COSMOS Uses Soft Mechanism Mixtures for PDE Forecasting

COSMOS implements a soft-routing neural operator that continuously blends four process-biased specialists to model compositional physics. The architecture uses a shared coarse latent grid and a zero-initialized residual to maintain high-resolution detail during long-horizon forecasting.