Abstract
Accurately forecasting the long-term evolution of turbulence represents agrand challenge in scientific computing and is crucial for applications rangingfrom climate modeling to aerospace engineering. Existing deep learning methods,particularly neural operators, often fail in long-term autoregressivepredictions, suffering from catastrophic error accumulation and a loss ofphysical fidelity. This failure stems from their inability to simultaneouslycapture the distinct mathematical structures that govern turbulent dynamics:local, dissipative effects and global, non-local interactions. In this paper,we propose the{\textbf{\underline{D}}}ifferential-{\textbf{\underline{I}}}ntegral{\textbf{\underline{N}}}eural {\textbf{\underline{O}}}perator (\method{}), anovel framework designed from a first-principles approach of operatordecomposition. \method{} explicitly models the turbulent evolution throughparallel branches that learn distinct physical operators: a local differentialoperator, realized by a constrained convolutional network that provablyconverges to a derivative, and a global integral operator, captured by aTransformer architecture that learns a data-driven global kernel. Thisphysics-based decomposition endows \method{} with exceptional stability androbustness. Through extensive experiments on the challenging 2D Kolmogorov flowbenchmark, we demonstrate that \method{} significantly outperformsstate-of-the-art models in long-term forecasting. It successfully suppresseserror accumulation over hundreds of timesteps, maintains high fidelity in boththe vorticity fields and energy spectra, and establishes a new benchmark forphysically consistent, long-range turbulence forecast.