A static–dynamic feature coupling framework for spatiotemporal multiphase flow prediction in heterogeneous reservoirs

Water flooding as an enhanced oil recovery technology ensures the stable oil production capacity and enhances the recovery rate of oil in reservoir engineering. It allows for better water flooding management to accurately and quickly predict the spatiotemporal evolution of the water–oil displacement. However, variations in geological permeability distributions significantly influence the flow dynamics and substantially increase the complexity of reservoir simulation. Deep learning methods offer efficient solutions for flow field prediction, but existing approaches suffer from accuracy degradation over long-term prediction and exhibit limited generalization capabilities. In this study, we propose the Spatio-Temporal Gated Fusion Network (STGFNet). The framework takes static geological permeability information and dynamic multiphase flow features as inputs to predict the spatiotemporal evolution of water–oil displacement. Specifically, STGFNet introduces a spatial gated residual module and a static–dynamic fusion convolutional long short-term memory network to address accuracy degradation and enhance the generalization capability. Experiments demonstrate that STGFNet can accelerate simulation and achieve an order of magnitude faster than traditional numerical solvers. Furthermore, it achieves accurate and generalizable predictions across diverse geological scenarios, achieving state-of-the-art performance compared to existing intelligent models. The proposed methodology could be a useful reference source for real-time modeling and operational decision support in complex subsurface environments.