Physically Constrained Hybrid Deep Learning for Seismic Inversion to Enhance Reservoir Fluid Discrimination.

This study presents a novel inversion framework that integrates deep learning with ensemble modeling, rigorously constrained by rock physics theory, to enhance the prediction of subsurface properties across varying seismic incidence angles. Utilized Gassmann and Aki-Richards equations to generate label data for the training process. Developed a weighting algorithm that assigns optimal weights to each base model based on its accuracy. Our hybrid deep learning–ensemble method outperforms conventional inversion and standard deep learning across seismic incidence angles, delivering more accurate and efficient reservoir characterization and enabling stronger predictions of fluid content and production behavior.