Localized Operator Learning with Adaptive Partition-of-Unity Mixture-of-Expert Networks
HiRefPOU uses learned spatial partitions to make neural operators handle localized PDE structure more accurately.
The paper introduces a partition-of-unity mixture-of-experts setup where geometry-aware gates blend local expert networks. Its main model, HiRefPOU, adds hierarchical parent-child partitions to DeepONets while keeping global continuity. The authors also apply the same idea to Fourier Neural Operators without changing the spectral layers. On heterogeneous Darcy and reaction-diffusion benchmarks, HiRefPOU beats global DeepONet and static POU-MoE baselines, with learned partitions lining up with interfaces and sharp solution changes. ArXiv · AI/CL/LG's note
The paper introduces a partition-of-unity mixture-of-experts setup where geometry-aware gates blend local expert networks. Its main model, HiRefPOU, adds hierarchical parent-child partitions to DeepONets while keeping global continuity. The authors also apply the same idea to Fourier Neural Operators without changing the spectral layers. On heterogeneous Darcy and reaction-diffusion benchmarks, HiRefPOU beats global DeepONet and static POU-MoE baselines, with learned partitions lining up with interfaces and sharp solution changes. ArXiv · AI/CL/LG's note
score 4