A Two-Stage Residual-Correction Method for Transferring Discontinuous Physical Quantities in Multiphysics Coupling

The transfer of physical quantities between non-matching meshes in multiphysics coupling simulations is challenged by the coexistence of smooth and discontinuous features, for which existing interpolation methods often struggle to achieve both high global accuracy and shape preservation across discontinuities. This paper proposes a two-stage hybrid interpolation method. In the first stage, a baseline neural network (MLP), equipped with an unsupervised discontinuity-aware soft indicator, Fourier positional encoding, and adaptive Huber-loss clipping, is used to extract the global low-frequency smooth component. In the second stage, linear-kernel RBF interpolation is employed to locally and accurately correct the residual error left by the first stage. Numerical experiments on a two-dimensional composite function, the ONERA M6 wing, and a trapezoidal-wing temperature field show that the proposed method reduces global interpolation error relative to the tested baselines and, in the tested cases, yields sharper transitions with smaller visible overshoots near strong discontinuities such as shocks. The method accurately reproduces the source data and, in the test cases, improves the accuracy of field transfer between non-matching meshes.