MW-Net: Multi-Wave U-Net with Cross-Wave Links for Multi-Scale Physical Dynamics

ICLR 2026 Conference Submission21339 Authors

Keywords: Physics simulation, U-Net, Dynamics learning, Surrogate model

Abstract: We propose Multi-Wave Network (MW-Net), a novel deep learning architecture for modeling the temporal evolution of complex, multi-scale physical systems. MW-Net extends the U-Net architecture by stacking multiple encoder–decoder “waves” (U-Net modules). Unlike prior stacked U-Net variants such as SineNet, which restrict skip connections to within each wave, MW-Net introduces skip connections both within and across successive waves at matching spatial resolutions. This design enhances hierarchical representation learning by enabling repeated interactions between feature representations at the same and different spatial scales, supporting progressive refinement of learned dynamics and offering explicit control over network depth through the number of stacked waves. We evaluate MW-Net on diverse physical systems: 2D Kolmogorov fluid turbulence, Hasegawa–Wakatani plasma turbulence, a shallow-water planetary atmosphere model, and buoyant smoke flows (2D and 3D). Across all cases, MW-Net consistently outperforms state-of-the-art baselines and achieves Pareto improvements in the accuracy–computational cost trade-off. While the best-performing baseline varied by task, MW-Net achieved substantially lower errors and up to 3× faster convergence in reaching low-error regimes under fixed learning schedules.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 21339