Spatiotemporal attention network for pore-scale capillary–viscous imbibition in complex porous media

Pore-scale multiphase displacement in heterogeneous fractured porous media is governed by coupled capillary, viscous, interfacial, and structural effects, leading to complex interfacial evolution and displacement behaviors, which require fine discretization and substantial computational cost in conventional simulations. In this work, an end-to-end spatiotemporal attention network (STA-Net) is proposed for time-dependent prediction of pore-scale capillary–viscous imbibition. STA-Net integrates a convolutional encoder–decoder, feature-wise linear modulation-based physics embedding, and a tailored spatiotemporal attention module. The attention mechanism is decoupled into spatial and temporal branches. Spatial attention captures nonlocal interactions induced by pore connectivity and flow redistribution, whereas causally masked temporal attention explicitly models history-dependent evolution by ensuring each state attends only to current and preceding flow states. This architecture enables direct prediction of temporally ordered saturation, velocity magnitude, and pressure fields. Validations against reference simulations under various porous structures, physical conditions, and time instances demonstrate that STA-Net accurately captures flow fields, displacement evolution, and transient interfacial dynamics. Ablation studies further show that STA-Net outperforms the convolutional baseline by ∼10% in terms of average R2, owing to the complementary contributions of spatial-context integration and temporal-state correlation. Leveraging this high-fidelity modeling, STA-Net enables rapid and dense exploration of the parameter space and physics-guided interpretation of displacement regimes, identifying how capillary and viscous effects govern regime transitions. These findings indicate that STA-Net provides an effective method for predicting and interpreting pore-scale multiphase flow evolution in complex porous media.