Composing interpretable numerical and learned modules with learned policies outperforms monolithic neural operators on PDEs, generalizes better to out-of-distribution cases, and lets you swap components (like boundary conditions) without retraining.
HyCOP learns to solve PDEs by composing simple, interpretable modules (like advection and diffusion) rather than training a single neural network. It learns a policy that decides which module to apply and for how long based on the current state, enabling better generalization to new scenarios and easier transfer to different problems.