Back to search

National Technical University of Athens (NTUA)

Computational cost reduction and stabilization of solvers of the continuous adjoint method in aerodynamic shape optimization, with or without uncertainties

Abstract

dc:description

The objective of this thesis is the development and assessment of techniques for the reduction of the computational cost of gradient-based aerodynamic shape optimization for flows that exhibit mild unsteadiness and/or in the presence of uncertainties. All these techniques are based on the continuous adjoint method and have been developed in the OpenFOAM environment, as a complement to the adjoint solver developed and made publicly available by the Parallel CFD & Optimization Unit of the National Technical University of Athens (PCOpt/NTUA). They are first validated in academic cases and are later used for aerodynamic shape optimization in industrial applications. It is well known that adjoint-based shape optimization using unsteady solvers is costly and/or memory demanding. For this reason, it is sparingly used in real-world applications. When mild unsteadiness is present, steady primal and adjoint solvers can be used instead provided that convergence difficulties are properly treated. In this thesis and for this type of applications, the steady primal and the corresponding (continuous) adjoint solvers are both stabilized by implementing the Recursive Projection Method (RPM), a stabilization method for iterative procedures. Upon completion of the optimization using steady solvers, unsteady re-evaluations of the optimized solutions should normally be used to confirm a reduction in the time-averaged objective function. Thus, shape optimization costs less, even by an order of magnitude, than an optimization using unsteady adjoint. This approach is further tested in flows with vortex shedding wherein steady flow solutions are computed with the help of the RPM. Although vortex shedding problems can be optimized with steady solvers, optimization using unsteady adjoint cannot be avoided. However, the latter starts with a good initialization, computed by the steady solvers, and this reduces a lot the overall CPU cost. In complex cases, the convergence difficulties of the primal and adjoint solvers are far more prevalent. For instance, the RPM may not necessarily ensure convergence of the adjoint solver on its own. This might be due to the lack of good grid quality, often in applications involving complex geometries (e.g. a car with wheels, side mirrors, surface details etc.), or the existence of many unstable modes causing rapid divergence of the solver. To overcome this difficulty, the controlled damping of the Adjoint Transposed-Convection (ATC) term in the adjoint momentum equations is occasionally employed along with the RPM. In the literature, the damping or, even, elimination of the ATC term is frequently used for the stabilization of continuous adjoint solvers. The effect of this treatment on the computed Sensitivity Derivatives (SDs) is studied in cases in which the RPM, on its own, makes the adjoint solver converge. Controlled/mild ATC damping (if needed), as applied in this thesis, proves to be harmless and greatly contributes to robustness. On the contrary, the uncontrolled damping or, even, elimination of the ATC, an action that would have been required in the absence of the RPM, may become harmful. In addition, the RPM stabilization helps avoiding the solution of the adjoint equations based on averaged primal fields. In the second part, this thesis tackles aerodynamic shape optimization in the presence of uncertainties. In such problems, Uncertainty Quantification (UQ) techniques are used to measure the effect of uncertain variables on a Quantity of Interest (QoI). The inclusion of UQ in a shape optimization loop may noticeably increase the computational cost. Two UQ methods, namely the First-Order Second-Moment (FOSM) and the Adjoint-assisted Polynomial Chaos Expansion (APCE) are studied. Both FOSM and APCE compute the first two statistical moments (mean value and standard deviation) of the QoI and the objective function to be minimized is their weighted sum. Gradient-based optimization with such an objective function requires mixed derivatives of the QoI with respect to the design and uncertain variables. However, in practice, only the projection of the matrix of mixed derivatives on some vectors needs to be computed, instead of the matrix itself. These projections are herein exploited in both UQ methods to reduce the CPU cost, yielding projected counterparts of these two methods. Their key point is that the cost of computing the projected matrix of mixed derivatives does not scale with the number of either the design or the uncertain variables. In the projected FOSM (pFOSM), the cost per optimization cycle with gradient-based methods is 4 Equivalent Flow Solutions (EFS), i.e. as if the flow equations are solved 4 times. The cost of the projected APCE (pAPCE) is equal to 4L EFS (L is the number of collocation points needed in the regression). The gain in CPU cost is significant even if only a few uncertain variables are involved. The pAPCE method is more expensive than pFOSM but computes more accurate statistical moments. Over and above to some laminar flow problems, pFOSM and pAPCE are demonstrated in aerodynamic shape optimization problems with turbulent flows, solved using the Spalart-Allmaras model and its adjoint. The industrial applications involve aerodynamic shape optimization of ground vehicles using steady solvers and the RPM and of an aircraft wing in the presence of flow uncertainties, to assess pFOSM and pAPCE.

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
  • Skamagkis, Themistoklis
  • Σκαμάγκης, Θεμιστοκλής

Subjects

dc:subject × 22

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/56939

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

Skamagkis, Themistoklis; Σκαμάγκης, Θεμιστοκλής. Computational cost reduction and stabilization of solvers of the continuous adjoint method in aerodynamic shape optimization, with or without uncertainties. National Technical University of Athens (NTUA), 2023. http://hdl.handle.net/10442/hedi/56939