Enhancing Precision in Data-Driven Mud Weight Mapping
K. A. Khemraev, Georgy Peshkov, N. V. Zinovich, N. Bukhanov · 2025
Abstract Overpressure zones, where subsurface fluid pressure exceeds normal hydrostatic levels due to geological processes, pose challenges in oil and gas drilling, impacting well stability, construction, economics, and safety. To mitigate well kick, engineers increase mud weight, indicating potential overpressure. However, detecting these zones is difficult due to the lack of well logs in intersecting areas. Therefore, knowing the correct mud weight density is crucial for drilling safely and efficiently. The preprocessing of seismic data begins with the extraction of seismic traces within a specified window for the target geological formation. These amplitude traces are used to train an autoencoder, which compresses the temporal interval along the target horizon, reducing its vertical dimensionality. We then employ machine learning techniques to identify the low-frequency trend between seismic data and preprocessed mud weight. This step generates a smoothed, geologically interpretable mud weight map. To refine our predictions, we use a geostatistical approach. This involves applying a Gaussian process regression model to enhance the prediction accuracy around already drilled wells. After preprocessing of well data with mud weight values, only 70% of these wells were used for building the low-frequency mud weight map. We then compressed (autoencoded) selected seismic interval in the vertical axis from 300 samples to an embedding of 16 layers. Well coordinates, seismic embedding and target two-way travel time horizon were used as input data for regression KNN model. The generated smooth mud weight map has mean average error around six pound cubic foot (PCF). Finally, incorporating all mud weight data from the wells in the analyzed formation and using kriging modeling, we improved the accuracy of the mud weight map prediction. Also, it was determined that the largest overpressure areas are associated with the structural highs. Our data-driven approach effectively predicts mud weight in overpressured formations by combining seismic data autoencoding and geostatistical methods. The resulted mud weight map minimizes the risks of kicks, demonstrating the method's practical reliability in oil and gas drilling operations.