Topology-Preserving Neural Operator Learning via Hodge Decomposition
International Conference on Machine Learning (ICML) 2026 · Regular Paper [Avg. Score: 5/6 (est. Top 1% via PaperCopilot)]
Introducing a hybrid Eulerian–Lagrangian architecture, Hodge Spectral Duality (HSD), that exploits Hodge orthogonality to decouple unlearnable topological features from learnable geometric dynamics. Combining discrete differential forms with an auxiliary ambient space, the method markedly improves accuracy, efficiency, and fidelity to physical invariants when solving field equations on geometric meshes.