A two‑stage multi‑grade deep learning method drastically improves neural PDE solvers.
The authors train shallow networks incrementally, freezing earlier grades and adding residual blocks to capture higher‑frequency features. In a second pass they unfreeze selected layers and retrain, using the first stage as a warm start. Theory shows each grade lowers the loss, and experiments on 1D, 2D and 3D viscous Burgers' equations report error reductions up to 60 times compared with conventional single‑grade training.
This matters because traditional deep networks for PDEs struggle with non‑convex optimisation and shock‑like solutions. By breaking the problem into manageable grades, the approach sidesteps deep‑network pitfalls and offers a more interpretable refinement path.
If the trick scales, it could narrow the gap between neural solvers and classic numerical schemes, though the added staging may offset some speed gains.