A Novel Approach for Rock Mechanical Parameter Prediction in Deep Shale Gas Reservoirs Based on VMD-CNN-BiLSTM-AT

The commercial development of deep shale gas reservoirs depends mainly on hydraulic fracturing. Concurrently, the accurate prediction of rock mechanical parameters is critical to informed decision-making and the optimization of hydraulic fracturing parameters. Traditional methods for developing deep shale reservoirs primarily rely on empirical formulas based on compressional and shear-wave velocities. However, acquiring shear wave data is challenging, and the accuracy of such formulas is often low. Conventional machine learning algorithms exhibit limited predictive accuracy and generalization due to the significant heterogeneity of rock-mechanical data. To efficiently predict rock mechanical parameters with high precision in deep shale formations and to improve guidance for optimizing hydraulic fracturing designs in the deep shale reservoirs of Western Chongqing, this study integrates logging data with laboratory rock-mechanical test data. This research moves beyond simplistic model stacking and utilizes a customized architectural design tailored to the three core characteristics of deep shale in the Western Chongqing Block: high pressure, high organic content, and high brittle mineral content. Specifically, the Variational Mode Decomposition (VMD) is utilized to isolate high-frequency logging fluctuations induced by the strong heterogeneity of high brittle mineral content. The Convolutional Neural Network (CNN) acts as a spatial feature extractor to capture the intricate spatial distribution patterns associated with high organic content. Subsequently, the Bidirectional Long Short-Term Memory (BiLSTM) network models the long-range sequential depth dependencies, reflecting the continuous geomechanical evolution under high-pressure compaction gradients. Finally, the Attention (AT) mechanism dynamically prioritizes the most sensitive logging responses to rock mechanical properties under these complex geological constraints. The proposed VMD-CNN-BiLSTM-AT model was then used to estimate Young’s modulus and Poisson’s ratio in the Western Chongqing Block. The testing phase yielded strong predictive performance, achieving R2 scores of 0.966 and 0.96 for Young’s modulus and Poisson’s ratio, respectively, which were approximately 10% higher than those of other conventional models. Therefore, the proposed model supports data-driven fracturing optimization in deep shale plays by precisely predicting mechanical properties.