas wells, such as undrained shear resistance and specific unit weight. These parameters are essential for designing conductor and surface casing sections in drilling operations and ensuring the structural integrity of wellhead systems and casing strings. This study used datasets of piezocone penetration tests (CPTu) conducted in the Campos basin in Brazil and supplied by the partner operator. Input data is used to generate 2D and 3D modeling of the spatial trend of undrained shear resistance and specific weight using the test data. The trend is assessed using Sparse Bayesian Learning and Gaussian Process Regression, which are probabilistic-based methods capable of dealing with data uncertainties. These methods, which represent the latest advancements in the field according to current literature, provide accurate and cost-effective estimates of soil properties. Analysis procedure consisted of assessing the plausibility of the models regarding the CPTu data used as input. The comparative analysis revealed that both SBL and GPR improved soil characterization compared to traditional methods. Both methods show significant capabilities to model spatial trends, and accuracy and performance issues are addressed for 3D modeling with different soil types. Depending on the initial data, both methods can present interesting results with similar computational cost but, in general, SBL takes longer time to be generated. Bayesian evidence is used as metric to compare both models as it is shown that is equivalent to a cross-validation procedure using the log posterior probability function as scoring function. The probabilistic nature of both techniques supports robust evaluation of uncertainties associated with spatial heterogeneity, facilitating reliability-based design and decision-making processes. The study highlights the potential of SBL as a powerful tool for geotechnical site characterization in well design. This innovative approach, validated using real-world data from industry partners, offers valuable insights for research and practical applications in diverse geological settings. The adoption of SBL and GPR in 3D geomechanical modeling represents a significant advancement over existing methods, enhancing understanding of soil behavior and supporting the development of more robust offshore structures.