Szczegóły publikacji
Opis bibliograficzny
Reliable physics-informed neural networks for Navier-Stokes simulations : can we trust AI-generated numerical simulations? / Tomasz SŁUŻALEC, Daria WÓJCIK, Carlos Uriarte, Marcin ŁOŚ, Anna PASZYŃSKA, Maciej PASZYŃSKI // Journal of Computational Science ; ISSN 1877-7503 . — 2026 — vol. 95 art. no. 102817, s. 1-8. — Bibliogr. s. 8, Abstr. — Publikacja dostępna online od: 2026-02-23
Autorzy (6)
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 167146 |
|---|---|
| Data dodania do BaDAP | 2026-05-29 |
| Tekst źródłowy | URL |
| DOI | 10.1016/j.jocs.2026.102817 |
| Rok publikacji | 2026 |
| Typ publikacji | artykuł w czasopiśmie |
| Otwarty dostęp |  |
| Czasopismo/seria | Journal of Computational Science |
Abstract
The Navier–Stokes equations are one of the most important computational problems. They are employed for modeling the flow of the air around the airplane, for the verification of the airplane’s safety. The Navier–Stokes equations or their simplified version, the shallow water equations, are used for tsunami prediction simulations. In the extended version, the Navier–Stokes–Boussinesq equations are employed for weather prediction simulations. In this paper, we show how challenging it is for PINNs with a robust loss function to solve the model Navier–Stokes problem, the lid-driven cavity flow, especially at high Reynolds numbers. If PINNs cannot solve a model lid-driven cavity problem, how can we trust the AI-generated numerical simulations? We investigate the common understanding and trust into AI generated numerical simulations among the computer science students, and we relate these findings to the conclusions from the numerical simulations. It is very important to understand the accuracy of AI simulators at a reliable level to prevent future disasters resulting from their misuse. Finally, we provide Google Colab and GitHub links to numerical codes, namely PINNs and CRVPINNs in Python, and a C++ FEM code in FeniCs.
