IGA-KAN: Isogeometric Analysis with Physics-Informed Closed-Form Kolmogorov-Arnold Networks for Forward and Inverse PDEs
Naraghi, Parand, Sadr and Rahmati · arXiv 2610.06348
The outcome is the verdict on the one claim below, under the protocol we fixed in advance. It is not a verdict on the paper as a whole.
Claim tested
Released code reproduces the paper's tabulated L2/H1 gains over baselines on nine rows.
What we ran and what came back
The authors' released notebooks, run twice at their default settings, reproduced the accuracy gains of the post-processing step on all nine tabulated problems, in both error norms. Gains ranged from 4.16 to 89.9 (L2) and 4.14 to 219.7 (H1) against our bounds of 4.15 and 4.05, and the two runs agreed to within 0.5%. A held-out geometry not in the paper (an annulus sector) also passed, with gains of 17.7 and 84.3. Read against the paper's quoted lower end of 4.2 instead of our bound, one row falls short: the flower geometry at 4.16. We set the bound at 4.15 after reading the authors' stored outputs, and say so in the protocol; the difference is at rounding level. The method is deterministic, both runs were on the same type of machine, the neural-network comparisons in the notebooks were not run, and the errors were computed with the library's own routine.
What this result supports
The released code is a reliable base for further work on the nine tabulated rows.
What it does not support
Generalisation beyond those rows; arm B (other settings) is separate.
Record
Protocol, frozen by git commit before the run and deposited afterwards (the archive was unreachable at the time): 10.5281/zenodo.23332639
The authors were sent this result. They have not reviewed this listing.
How the protocol works is on the method page. If you are an author of this paper and something here is wrong, write to customers@aegismind.app and we will correct it on this page.