Identifying changing partial differential equations using Sampled Local WeakIdent

What the math gives to ML

The paper's transferable contribution is a localized weak-form identification procedure that combines cheap patchwise equations with global sampling and consensus. Instead of differentiating noisy data pointwise, it projects temporal derivatives and candidate operators against compactly supported test functions, producing stable linear measurements. For neural PDE models, this suggests a robust training and architecture-selection mechanism: sample spatial patches, compute weak residuals for candidate operator supports, and use frequency-consistent supports to infer spatial regions or route inputs to different local experts. The strongest near-term target is a PINN or neural operator for heterogeneous systems where the governing operator changes across space.