Predicting mud weight in carbonate formations using seismic data: A data-driven approach

Abstract

Accurate mud weight prediction is crucial for safe and efficient drilling operations in overpressured formations. Especially. Traditional approaches heavily rely on manual seismic interpretation, which is limited by data noise and high dimensionality, and the need for well log data. This study aims to address these challenges by integrating advanced machine learning techniques with seismic and mud weight data, focusing on developing an automated, data-driven workflow that can reliably predict appropriate mud weight trends across large-scale geological fields. The methodology employs a neural network (NN) autoencoder for seismic dimensionality reduction, retaining critical geological features in latent layers. Optimized input features are selected using statistical and interpretability-driven techniques. A k-nearest neighbors model, tuned through grid search, serves as the predictive core, with kriging applied to refine spatial predictions and reduce errors. The approach demonstrates improved geological interpretability and accuracy compared to conventional methods. Applied to a large carbonate field, the workflow effectively predicts spatial mud weight variations, highlighting its scalability and reliability. By combining autoencoding, feature selection, and geostatistical refinement, this methodology offers a robust and interpretable framework for tackling complex geological challenges in drilling operations.