Enhancing Rock Strength Prediction and Features Selection by Coupling Well Log Data and Deep Learning Approaches in Reservoir Geomechanics

ABSTRACT Accurate rock strength parameters play a vital role in petroleum and mining operations for sustainable drilling activities, and wellbore stability analysis. This study investigates the applicability of deep learning techniques for assessing data‐driven models, and to conduct parametric sensitivity examination for feature attributes ranking to predict compressive strength (UCS) of clastic sedimentary rocks. The predictor variables of well logs data such as formation density, gamma‐ray (GR), compressional acoustic wave velocity (V p ), and share wave velocity (V s ) are utilized. The convolutional neural network (CNN), multi‐layer perception‐based neural network (MLPNN), and transformer‐based predictive models’ outcomes are assessed for checking the models’ reliability, robustness and feature selections using statistical performance indices. The Taylor diagram also employed to examine the importance of the variations in model outcomes. The variables’ dimensionality reduction and features importance are measured applying Shapley additive explanation (SHAP), and three filter approaches. Based on the findings, the transformer model outperformed the MLPNN and CNN models, with high accuracy (correlation coefficient of 0.99) and root mean square error of 1.3 MPa. For instance, GR and V p are the most significant predictor variables to obtain UCS for the studied field, compared and verified by SHAP and filter methods of mutual information, Fisher score, and Chi‐Square test. In the light of these findings, it is revealed that these novel deep learning approaches provide valuable insights into model development and feature attributes selection to estimate rock strength parameters and optimize drilling parameters for wellbore stability analysis to reduce operational risk in petroleum and minerals exploration.