Investigations on Physics-Informed Neural Networks for Aerodynamics

Guillaume Coulaud,Maxime Le,Régis Duvigneau

2024-03-26

Abstract:Physics-Informed Neural Networks (PINNs) have recently emerged as a novel approach to simulate complex physical systems on the basis of both data observations and physical models. In this work, we investigate the use of PINNs for various applications in aerodynamics and we explain how to leverage their specific formulation to perform some tasks effectively. In particular, we demonstrate the ability of PINNs to construct parametric surrogate models, to achieve multiphysic couplings and to infer turbulence characteristics via data assimilation. The robustness and accuracy of the PINNs approach are analysed, then current issues and challenges are discussed.

Analysis of PDEs,Optimization and Control,Classical Physics

What problem does this paper attempt to address?

This paper discusses the application of Physics-Informed Neural Networks (PINNs) in fluid mechanics, particularly for non-classical problems. PINNs combine data observations and physical models to construct hybrid models, which can be used for constructing parameterized surrogate models, implementing multi-physics coupling, and inferring turbulence characteristics through data assimilation. The paper analyzes the robustness and accuracy of the PINNs method, and discusses current issues and challenges.

Firstly, PINNs aim to simulate physical systems using multilayer perceptron neural networks, which are trained to conform to known physical rules and observation data. They are particularly suitable for dealing with problems that require real-time response to a series of physical parameters and facilitate interdisciplinary coupling. In addition, PINNs can estimate turbulence characteristics using data and models.

In terms of methodology, the core of PINNs is a multilayer perceptron that is trained based on physical rules and data. The loss function includes terms related to PDE residuals, initial and boundary conditions, and is minimized using gradient descent. Automatic differentiation is used to compute the loss terms, while optimization of network parameters may involve different optimization algorithms such as ADAM and L-BFGS.

The paper demonstrates the applications of PINNs in parameterized simulations (such as flow simulations within different ranges of physical parameters), multi-physics coupling (such as heat conduction problems), and data assimilation (such as estimating turbulence characteristics through observation data). These applications illustrate the flexibility of PINNs in solving complex problems, but also highlight the challenges of convergence, accuracy assurance, and loss function optimization.

In conclusion, the paper studies the effectiveness of PINNs in solving non-classical problems in fluid mechanics and proposes possible future research directions and improvement strategies.