Efficient Pareto-Front Generation for Electric Machines using IGA and Second Order Derivatives

What the math gives to ML

The paper's transferable asset is a predictor-corrector continuation method for tracing constrained Pareto fronts, combined with second-order Newton corrections and implicit derivative information. Instead of independently training a neural model for many scalarizations, one can warm-start a sequence of models as an objective threshold or loss weight changes, predicting the next solution from the local tangent and correcting it with a few Newton-like steps. The implicit-state formulation also suggests replacing expensive full Hessians with Hessian-vector products and Krylov solves. This is most promising for multi-objective training, constrained fine-tuning, and model compression, where many high-quality trade-off solutions are otherwise obtained by repeated retraining.