Real Time Wellbore Stability Prediction for the Natural Fractured Reservoir with Physical-Based Machine Learning Techniques

Abstract

Wellbore stability remains a persistent and costly drilling challenge, especially in naturally fractured reservoirs, where complex interactions among in-situ stress, geomechanical properties, drilling fluids, and operational parameters govern failure. Traditional analytical and numerical models are time-consuming and mainly applicable to post-drilling analysis, thus ill-suited for real-time decision support. To overcome this, we propose a physics-informed machine learning framework that integrates insights from a chemo-hydro-mechanical (CHM) model—specifically designed for fractured formations—with high-resolution logging-while-drilling (LWD) data. A hybrid LSTM-DNN network is developed: the LSTM branch extracts temporal evolution patterns from historical drilling sequences, while the DNN branch captures instantaneous features from current operating conditions. Their fused representations enable simultaneous real-time prediction of pore pressure and collapse pressure. Validated on deep drilling data from the Sichuan Basin and compared against standalone LSTM and DNN models, the fusion model demonstrates faster convergence and superior accuracy on both training and validation sets. In a 150-step dynamic real-time test, predictions track field measurements closely, achieving relative errors of 4.53% and 5.68% for pore and collapse pressures, respectively. This work provides a robust, physically grounded deep-learning solution that facilitates timely operational adjustments, thereby enhancing safety and efficiency in complex fractured reservoirs.