HyPINO: Multi-Physics Neural Operators via HyperPINNs and the Method of Manufactured Solutions

Read the paper Explore the code Download model weights

HyPINO is a multi-physics neural operator framework that generates Physics-Informed Neural Networks (PINNs) with a Swin Transformer–based hypernetwork and is trained end-to-end via the Method of Manufactured Solutions (MMS). Its main features are:

  • Zero-shot generalization across parametric 2D PDE families (elliptic, parabolic, hyperbolic)
  • Mixed boundary condition support with Dirichlet, Neumann, and interior boundaries
  • Swin Transformer hypernetwork that generates task-specific PINNs
  • Residual-based iterative refinement for test-time accuracy improvement
  • PINN initialization for faster convergence and stronger fine-tuning performance

This in-browser demo lets you run the HyPINO model without installing any software. Configure linear 2D second-order PDEs—elliptic, parabolic, or hyperbolic, on complex domains, then generate predictions directly in the tab, and export the resulting weights plus ready-to-use PyTorch helpers for downstream adaptation.

PDE inputs

Select a preset or insert custom coefficients and boundary information.

This is a list of presets. Clicking on one of them will autofill coefficients and grid inputs.

Input preview

Check masks, boundary values, and forcing before refining.

Select a preset or upload files, then use the Load data button above to inspect the inputs.

Run inference

Choose number of refinement iterations and launch prediction. Set to 0 for no refinement.

Idle

Cite HyPINO

Copy the BibTeX entry for the demo paper.