National Technical University of Athens (NTUA)
Unsteady RANS & (D)DES continuous adjoint and compression techniques in aerodynamic shape optimization
Abstract
dc:descriptionThis PhD thesis focuses on aerodynamic gradient-based shape optimization in problems governed by unsteady PDEs, such us the unsteady Navier-Stokes equations, in which gradients with respect to the design variables are computed using the adjoint method. One of the major challenges hindering the widespread use of unsteady adjoint in industry is the backward in time integration of the adjoint equations. During this, previously computed instantaneous flow fields must be available at each time-step, notably increasing memory requirements and/or computational cost. Treatments of this matter include (a) the full storage of the instantaneous flow field time-series, which, despite having the minimum cost in the absence of flow recomputations, is infeasible in large-scale problems due to excessive memory demands, and (b) a zero-storage policy combined with repetitive flow recomputations starting always from the initial state, which drastically increases the computational overhead. The current state-of-the-art technique for handling unsteady adjoint is binomial check-pointing, which reduces memory demands at the expense of a controllable number of flow recomputations. The first objective of this PhD thesis is to reduce both the memory requirements and the computational cost of unsteady adjoint by means of lossy and lossless data compression techniques. These are built by synergistically implementing (a) the incremental Proper Generalized Decomposition (iPGD), (b) the ZFP and (c) the Zlib algorithms, giving rise to the so-called iPGDZ+ compression kernel. Standard PGD requires the complete time-history of the flow to be available prior to compression. It is though obvious that, if the available hardware has sufficient capacity to store the complete flow solution uncompressed, there is no reason for the fields to be compressed. This is the motivation for developing and using its incremental variant iPGD) which compresses snapshots incrementally, each time the flow field at a new time-instant is computed. In this thesis, iPGD is revisited, improved and extended to flow fields computed on unstructured grids. Given that ZFP was initially designed for structured data, three variants of the algorithm are developed to optimize its application on unstructured grids. Two different implementations of the iPGDZ+ algorithm are developed: (a) the Compressed Full Storage (CFS) technique, in which the iPGDZ+ kernel compresses the entire time-history of the flow in memory, thereby avoiding costly flow recomputations, and (b) the Compressed Coarse-grained Check-Pointing (3CP) technique, which combines iPGDZ+ with binomial check-pointing for even greater memory savings, albeit with an increased cost compared to CFS due to flow recomputations. While this thesis focuses solely on aerodynamic shape optimization using continuous adjoint, the devised compression techniques are general and can readily be applied to any gradient-based optimization problem modeled by unsteady PDEs, using either continuous or discrete adjoint. Herein, the CFS and 3CP techniques are assessed in terms of memory savings, computational cost and representation accuracy and compared with binomial check-pointing. The second objective of this thesis is to extend the continuous adjoint method to higher fidelity eddy-resolving flow models, for the first time in the literature. This involves the development of the continuous adjoint to the incompressible Detached Eddy Simulation (DES) and Delayed-DES (DDES) models, both based on the Spalart-Allmaras turbulence model. It is shown that solving the adjoint to the turbulence model equation is crucial for the computation of accurate derivatives of the objective function with respect to the design variables, which closely match reference values obtained using Finite Differences. In contrast, derivatives computed under the “frozen turbulence” assumption, which neglects variations in turbulent viscosity due to changes in the design variables, are prone to inaccuracies. As (D)DES models are inherently unsteady, the corresponding adjoint methods can also benefit from compression techniques developed in this thesis. All methods are implemented as an extension to the publicly available adjointOptimisation library of OpenFOAM, developed by the PCOpt/NTUA. They are first applied to academic aerodynamic shape optimization problems and, then, across a wide range of industrial automotive problems governed by the Unsteady Reynolds-Averaged Navier-Stokes (URANS) and/or (D)DES flow models. These applications are concerned with the shape optimization of various ground vehicles for minimum drag. In all cases, the CFS and 3CP techniques reduce the computational cost per optimization cycle by approximately 30% and 5%, respectively, compared to check-pointing. Additionally, CFS reduces memory footprint by a factor of 5 to 35 compared to check-pointing (or 350 to 1200 compared to full storage without compression) in the URANS cases and by a factor of 1.2 to 8 compared to check-pointing (or 70 to 120 compared to full storage) in the DDES cases. Meanwhile, 3CP achieves even higher memory savings, ranging from ÷30 to ÷60 compared to check-pointing (or ÷1900 to ÷2500 compared to full storage) in the URANS cases and from ÷15 to ÷40 compared to check-pointing (or approximately ÷1000 compared to full storage) in the DDES cases. To extend the application domain, (some of the) compression techniques are also implemented in the GPU-enabled solver PUMA, also developed by the PCOpt/NTUA. This allows the use of the developed compression techniques in computing objective function gradients in the case of a high aspect-ratio wing aircraft model, governed by the compressible URANS equations and featuring shock waves, thereby confirming the versatility of the developed compression tools.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Margetis, Andreas-Stefanos
- Μαργέτης, Ανδρέας-Στέφανος
Subjects
dc:subject × 22- Αεροδυναμική βελτιστοποίηση μορφής
- Υπολογιστική ρευστοδυναμική
- Χρονικά μη-µόνιµη συνεχής συζυγής μέθοδος
- Τεχνικές συµπίεσης δεδοµένων
- Σταδιακός ιδιο-γενικευµένος διαχωρισµός
- ZFP
- DES
- DDES
- Spalart-Allmaras
- OpenFOAM
- Aerodynamic shape optimization
- Computational fluid dynamics
- Continuous adjoint
- Unsteady adjoint
- Data compression
- Incremental proper generalized decomposition
- Επιστήμη Μηχανολόγου Μηχανικού
- Επιστήμες Μηχανικού και Τεχνολογία
- Μηχανολογία
- Mechanical Engineering
- Engineering and Technology
- Mechanics
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/57297