Authors:
Philipp Moser
;
Wolfgang Fenz
;
Stefan Thumfart
;
Isabell Ganitzer
and
Michael Giretzlehner
Affiliation:
Research Department Medical Informatics, RISC Software GmbH, Softwarepark 32a, 4232 Hagenberg, Austria
Keyword(s):
Physics-Informed Neural Networks, Fluids, Navier-Stokes, Deep Learning.
Abstract:
Physics-Informed deep learning methods are attracting increased attention for modeling physical systems due to their mesh-free approach, their straightforward handling of forward and inverse problems, and the possibility to seamlessly include measurement data. Today, most learning-based flow modeling reports rely on the representational power of fully-connected neural networks, although many different architectures have been introduced into deep learning, each with specific benefits for certain applications. In this paper, we successfully demonstrate the application of physics-informed neural networks for modeling steady and transient flows through 3D geometries. Our work serves as a practical guideline for machine learning practitioners by comparing several popular network architectures in terms of accuracy and computational costs. The steady flow results were in good agreement with finite element-based simulations, while the transient flows proved more challenging for the continuou
s-time PINN approaches. Overall, our findings suggest that standard fully-connected neural networks offer an efficient balance between training time and accuracy. Although not readily supported by statistical/practical significance, we could identify a few more complex architectures, namely Fourier networks and Deep Galerkin Methods, as attractive options for accurate flow modeling.
(More)