GeoIncNO: Geometry-Aware Neural Operator for Stable Long-Horizon PDE Prediction
A recent preprint on arXiv (2608.11237) presents GeoIncNO, an innovative incremental neural operator that incorporates geometric awareness to enhance the stability of long-term predictions for partial differential equations (PDEs). This technique tackles the issue of local error accumulation in autoregressive models, which can lead to problems like spectral inconsistency, phase misalignment, or mean drift. Unlike traditional methods that enhance state representations or operator frameworks, GeoIncNO organizes the latent transition increment by forecasting latent increments for residual improvement and employing lightweight low-rank projectors to manage channel coupling based on the increment's spectral energy distribution. Furthermore, it features a mechanism aimed at minimizing reconstruction errors in physical space. The paper is introduced as a new submission on arXiv, with the abstract outlining the challenges and the proposed approach.
Key facts
- GeoIncNO is a geometry-aware incremental neural operator for PDE prediction.
- It targets long-horizon autoregressive prediction issues like spectral inconsistency, phase misalignment, and mean drift.
- The method predicts latent increments for residual advancement.
- It uses lightweight low-rank projectors to regulate channel coupling in active frequency bands.
- The frequency bands are derived from the increment spectral energy distribution.
- It also introduces a mechanism to reduce physical-space reconstruction errors.
- The paper is available on arXiv with ID 2608.11237.
- The announcement type is 'new'.
Entities
Institutions
- arXiv