Adaptive Learning Method for Predicting Geomechanical Parameters from Heterogeneous Formation Images Based on Deep Learning

Abstract This study introduces an innovative approach utilizing the ResNet50 deep learning model to enhance the accuracy of geomechanical parameter characterization in heterogeneous formations. The method employs a dual-module architecture, combining "feature extraction" and "regression mapping," to improve the precision of key parameters such as Poisson's ratio, elastic modulus, and porosity. By leveraging ResNet50 for high-dimensional feature extraction from logging images and integrating a customized regression network, the method achieves a resolution improvement from 12.5 cm to 7 cm, with prediction accuracy exceeding 83%. This advancement overcomes the limitations of traditional methods in complex reservoir characterization. The results highlight the model's ability to capture detailed formation information, providing critical insights for geological interpretation and reservoir evaluation. By addressing cross-scale challenges and refining formation information technology, this approach significantly enhances the efficiency of shale reservoir development, demonstrating the potential of deep learning in advancing geomechanical parameter characterization.