Machine Learning-Based Multiscale Geomechanical Modeling

Traditional three-dimensional geomechanical modeling is hindered by several limitations, including time-consuming workflows, challenges in effectively updating models, constraints in available computational resources, and issues related to software compatibility. To address these challenges, this study introduces a machine learning-based multiscale (well-log and seismic scales) geomechanical modeling approach. It leverages the complementary advantages of well-log data (providing vertical resolution at borehole locations) and seismic data (offering lateral continuity), where machine learning algorithms utilize the spatial constraints from seismic data to guide interpolation and extrapolation of well-log information in 3D space. This methodology integrates geological structural data, well-log information, and seismic data, leveraging machine learning techniques to construct intelligent digital models that can represent subsurface geomechanical behavior. To evaluate prediction performance at both well-log and seismic scales, we conducted extensive model training and completed the method selection. Results indicate that the multitask learning approach achieved R 2 scores of 0.9826 at the well-log scale and 0.9379 at the seismic scale, with mean absolute percentage errors in blind-zone prediction falling within acceptable engineering standards. In the application test for predicting geomechanical parameters of planned wells, the average prediction error for key mechanical parameters is 10.17%, validating the method’s applicability to field development planning, reserve estimation, and drilling decision-making. Additionally, this study implemented interactive visualization using the VisPy library with support for synchronized display of multiscale data and real-time zooming functionality, providing extensible technical support for intelligent oilfield exploration and decision-making. In summary, this research can be applied to scenarios including geomechanical parameter prediction for planned wells, well-log parameter prediction, and seismically driven geomechanical modeling.