Learning Pore-scale Multiphase Flow from 4D Velocimetry

Multiphase flow in porous media underpins subsurface energy technologies, including geological $CO_{2}$ and underground hydrogen storage, yet pore-scale dynamics in realistic three-dimensional materials remain difficult to predict. Here we introduce a multimodal machine learning framework that learns multiphase pore-scale flow directly from time-resolved four-dimensional micro-velocimetry measurements. The model couples a graph network simulator for Lagrangian tracer-particle dynamics with a 3D U-Net for voxelized fluid--fluid interface evolution, using the imaged pore geometry as a shared boundary constraint. Trained autoregressively on capillary-dominated drainage sequences ($Ca \approx 10^{-6}$), the framework captures transient, nonlocal flow perturbations and abrupt interface rearrangements (Haines jumps) over rollouts spanning seconds of physical time, compressing hour-to-day-scale direct numerical simulations to seconds of inference. The framework generalizes to distinct experimental boundary conditions and, zero-shot, to a structurally different rock type, demonstrating that experimental 4D velocimetry can underpin physically grounded surrogate models for rapid digital experiments in subsurface energy applications.