Geometry-aware Incremental Neural Operator for Long-Horizon PDE Prediction

This paper addresses the challenge of long-horizon autoregressive prediction for partial differential equations using neural operators. The approach introduces a geometry-aware incremental mechanism that structures the latent transition increment, effectively suppressing error accumulation from spectral inconsistency, phase misalignment, and mean drift. By improving the temporal consistency of the learned operator, the method enhances stability and accuracy for multi-step PDE predictions.