Picking the Branch Is Not Solving the Problem: What an Imperfect Neural Operator Buys a Physics-Informed Network, Where Its Basin Information Comes From, and the Comparators That Decide Whether the Hybrid Pays

2026-10-28T13:16:16Z https://lifeofresearch.com/oai2

oai:lifeofresearch.com:151 2026-10-28T09:48:01Z type:preprint openaire

10.5281/zenodo.23669446 https://lifeofresearch.com/preprints/151 Pranay Mahendrakar 0009-0003-7224-029X SONYTECH Picking the Branch Is Not Solving the Problem: What an Imperfect Neural Operator Buys a Physics-Informed Network, Where Its Basin Information Comes From, and the Comparators That Decide Whether the Hybrid Pays Life of Research 2026 physics-informed neural networks neural operators basin of attraction numerical continuation solution multiplicity warm start scientific machine learning baselines Preprint Creative Commons Attribution 4.0 (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research Physics-informed neural networks (PINNs) can minimise a PDE residual and still land on the wrong solution, while neural operators are cheap to evaluate but lose accuracy outside their training range. A 2026 study joins the two: a crude, extrapolated operator is distilled into a fresh PINN, and the authors conclude that a prior need not be accurate, only able to identify the correct basin from which residual minimisation recovers the solution. In this analysis the author reviews and synthesises published results to show that the principle is sound but old, and that the evidence offered for it leaves open the three questions that decide whether the hybrid is worth building: where the basin information came from, whether the comparison baseline was the one practitioners would use, and whether a PINN is the right refiner once a basin is known. For the bifurcating Allen-Cahn case, the author derives a one-mode estimate of the branch amplitude that agrees with the study's reference peak to within about 3 percent. On that estimate, the seed function hand-placed in the operator's ansatz is closer to the target than the extrapolated teacher. For the lid-driven cavity at Reynolds number 3200, each of the published arms collected here that reaches good accuracy walks a path in Reynolds number, and a curriculum-trained PINN with no operator reports a lower error than the hybrid. Eighteen published results are consolidated in a table, a five-arm comparator procedure is proposed for licensing a basin-selection claim, and the opposite failure is recorded as well: an accurate operator can still start Newton's method outside its basin.