New Physics-Informed Schrödinger Bridge method enables efficient data assimilation from sparse observations without requiring full high-fidelity training data

PICSB (Physics-Informed Conditional Schrödinger Bridge) reconstructs complete spatiotemporal fields from sparse measurements while respecting PDE constraints

Eliminates need for expensive per-instance test-time optimization by using amortized reconstruction with generative models

Leverages low-fidelity simulations as informative priors in multi-fidelity settings, reducing dependence on scarce high-fidelity supervision during training

Addresses practical real-world limitations where full-field high-fidelity training data is unavailable or prohibitively expensive to obtain