Learning Physics from an Imperfect Ancestor
A rough neural operator can still steer a PINN into the right physics solution.
The paper proposes a three-stage setup: freeze a physics-informed neural operator’s spatial basis, extrapolate its solution branch, then distill that field into a new PINN. The claim is not that the operator is accurate out of distribution, but that it can point optimization toward the correct basin. In Allen-Cahn tests, this keeps the PINN from settling on the trivial zero solution even though that solution satisfies the residual. In lid-driven cavity flow at Re = 3200, the method reaches the intended physical state with fewer parameters and optimization steps than cited baselines. ArXiv · AI/CL/LG's note
The paper proposes a three-stage setup: freeze a physics-informed neural operator’s spatial basis, extrapolate its solution branch, then distill that field into a new PINN. The claim is not that the operator is accurate out of distribution, but that it can point optimization toward the correct basin. In Allen-Cahn tests, this keeps the PINN from settling on the trivial zero solution even though that solution satisfies the residual. In lid-driven cavity flow at Re = 3200, the method reaches the intended physical state with fewer parameters and optimization steps than cited baselines. ArXiv · AI/CL/LG's note
score 4