AI/ML Method for Seismic Well Tie Support on the OSDU Platform: Predicting Missing Wireline and Checkshot Data Using Well Borehole, Mudlog, and Seismic Data

Abstract In this study, we introduce an AI/ML method for predicting missing wireline and checkshot data to support seismic well tie workflows. Well tie seismic is a crucial aspect of seismic interpretation, providing seismologists and geophysicists with a powerful tool to correlate well data with seismic volumes, thereby enhancing the precision of 2D/3D seismic data interpretation. Our approach leverages existing well borehole, mudlog, and seismic data to predict a series of critical datasets, including petrophysical sonic logs and time-depth pair data. The input data for this model comprises conventional wireline logs (gamma ray, resistivity, density, and neutron logs), Vertical Seismic Profile (VSP) datasets, seismic reflectivity, velocity cubes from Pre-Stack Depth Migration (PreSDM) datasets, mudlogs, and deviation surveys. Well log sonic curves (DT), such as Compressional Sonic (DTC) and Shear Sonic (DTS), are essential for well-seismic tie workflows and geomechanical interpretations but are often not available due to additional costs associated with special wireline services. Similarly, checkshot data is costly and typically limited to selected wells. Our model employs AI/ML techniques, including Random Forest and Gradient Boost Regressors, to accurately predict these missing datasets. The dataset used in this model includes 24 potential wells and seismic volumes from the VOLVE dataset. The initial phase involved data gathering, sanitization, and alignment to reference MD values. The dataset was split into training and testing sets to build a generalized data-driven model using Linear Regression, Random Forest, and Gradient Boost algorithms. The trained model synthesizes the required target data, including DT curves, TD pairs, and wavelets, which were tested directly on the target seismic reflectivity cube. The final model was validated by testing various sets of VOLVE wells and seismic datasets both horizontally and vertically. The accuracy of the model was further confirmed using a new well not included in the training dataset. Our model achieved an R² of 0.91 for predicting sonic logs. Cross-validation using leave-one-out methodology yielded a performance score of 0.8 ± 0.2. We then used the predicted sonic logs to estimate TZ Time, achieving a validation score of 0.9 ± 0.05. Feature importance was determined using a Random Forest algorithm, and a light GBM regressor was employed to predict final model scores using the 10 most significant features. The final TZ Time predictions incorporated the predicted sonic logs as input data. Our method demonstrates the capability to fill gaps in missing well data, such as DT family curves and time-depth relationships, without incurring the additional costs of DT curve services and VSP/Checkshots surveys. These datasets can be directly applied in seismic well tie workflows to produce synthetic curves that match the frequency wavelets of seismic data. Experiments were conducted on the OSDU (Open Subsurface Data Universe) platform, showcasing its potential as a unified data framework that facilitates collaboration between seismic and petrophysical experts. This study proposes a comprehensive workflow for AI/ML projects within the OSDU environment, highlighting its significance as a foundational tool for future research and development in geophysical and petrophysical integration. This approach ultimately supports G&G operations by providing more accurate predictions and enhancing the efficiency of seismic interpretation processes.