Perturbative-NeuSA: A Structured Spectral Framework for Time-Dependent PDEs
arXiv:2607.24345
2026
Dynamics
1 ideas extracted · analyzed Aug 31, 2026
What the math gives to ML
The paper offers a hybrid-dynamics principle: represent the state as a cheaply computed background plus a perturbation, apply known dynamics and background defects explicitly, and train a neural component only on the unresolved remainder. The transferable asset is the algebraic separation between structured dynamics, background error, linearized correction, and nonlinear closure. This can reduce optimization burden and prevent a model from relearning dominant dynamics that are already available analytically. The most practical transfer is a residual neural simulator or neural operator with an explicit closure gate that suppresses learning when the background already explains the target trajectory.
Ideas from this paper
✓✓ Beats tuned baseline
2026
Replace full-state prediction in a neural simulator or neural operator with prediction of a perturbation around a cheap structured background trajectory. Compute the background defect and known linearized or nonlinear corrections explicitly, and let the neural closure model only the remaining residual. Add a residual-magnitude gate so the learned closure is suppressed when the structured solver already explains the target dynamics.
Useful8/10
Difficulty5/10
Novelty6/10