Flux-form spatiotemporal neural operators for coarse-grained dynamics of multiscale PDEs
arXiv:2608.18148
2026
Dynamics
2 ideas extracted · analyzed Sep 1, 2026
What the math gives to ML
The paper offers a transferable flux-form neural operator for coarse-grained multiscale dynamics, combining causal temporal memory with spatial Fourier mixing and an explicit divergence-form update. The key mechanism is structural: predicting a flux and applying a discrete divergence enforces local conservation and prevents arbitrary pointwise corrections from accumulating during autoregressive rollout. A second useful mechanism is selecting the history-window length from the decorrelation time of the unresolved closure injection rather than tuning memory heuristically. These mechanisms are especially promising for neural PDE solvers, state-space models, and long-horizon sequence predictors where conservation, finite memory, and rollout stability matter.
Ideas from this paper
✓✓ Beats tuned baseline
2026
Replace pointwise prediction of the next field with prediction of a learned flux followed by a discrete divergence. Combine Fourier spatial mixing with a causal temporal kernel over the recent resolved-history slab, so the model learns finite-memory closure effects while preserving local conservation exactly under periodic or compatible boundary conditions. The architecture should reduce spurious mass drift and improve autoregressive rollout stability on coarse-grained PDE data.
Useful8/10
Difficulty5/10
Novelty6/10
✗ Mechanism failed
2026
Choose the neural operator's input-history length from the measured correlation time of the unresolved closure signal produced by coarse-graining. This avoids under-memory, which causes systematic closure error, and over-memory, which increases attention cost and can destabilize training. The same diagnostic can drive adaptive memory truncation across physical regimes.
Useful7/10
Difficulty4/10
Novelty7/10