Neural operators for PDEs often learn a single map, but real physics combines multiple processes. Mixture-of-experts operators typically use hard top-K routing, which forces a discrete choice between concurrent mechanisms. COSMOS proposes soft continuous gating where all specialists remain active and their outputs are mixed by a learned gate, better matching the blended nature of real physics.
The authors introduce a compositional benchmark with known mixture weights to evaluate routing. On this benchmark, dense soft routing achieves 2.2 times lower rollout error than hard top-1 routing with the same feature fusion. Using the true mixture weights at inference further reduces error to 0.034, revealing limitations of the learned gate. However, an initial advantage over a tuned FNO on diffusion-reaction was overturned by a larger seed audit, showing that small-seed comparisons can be unreliable. The learned routing does not align with the intended physical mechanisms, so the model claims accuracy rather than interpretability.