Back to search

National Technical University of Athens (NTUA)

Topology and shape optimization in fluid mechanics & conjugate heat transfer using continuous adjoint with consistent discretization

Abstract

dc:description

This PhD thesis develops methods to perform adjoint-based Topology Optimization (TopO) in Computational Fluid Dynamics (CFD) problems with (or without) Conjugate Heat Transfer (CHT). All methods and tools are developed within the open–source CFD code of OpenFOAM, extending the adjointOptimisation library programmed and made publicly available by the Parallel CFD & Optimization Unit of the NTUA. A key contribution of this thesis is to alleviate some of the well–known weaknesses of the widely–used density–based TopO (denTopO) approach, related to the accuracy of the corresponding analysis code. denTopO is performed on a static CFD grid by introducing a single design variable per grid cell which stands for an impermeability and indicates whether each cell is occupied by fluid or solid material. Ideally, binary designs, where all the design variables take on 0/1 values, are expected. In view of this and to avoid well–known checkerboard pathologies, the filtering of the design variables based on a Helmholtz PDE is performed. This yields a smoother field that is later projected using a smooth/differentiable Heaviside function to become sharper. The so-computed field is introduced to the flow equations through the Brinkman penalization terms, standing for blockage forces inside the solid parts of the domain. The flow is modeled by the Reynolds-Averaged Navier–Stokes (RANS) equations coupled with the Spalart–Allmaras turbulence model. CHT problems require solving the temperature equation over the fluid and solid domains. Due to the one–way coupling between the mean flow and the temperature equation, the latter is solved as a post–processing step by making use of proper interpolation schemes for the thermophysical properties based on the values of the design variables. This thesis relies upon the continuous adjoint method which computes the gradients of the objective and constraint functions with respect to (w.r.t.) the design variables. This computation comes at a cost that does not depend on the number of design variables which, in TopO, is equal the number of grid cells.The ability of the denTopO method to compute innovative designs and unexpected topologies comes with lower accuracy of the CFD/CHT solution, as the denTopO solver is unable to impose exact conditions on the Fluid Solid Interfaces (FSIs). Instead, it models solid areas as porous regions of high impermeability, without being able to avoid spurious/unwanted flow leakage. Extracting the FSI from the TopO solution is a task that, depending on the way it is performed and the formulation of TopO (e.g. density–based, level–set etc.), might lead to slightly different optimized solutions. For this reason, and due to the aforementioned inaccuracies of denTopO, performances may change if the optimized solutions from denTopO are re–evaluated on body–fitted grids. As a first remedy, this thesis proposes to compute the FSI upon completion of denTopO and, then, generate body–fitted grids in order to proceed with Shape Optimization (ShpO) as a means to further refine the denTopO solutions; however, this increases the computational cost. To eliminate the lack of accuracy of the denTopO solver in the near wall regions and avoid continuing with ShpO, this thesis proposes and implements the cut–cell method which computes the intersections of the FSI with the cells of the background CFD grid in each cycle of TopO. This leads to the cut–cell TopO method which is still based on the impermeability field of denTopO. However, in contrast to denTopO, this is used exclusively to compute the FSI. The solution of the flow and CHT equations is, thus, performed, on the cut–cell grid, allowing the imposition of exact boundary conditions on the FSI. Moreover, cut–cell TopO applies h-refinement close to the computed FSIs. ShpO is no more needed, and accurate performance values as well as a clear FSI become available. To update the design variables, the derivatives of the objective and constraints w.r.t. the node positions on the FSIs are computed first, and are then transformed into derivatives of the same variables w.r.t. cell impermeabilities by applying the chain rule of differentiation. The superiority of the cut–cell TopO method w.r.t. the ‘‘standard’’ denTopO method is assessed by comparing the results in different test cases. The next part of this dissertation pertains to the TopO of two-fluid heat exchangers, performed by extending the previously developed denTopO and cut–cell TopO tools. In such cases, it is important to ensure the non-mixing of the two fluids. The developed framework relies on a single field of design variables to parameterize regions occupied by solid material or either of the two working fluids. Compared to mono–fluid TopO, the bi–fluid TopO method incorporates the filtering and projection operators in two steps. The second step controls the user-specified minimum solid thickness separating the two fluids, and yields two distinct fluid indicator fields. In denTopO, these are used to impose the Brinkman terms in the flow equations, solved separately for each fluid. In cut–cell TopO, the two fluid indicators are used to compute the FSI and generate cut–cells, without affecting directly the governing equations. The developed tools for TopO, either based on the ‘‘standard’’ density approach or the cut–cell method, are assessed in problems of academic and of practical interest. 2D internal flow problems targeting min. total pressure drop and/or max. heat exchange are tackled. The application of mono–fluid TopO for the minimization of total pressure losses inside an HVAC duct, bringing cool air from the front console to the rear part of a passenger car, is also worked out. For bi–fluid TopO, both the density-based and the cut-cell method are assessed in the optimization of a compact heat exchanger with many inlets and outlets, for the two working fluids.As mentioned before, both ShpO and TopO are supported by continuous adjoint. Selecting proper discretization schemes for the solution of the continuous adjoint problem affects the accuracy of the computed derivatives. On the other hand, discrete adjoint solvers provide consistent gradients, as the discretized primal residuals are differentiated. This PhD thesis develops the ‘‘Think Discrete-Do Continuous’’ (TDDC) adjoint for the incompressible RANS equations solved in OpenFOAM using the SIMPLE algorithm for the pressure–velocity coupling. In the TDDC adjoint, the development starts by hand–differentiating the discretized residuals of the primal solver, as in discrete adjoint. The terms emerging from this development are appropriately re–arranged to reverse–engineer consistent fluxes for the discretization of all terms in the continuous adjoint PDEs. The new adjoint solver combines the best of discrete and continuous adjoint, as it gives exact/consistent gradients of the objective function which are in perfect agreement with those computed by Finite Differences (FDs) and maintains a low memory footprint. This dissertation provides two ways of solving the set of adjoint PDEs. The first is based on a fixed–point iterative scheme derived by first rewriting the segregated SIMPLE algorithm in the form of a fully–coupled left–preconditioned Fixed Point Iteration (FPI). The transpose of the derived preconditioning matrix is then used for the adjoint FPI. Even in fully coupled form, the preconditioner finally acts on the adjoint residuals in a segregated manner. This leads to the adjoint to the SIMPLE algorithm that performs the exact same operations as the primal algorithm but in reverse order. This strategy is duality–preserving as the adjoint solver inherits the convergence characteristics of the primal FPI. However, in real–world applications the primal solver is likely not to converge tightly/monotonically, but enter Limit Cycle Oscillations (LCO). In such cases, solving the steady–state adjoint problem by making a linearization around the last snapshot of the flow solution is likely to diverge since the adjoint FPI (derived directly from the primal FPI) might not be contractive, i.e. might possess eigenvalues of magnitude greater than 1. To address this issue, the solution of the adjoint problem in a fully coupled manner by making use of Krylov subspace linear solvers is proposed. The full adjoint system is solved using the Flexible variant of the Generalized Minimal Residual method with Deflated Restarting (FGMRES-DR) between restarts to avoid residual stagnation. The linear solver is implemented in OpenFOAM and makes use of the developed adjoint SIMPLE algorithm as preconditioner. The coupled Krylov adjoint solver is robust enough to converge even in problems where the primal solver enters LCO, with enough, though, Krylov basis vectors. Occasionally, the coupled adjoint solver may significantly reduce the cost of solving the adjoint problem, at the expense of a larger memory footprint.

Degree

thesis:*
Grantor dc:publisher
National Technical University of Athens (NTUA)
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Galanos, Nikolaos
  • Γαλανός, Νικόλαος

Subjects

dc:subject × 28

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/58672

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

Galanos, Nikolaos; Γαλανός, Νικόλαος. Topology and shape optimization in fluid mechanics & conjugate heat transfer using continuous adjoint with consistent discretization. National Technical University of Athens (NTUA), 2025. http://hdl.handle.net/10442/hedi/58672