DSO: Dual-Scale Neural Operators for Stable Long-term Fluid Dynamics Forecasting
New neural operator solves long-term stability issues in fluid dynamics forecasting, outperforming existing models.
A research team has introduced DSO (Dual-Scale Neural Operator), a breakthrough architecture designed to solve the persistent challenge of long-term stability in AI-driven fluid dynamics forecasting. Traditional neural operators, which model systems governed by partial differential equations (PDEs), often fail over extended simulations due to two critical issues: local detail blurring (losing fine vortex structures) and global trend deviation (drifting from true motion paths). The researchers identified that these failures stem from treating local and global physical information uniformly, despite their fundamentally different evolution characteristics.
DSO's novel solution is an explicit architectural decoupling. It processes fine-grained local features through depthwise separable convolutions, while handling long-range global aggregation with an MLP-Mixer module. This design is grounded in physical observation: experiments on vortex dynamics showed that nearby perturbations affect local structure, while distant ones influence global trends. In extensive testing on turbulent flow benchmarks, DSO demonstrated state-of-the-art accuracy and robust stability, reducing prediction error by over 88% compared to existing neural operator models. This represents a significant leap toward reliable, long-horizon simulation of complex physical systems.
- DSO architecture decouples processing into local (depthwise convolutions) and global (MLP-Mixer) modules based on physical principles
- Solves two key failure modes: local detail blurring and global trend deviation in extended simulations
- Achieves 88% lower prediction error than previous neural operators on turbulent flow benchmarks
Why It Matters
Enables more accurate, stable simulations for weather prediction, aerospace design, and industrial fluid systems.