Discrete - Residual Loss for training PINNs
Автор: OpenSteam
Загружено: 2026-02-20
Просмотров: 12
Описание:
In this video, we explore a discrete Finite Volume (FVD) residual loss for training Physics-Informed Neural Networks (PINNs), replacing traditional Automatic Differentiation (AD).
We explain why FVD reduces computational cost, accelerates convergence, and performs better in nonlinear and high Reynolds number fluid flow problems. Through experiments on Kovasznay flow, lid-driven cavity flow, and flow past a cylinder, we compare training time, convergence behavior, and accuracy.
Results show that FVD-based PINNs train faster, require fewer epochs, and achieve competitive or superior accuracy compared to standard AD-based approaches.
link - https://www.iccs-meeting.org/archive/...
Повторяем попытку...
Доступные форматы для скачивания:
Скачать видео
-
Информация по загрузке:
