Back to search

National Technical University of Athens (NTUA)

The continuous adjoint method with consistent discretization schemes for transitional flows and the Use of deep neural networks in shape optimization in fluid mechanics

Abstract

dc:description

In this PhD Thesis, methods and tools for shape optimization in fluid dynamics are developed, assessed and applied. Emphasis is laid on (a) the mathematical formulation, programming and assessment of the continuous adjoint method for shape optimization in transitional flows of compressible fluids, (b) the development of consistent discretization schemes for the adjoint PDEs and adjoint boundary conditions involved in continuous adjoint and (c) the use of Deep Neural Networks (DNNs) in flow predictions and demonstration of the expected gain in shape optimization. The two first subjects are related to Gradient-Based (GB) optimization, whereas the last one is associated with the more efficient use of Evolutionary Algorithms (EAs).Throughout this Thesis, the in–house GPU–enabled (U)RANS solver PUMA, developed by the Parallel CFD & Optimization Unit of the NTUA, is used. The compressible flow solver of PUMA, previously developed for turbulent flows, is extended to transitional flows. The γ-Reθt transition model is coupled with both the k−ω SST and the Spalart–Allmaras turbulence models. A series of 2D or 3D cases, in internal and external aerodynamics, are used for the purpose of comparison either with experimental or numerical results. These cases include the transitional flow over flat-plates and the flow through a highly loaded transonic turbine blade which is evaluated using aerothermal criteria. Regarding external aerodynamics, an isolated airfoil and two wings, all with Natural Laminar Flow are used. The adjoint method is used to compute the Sensitivity Derivatives (SDs) of any objective function with respect to (w.r.t.) the design variables in the most efficient way, in a GB optimization. Herein, the continuous adjoint method is mathematically developed for the γ−Reθt transition model coupled with the Spalart–Allmaras turbulence model, for the first time in the corresponding literature. The ‘‘frozen transition’’ adjoint (i.e. skipping the adjoint to the transition model equations) is evaluated as a possible cheaper alternative and is shown to noticeably damage the accuracy of SDs. The main advantages of the continuous adjoint method is the clear physical insight into the adjoint PDEs and their boundary conditions and, in particular, the low memory footprint of the adjoint code. The discretization schemes used for the adjoint equations can significantly affect the accuracy of SDs; ideally these should be consistent with the discretized flow (primal) problem. In discrete adjoint, consistency is not an issue; the main weaknesses are the excessive memory footprint of the adjoint code and the lack of a clear interpretation of the adjoint equations. This PhD Thesis bridges the gap between the two adjoint variants by proposing consistent discretization schemes (inspired by discrete adjoint) for the continuous adjoint PDEs and their boundary conditions, with a clear physical meaning. The new Think-Discrete Do-Continuous (TDDC) adjoint is developed for the time-marching hyperbolic-type solver of compressible fluid flows of the PUMA code. The TDDC adjoint computes SDs as accurate as discrete adjoint, ensuring though low memory footprint and physical insight into the adjoint problem. The TDDC adjoint outperforms the continuous adjoint that makes use of the ‘‘standard’’ discretization schemes and the more accurate SDs result in improved optimization behavior. The continuous adjoint method for transitional flows, that also makes use of the developed consistent discretization schemes, is used to carry out two industrial optimizations, that of a wing and that of a high aspect-ratio wing business jet configuration. In its last part, which focuses on cost reduction in shape optimization problems, this PhD Thesis develops and uses DNNs for flow prediction and shape optimization. A first approach is to train DNNs and use them to replace the numerical solution of the turbulence and/or transition model PDEs by providing the turbulent viscosity field in each pseudo–time iteration of the RANS equations’ solver. The optimal DNN configuration (architecture and input data) is found from a Metamodel–Assisted Evolutionary Algorithm (MAEA). The resulting cost–effective solver is used as an evaluation tool during the shape optimization of an isolated airfoil, a turbomachinery cascade and a car model using the MAEA of the EASY (Evolutionary Algorithm SYstem) platform. A second approach is to use DNNs in multi-disciplinary and multi-row turbomachinery shape optimization problems. The use of the two-branch λ−DNN architecture is proposed; its distinguishing feature is the separate processing of input data of different nature. Consequently, the λ−DNN is ideal for use in multidisciplinary problems by replacing one or more disciplines. Herein, it is used in Conjugate Heat Transfer (CHT) shape optimization problems by replacing the numerical solution of the heat conduction equation solver in the solid material. The new, less computationally expensive solver is used as an evaluation tool in two CHT shape optimizations: an S-Shaped cooling channel and an internally cooled turbine blade-airfoil. Finally, for multi-row turbomachinery optimization problems, a DNN surrogate to the mixing plane technique is proposed. Overall, the use of DNN-assisted solvers reduces the cost of optimizations using MAEAs and the EASY platform, in specific.

Degree

thesis:*
Grantor dc:publisher
National Technical University of Athens (NTUA)
Year dc:date
2023

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Kontou, Marina
  • Κοντού, Μαρίνα

Subjects

dc:subject × 20

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/56951

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Kontou, Marina; Κοντού, Μαρίνα. The continuous adjoint method with consistent discretization schemes for transitional flows and the Use of deep neural networks in shape optimization in fluid mechanics. National Technical University of Athens (NTUA), 2023. http://hdl.handle.net/10442/hedi/56951