Stable Long-Horizon PDE Forecasting via Latent Structured Spectral Propagators
Stable Long-Horizon PDE Forecasting via Latent Structured Spectral Propagators
Authors: Xiaoxiao Lu, Ye Yuan, Jiahao Shi Date: 2026-05-11 Paper ID: openalex:2605.10154
Summary
This paper addresses error accumulation and dynamic drift in autoregressive PDE forecasting by proposing the Structured Spectral Propagator (SSP). The framework utilizes an analysis-propagation-synthesis design that maps inputs to a time-consistent latent space, isolates recurrent dynamics from spatial details, and evolves spectral modes using a frequency-conditioned linear backbone. This explicit architectural bias significantly improves stability and reduces relative L2 errors for long-horizon temporal extrapolation.
Key Contributions
- Proposes a Structured Spectral Propagator (SSP) that decouples latent dynamics from spatial reconstruction to mitigate error accumulation in PDE forecasting.
- Introduces an analysis-propagation-synthesis framework that uses a frequency-conditioned linear backbone and nonlinear spectral closure to maintain stability.
- Demonstrates up to 48.9% reduction in relative L2 error compared to existing state-of-the-art neural operator surrogates on long-horizon PDE benchmarks.
Open Questions & Future Work
Key Concepts
- structured-spectral-propagator: A neural forecasting framework that models PDE dynamics by evolving spectral modes within a propagation-oriented latent space.
Archivist Review
The Structured Spectral Propagator was approved as it represents a robust architectural paradigm for PDE surrogates by separating latent state representation from spectral evolution. The open question regarding long-horizon stability in PDE forecasting addresses a fundamental bottleneck in the field of neural operators. No datasets were approved as none were specifically named or highlighted as a unique benchmarking contribution.
Approved Concepts
- Structured Spectral Propagator: It is the core architectural innovation that enables stable, long-horizon PDE forecasting by decoupling latent dynamics from reconstruction.
Approved Open Questions
- Stability in Long-Horizon Forecasting: Long-horizon forecasting is essential for scientific utility, and the current reliance on finite-time supervision makes generalization to long-time dynamics a significant bottleneck in deploying neural surrogates for real-world physical systems.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.10154
- paper_id
- 2605.10154
- paper_source
- openalex
- domain
- nlp
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- structured-spectral-propagator
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-14T06:10:00Z