Derivative-Informed Training of Neural Operators On-the-Fly via Sketched Tangent Consistency
My rating
The work targets neural operators used as differentiable surrogates for partial differential equation inverse problems, constrained optimization, design, and control. It directly supervises input-output sensitivities through derivative-informed training and proposes generating sketched tangent consistency information on-the-fly. This avoids dependence on offline-generated derivative labels that can be slow, storage-intensive, and difficult to adapt across datasets, resolutions, or perturbation bases.
Tags: