Double screening in the training dynamics of variational physics-informed neural networks for heterogeneous coupled parabolic systems

What the math gives to ML

The paper identifies a useful failure mode in coupled neural training: when two output or parameter groups have very different differential scales, the joint Gram or Hessian matrix becomes ill-conditioned even though continuous gradient flow itself does not slow down. Its transferable mathematical asset is the Schur-complement view of coupled positive-definite blocks: eliminate the response of one group and optimize the residual curvature seen by the other. This suggests a block preconditioner or blockwise optimizer for multi-output networks and multi-task models whose parameter groups have strongly different gradient spectra. The first implementation should target small output heads or explicitly separated parameter groups, where block curvature can be estimated cheaply and the predicted reduction in condition number is directly measurable.