SRPO: Deriving Neural Operators from Symmetry-Reduced Polynomial Interactions
Abstract Neural operators typically learn mappings from predefined computational primitives. We instead derive the architecture from local PDE evolution, representing each update as a multivariate function and reducing redundancy before learning coefficients. This yields the Symmetry-Reduced Polynomial Operator (SRPO), combining relative polar coordinates, angular harmonics, low-rank radial/channel couplings, and multiscale first-, second-, and third-order interactions. Across five two-dimensional PDE benchmarks spanning reaction–diffusion, Navier–Stokes, shallow-water dynamics, and wave propagation in heterogeneous media, SRPO-P3 achieves the lowest held-out error on four tasks and ranks second on the fifth with 1.70 million parameters—about 2.5% of the full-spectrum Fourier Neural Operator (FNO) baseline. It reduces error by 38.0%–47.7% against the strongest non-SRPO baselines on Navier–Stokes and both wave benchmarks. Autoregressive rollouts and ablations indicate complementary contributions from multiscale information, angular harmonics, and higher-order interactions. These results show that PDE-informed structural reduction can yield compact operators with strong predictive accuracy.
Authors
- Xian He
- Lufei Li
- Fuzhen Pang
- Haichao Li
- Yuan Du
Institutions
- Harbin Engineering University (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-16
- DOI
- https://doi.org/10.21203/rs.3.rs-11488718/v1
- Primary Topic
- Model Reduction and Neural Networks
- Type
- preprint