Ambient Physics: Training Neural PDE Solvers with Partial Observations
In many scientific contexts, obtaining complete observations of PDE coefficients and solutions proves expensive, hazardous, or unfeasible. Recent diffusion-based approaches can reconstruct fields from incomplete data, yet require fully observed examples during training. We present Ambient Physics, a framework enabling …