Improving numerical fluid flow modeling by resolving under-resolution pores from X-ray images of porous media: a hybrid deep reinforcement learning framework

Numerical modeling based on the 3D X-ray microtomography data is a cornerstone for comprehending fluid transport dynamics in porous materials. However, inaccuracies in image processing and pore morphology delineation introduce significant errors into the simulation outcomes. In this study, we introduce a deep reinforcement learning (DRL)-based framework designed to refine pore morphology by accurately unveiling under-resolution, micro-links, and complex pore boundaries. The framework uses a smart agent and its action space to iteratively adjust the pore space morphology via targeted morphological operations. This process is dynamically guided by a reward function that minimizes deviations between simulation results and corresponding experimental measurements. The framework operates using three distinct simulation modalities: Pore Network Modeling (PNM), a Convolutional Neural Network (CNN), or a hybrid algorithm. The CNN, trained on 1,176 3D images sourced from 21 unique rock samples, achieves a dramatic reduction in PNM computational time from 134 to 0.44 seconds per step. Crucially, the hybrid algorithm intelligently leverages PNM and CNN interchangeably to optimize the trade-off between computational expense and predictive accuracy. Evaluation on a carbonate sample demonstrated unprecedented improvements: permeability error was reduced from 32.16% (base PNM) to 0.79%, and effective porosity error from 40% to 0.32%. Furthermore, for a tight sandstone sample, the framework established fluid flow paths, increasing computed permeability from zero to 0.52 millidarcy (experimental value: 0.104 millidarcy). These results demonstrate this framework's ability to systematically refine pore morphology and revolutionize simulation accuracy, offering transformative potential for applications in hydrology, CO 2 sequestration, and subsurface energy storage.