Bayesian-Based Approach in Soil Characterization for Top-Hole Design

Abstract

This paper addresses Bayesian-based data-driven site characterization methods for estimating soil parameters used in top-hole casing design. Different models are applied to datasets of piezocone tests (CPTu) conducted in Brazilian fields and their performance is compared to some characterization techniques currently employed by the oil and gas industry. Regression models are crucial for soil characterization for top-hole casing design. Data-driven methods consist of a powerful tool for this purpose, allowing handling uncertainties originating from soil variability. This paper addresses machine learning models devoted to sparse data, namely Geotechnical lasso (Glasso) and Gaussian Process Regression (GPR) from a Bayesian perspective considering prior knowledge of site information. This approach provides statistical information on soil parameters like undrained shear strength, supporting structural analysis of wellhead systems, and conductor/surface casing strings. Datasets were collected from in-situ CPTu tests conducted in the Campos basins, in eastern Brazil. The first case study addressed primary CPTu data, including cone tip resistance, friction sleeve, and total pore pressure, to characterize undrained shear resistance as a random variable. In this context, the parameter was modeled using Phoon´s modified Bartlett test to calculate sample sizes that ensure stationarity. This approach presented relevant results in assessing the probability of failure for conductor and surface casing design based on the operator´s internal design criteria. As these applications, more robust techniques were used to improve data characterization. Glasso and GPR models are used to model parameter tendencies and evaluation of the random data. These methods differ in prior probability density functions adopted - Laplace and Gaussian, respectively - and they are compared with regression techniques widely used in design practice. All the models were trained to estimate undrained shear resistance. Preliminary results confirm that the technique improves the characterization of soil strata and undrained shear strength, with a beneficial effect on the analysis of offshore top-hole structural design cases. Some metrics such as Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) indicate that GPR outperforms other methods. This is an innovative methodology applied in real-case scenarios with data ceded from the partner operator. The formulation evaluates uncertainties associated with the spatial heterogeneity of the material, continuously improving robustness with each new project data. This enables a better understanding of soil behavior in specific oilfields and can assist the decision-making process in well design, improving operational safety. Furthermore, the results of the statistical modeling support reliability-based analysis to deliver probability-based indicators for well integrity design.