Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Civil and Environmental Engineering | en_US |
| dc.contributor.advisor | Ni, Yi-qing (CEE) | en_US |
| dc.contributor.advisor | Chen, Zheng-wei (CEE) | en_US |
| dc.creator | Rui, Enze | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/13855 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | Development of ameliorative methods for reynolds-averaged navier-stokes simulations based on physics-informed machine learning | en_US |
| dcterms.abstract | Reynolds-averaged Navier-Stokes (RANS) simulation is a widely employed numerical approach for turbulence modeling. The computational fluid dynamics (CFD) method has long been the predominant approach in RANS simulation, owing to its intuitiveness, ease of implementation, and high accuracy in modeling. Some well-known CFD methods such as the finite element method and finite volume method have been adopted in RANS simulation in past decades. However, as research progresses, the shortcomings of the CFD-based RANS simulations are gradually becoming apparent. Several consensus problems are enumerated here. For instance, the CFD methods require manual meshing, which may induce various kinds of grid generation issues. In addition, the RANS equations incorporate the Reynolds stress terms as additional unknown variables during the averaging process. The Reynolds stress modeling resulting from the action of Reynolds averaging contributes to the non-universality of RANS simulations. | en_US |
| dcterms.abstract | In recent years, a novel machine learning-based solver for partial differential equations (PDEs), i.e., physics-informed neural network (PINN), has emerged. Since its inception, it has shown a considerable impact in the field of fluid mechanics. Researchers have also been identifying the potential of PINN’s applications in RANS simulations. The advantages of this method over CFD methods are evident. Firstly, being a meshless approach, it does not encounter any grid-related issues. For instance, the partial derivative terms in the PDEs are calculated using the automatic differentiation function in a PINN, eliminating any truncation error that may arise from grid methods. In addition, PINN has a neural network foundation, while it can integrate data information into its simulation, and thus derive physics-based data-driven solutions. | en_US |
| dcterms.abstract | Nevertheless, it must be acknowledged that while PINN possesses certain advantages over conventional CFD methods when solving the RANS equations, this emerging machine learning solver for PDEs still faces several challenges. First and foremost, the convergence performance of a PDE solver is a crucial indication for assessing its effectiveness in solving equations. However, studies have shown that minimizing the loss of a PINN may sometimes be challenging using these gradient-based methods. Secondly, there is the issue of the limitation of the nonlinear expression and feature learning capabilities for PINN-based RANS simulations. Learning high-frequency features using neural networks has been found to be difficult in previous research. Furthermore, although grid issues do not plague PINNs, the limited applicability of the RANS turbulence models under various flow conditions still exists. One of the most difficult issues to solve is still the search for a universal turbulence simulation approach in PINN-based RANS simulations. | en_US |
| dcterms.abstract | The entire thesis is divided into seven chapters, with the first chapter being an introduction to the entire thesis and the second chapter being a review of existing methods. Based on the discussions in Chapters 1 and 2, this thesis proposes four amelioration measures grounded in the PINN framework, namely Dynamic Prioritization in Chapter 3, Multifidelity Modeling in Chapter 4, Quantum Layer Integration in Chapter 5, and Weighted Sum Turbulence Model in Chapter 6. These amelioration measures are dedicated to alleviating the aforementioned key issues in PINN-based RANS simulations. Some are used to improve the convergence performance of a PINN, some can accelerate the learning ability of a PINN for extracting high-frequency features, and some can alleviate the issue of poor applicability of RANS turbulence models. These proposed methods have been validated by using experiments and achieved satisfactory results, promoting the further developments of physics-informed machine learning methods in turbulence modeling. At the end of this research, the future work is prospected, and the future research direction is pointed out. | en_US |
| dcterms.extent | xxxv, 213 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2025 | en_US |
| dcterms.educationalLevel | Ph.D. | en_US |
| dcterms.educationalLevel | All Doctorate | en_US |
| dcterms.accessRights | open access | en_US |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13855