Post-seismic deformation and glacial isostatic adjustment are two processes by which the Earth deforms viscoelastically. In both cases, the details of the deformation depend on the rheological structure of the Earth as well as the forcing, which is the earthquake and further movement on the fault in the case of post-seismic deformation, and the change in load on the surface of the Earth due to the redistribution of water and ice mass in the case of glacial isostatic adjustment. It is therefore possible to learn about the Earth's rheological structure and the processes' respective forcings from measurements of the deformation. In order to use measurements in this way, it is first necessary to have a method of forward modelling the processes, that is, calculating the deformation due to a given forcing and in an earth model with a given structure. Given this, a way of calculating derivatives of measurements of the deformation with respect to the parameters of interest is then desirable. In this dissertation, the adjoint method is used. This, for the first time, enables efficient calculation of continuous derivatives, which have many potential applications. Firstly, they can be used within a gradient-based optimisation method to find a model which minimises some data misfit function. The derivatives can also be used to quantify the uncertainty in such a model and hence to provide understanding of which parts of the model are well constrained. Finally, they enable construction of measurements which provide sensitivity to a particular part of the model space. In this dissertation, new methods for forward modelling both post-seismic deformation and glacial isostatic adjustment are presented. The adjoint method is also applied to both problems. Numerical examples are presented in spherically symmetric earth models and, in the case of glacial isostatic adjustment, models with laterally varying rheological structure. Such examples are used to illustrate the potential applications of the developments made within this dissertation.

This thesis investigates stabilisation of the SIMPLE-family discretisations for incompressible flow and their discrete adjoint counterparts. The SIMPLE method is presented from typical \prediction-correction" point of view, but also using a pressure Schur complement approach, which leads to a wider class of schemes. A novel semicoupled implicit solver with velocity coupling is proposed to improve stability. Skewness correction methods are applied to enhance solver accuracy on non-orthogonal grids. An algebraic multi grid linear solver from the HYPRE library is linked to flow and discrete adjoint solvers to further stabilise the computation and improve the convergence rate. With the improved implementation, both of flow and discrete adjoint solvers can be applied to a wide range of 2D and 3D test cases. Results show that the semi-coupled implicit solver is more robust compared to the standard SIMPLE solver. A shape optimisation of a S-bend air flow duct from a VW Golf vehicle is studied using a CAD-based parametrisation for two Reynolds numbers. The optimised shapes and their flows are analysed to con rm the physical nature of the improvement. A first application of the new stabilised discrete adjoint method to a reverse osmosis (RO) membrane channel flow is presented. A CFD model of the RO membrane process with a membrane boundary condition is added. Two objective functions, pressure drop and permeate flux, are evaluated for various spacer geometries such as open channel, cavity, submerged and zigzag spacer arrangements. The flow and the surface sensitivity of these two objective functions is computed and analysed for these geometries. An optimisation with a node-base parametrisation approach is carried out for the zigzag con guration channel flow in order to reduce the pressure drop. Results indicate that the pressure loss can be reduced by 24% with a slight reduction in permeate flux by 0.43%.

京都大学 / 0048 / 新制・課程博士 / 博士(理学) / 甲第18081号 / 理博第3959号 / 新制||理||1571(附属図書館) / 30939 / 京都大学大学院理学研究科地球惑星科学専攻 / (主査)准教授 宮﨑 真一, 教授 福田 洋一, 教授 平原 和朗 / 学位規則第4条第1項該当 / Doctor of Science / Kyoto University / DGAM

Improving the design of flow channels in a liquid-cooled heat sink is critical for boosting the capabilities of electronic components as well as reducing energy usage by the pump. This work explores the use of topology optimization to minimize the pressure difference across a heat sink and consequently, the energy used to supply the liquid. Topology optimization involves solving mathematical equations to obtain the optimal design for a defined cost function, here the total pressure difference between the inlet and outlet. A design variable called the porosity is defined for each cell in the mesh. The porosity features in a sink term in the momentum equation, which 'solidifies' cells by velocity suppression when deemed to be counterproductive to the cost function. The adjoint method of topology optimization, in particular, is a well-established tool for use in flow network problems and includes non-physical parameters such as the adjoint velocity and pressure. The method isn't without its drawbacks, such as the numerical instability of the adjoint equations, and the absence of boundary layers or wall functions at the interface of high and low porosity. The strength of the adjoint method lies in the ease with which it calculates the gradient of the cost function with respect to the porosity. When applied to the geometries in this work, it is observed that the problem is non-convex and results in multiple optimums with similar cost values. Thus the objective becomes seeking solutions with the simplest shape and at the same time having a minimized pressure difference. Interesting techniques are tested, namely an interpolation function, a velocity tolerance, and a volume constraint. The work is accomplished by modifying an existing adjoint optimization solver in the open-source CFD software, OpenFOAM.

Continuous adjoint methods and optimal control theory are applied to a pressure-matching inverse design problem of quasi 1-D nozzle flows. Pontryagin’s Minimum Principle is used to derive the adjoint system and the reduced gradient of the cost functional. The properties of adjoint variables at the sonic throat and the shock location are studied, revealing a logarithmic singularity at the sonic throat and continuity at the shock location. A numerical method, based on the Steger-Warming flux-vector-splitting scheme, is proposed to solve the adjoint equations. This scheme can finely resolve the singularity at the sonic throat. A non-uniform grid, with points clustered near the throat region, can resolve it even better. The analytical solutions to the adjoint equations are also constructed via Green’s function approach for the purpose of comparing the numerical results. The pressure-matching inverse design is then conducted for a nozzle parameterized by a single geometric parameter. In the second part, the adjoint methods are applied to the problem of minimizing drag coefficient, at fixed lift coefficient, for 2-D transonic airfoil flows. Reduced gradients of several functionals are derived through application of a Lagrange Multiplier Theorem. The adjoint system is carefully studied including the adjoint characteristic boundary conditions at the far-field boundary. A super-reduced design formulation is also explored by treating the angle of attack as an additional state; super-reduced gradients can be constructed either by solving adjoint equations with non-local boundary conditions or by a direct Lagrange multiplier method. In this way, the constrained optimization reduces to an unconstrained design problem. Numerical methods based on Jameson’s finite volume scheme are employed to solve the adjoint equations. The same grid system generated from an efficient hyperbolic grid generator are adopted in both the Euler flow solver and the adjoint solver. Several computational tests on transonic airfoil design are presented to show the reliability and efficiency of adjoint methods in calculating the reduced (super-reduced) gradients. / Ph. D.

Adaptivity in both space and time has become the norm for solving problems modeled by partial differential equations. The size of the discretized problem makes uniformly refined grids computationally prohibitive. Adaptive refinement of meshes and time steps allows to capture the phenomena of interest while keeping the cost of a simulation tractable on the current hardware. Many fields in science and engineering require the solution of inverse problems where parameters for a given model are estimated based on available measurement information. In contrast to forward (regular) simulations, inverse problems have not extensively benefited from the adaptive solver technology. Previous research in inverse problems has focused mainly on the continuous approach to calculate sensitivities, and has typically employed fixed time and space meshes in the solution process. Inverse problem solvers that make exclusive use of uniform or static meshes avoid complications such as the differentiation of mesh motion equations, or inconsistencies in the sensitivity equations between subdomains with different refinement levels. However, this comes at the cost of low computational efficiency. More efficient computations are possible through judicious use of adaptive mesh refinement, adaptive time steps, and the discrete adjoint method. This dissertation develops a complete framework for fully discrete adjoint sensitivity analysis and inverse problem solutions, in the context of time dependent, adaptive mesh, and adaptive step models. The discrete framework addresses all the necessary ingredients of a state–of–the–art adaptive inverse solution algorithm: adaptive mesh and time step refinement, solution grid transfer operators, a priori and a posteriori error analysis and estimation, and discrete adjoints for sensitivity analysis of flux–limited numerical algorithms. / Ph. D.

Originally developed in the 1970s for the optimal control of systems governed by partial differential equations, the adjoint method has found several successful applications, e.g., in meteorology with large-scale 3D or 4D atmospheric data assimilation schemes, for carbon cycle data assimilation in biogeochemistry and climate research, or in oceanographic modelling with efficient adjoint codes of ocean general circulation models. Despite the variety of applications in these research fields, adjoint methods have only very recently drawn attention from the ocean acoustics community. In ocean acoustic tomography and geoacoustic inversion, where the inverse problem is to recover unknown acoustic properties of the water column and the seabed from acoustic transmission data, the solution approaches are typically based on travel time inversion or standard matched-field processing in combination with metaheuristics for global optimization. In order to complement the adjoint schemes already in use in meteorology and oceanography with an ocean acoustic component, this thesis is concerned with the development of the adjoint of a full-field acoustic propagation model for shallow water environments. In view of the increasing importance of global ocean observing systems such as the European Seas Observatory Network, the Arctic Ocean Observing System and Maritime Rapid Environmental Assessment (MREA) systems for defence and security applications, the adjoint of an ocean acoustic propagation model can become an integral part of a coupled oceanographic and acoustic data assimilation scheme in the future. Given the acoustic pressure field measured on a vertical hydrophone array and a modelled replica field that is calculated for a specific parametrization of the environment, the developed adjoint model backpropagates the mismatch (residual) between the measured and predicted field from the receiver array towards the source. The backpropagated error field is then converted into an estimate of the exact gradient of the objective function with respect to any of the relevant physical parameters of the environment including the sound speed structure in the water column and densities, compressional/shear sound speeds, and attenuations of the sediment layers and the sub-bottom halfspace. The resulting environmental gradients can be used in combination with gradient descent methods such as conjugate gradient, or Newton-type optimization methods tolocate the error surface minimum via a series of iterations. This is particularly attractive for monitoring slowly varying environments, where the gradient information can be used to track the environmental parameters continuously over time and space. In shallow water environments, where an accurate treatment of the acoustic interaction with the bottom is of outmost importance for a correct prediction of the sound field, and field data are often recorded on non-fully populated arrays, there is an inherent need for observation over a broad range of frequencies. For this purpose, the adjoint-based approach is generalized for a joint optimization across multiple frequencies and special attention is devoted to regularization methods that incorporate additional information about the desired solution in order to stabilize the optimization process. Starting with an analytical formulation of the multiple-frequency adjoint approach for parabolic-type approximations, the adjoint method is progressively tailored in the course of the thesis towards a realistic wide-angle parabolic equation propagation model and the treatment of fully nonlocal impedance boundary conditions. A semi-automatic adjoint generation via modular graph approach enables the direct inversion of both the geoacoustic parameters embedded in the discrete nonlocal boundary condition and the acoustic properties of the water column. Several case studies based on environmental data obtained in Mediterranean shallow waters are used in the thesis to assess the capabilities of adjoint-based acoustic inversion for different experimental configurations, particularly taking into account sparse array geometries and partial depth coverage of the water column. The numerical implementation of the approach is found to be robust, provided that the initial guesses are not too far from the desired solution, and accurate, and converges in a small number of iterations. During the multi-frequency optimization process, the evolution of the control parameters displays a parameter hierarchy which clearly relates to the relative sensitivity of the acoustic pressure field to the physical parameters. The actual validation of the adjoint-generated environmental gradients for acoustic monitoring of a shallow water environment is based on acoustic and oceanographic data from the Yellow Shark '94 and the MREA '07 sea trials, conducted in the Tyrrhenian Sea, south of the island of Elba. Starting from an initial guess of the environmental control parameters, either obtained through acoustic inversion with global search or supported by archival in-situ data, the adjoint method provides an efficient means to adjust local changes with a couple of iterations and monitor the environmental properties over a series of inversions. In this thesis the adjoint-based approach is used, e.g., to fine-tune up to eight bottom geoacoustic parameters of a shallow-water environment and to track the time-varying sound speed profile in the water column. In the same way the approach can be extended to track the spatial water column and bottom structure using a mobile network of sparse arrays. Work is currently being focused on the inclusion of the adjoint approach into hybrid optimization schemes or ensemble predictions, as an essential building block in a combined ocean acoustic data assimilation framework and the subsequent validation of the acoustic monitoring capabilities with long-term experimental data in shallow water environments.

Problemas de otimização se fazem cada vez mais presentes nos mais diversos ramos da Engenharia. Encontrar configurações ótimas para um determinado problema significa, por exemplo, melhorar desempenho, reduzir custos entre outros ganhos. Existem hoje diversas maneiras de atacar um problema de otimização, cada qual com suas particularidades, vantagens e desvantagens. Dentre os métodos de otimização que utilizam gradientes de sensibilidade, o cálculo numérico dos mesmos consiste em uma importante etapa do projeto que, dependendo do problema, pode acarretar em custos computacionais muito elevados inviabilizando a abordagem escolhida. Este trabalho visa desenvolver e apresentar uma nova metodologia para o cálculo desses gradientes de sensibilidade, com base no Método Adjunto. O Método Adjunto é um método amplamente estudado e com diversas aplicações principalmente em Engenharia Aeronáutica. Nesse trabalho, todo o conhecimento prévio é utilizado para a derivação do método para aplicá-lo a escoamentos viscosos e incompressíveis. É desenvolvido também o cálculo do gradiente de sensibilidade com respeito a parâmetros geométricos e não geométricos. Para validar a metodologia proposta são feitas simulações numéricas das equações governantes do escoamento e adjuntas utilizando dois códigos computacionais distintos, SEMTEX e FreeFem++, o primeiro baseado no Método dos Elementos Espectrais e o segundo no Método dos Elementos Finitos, mostrando assim a independência do Método Adjunto na sua formulação contínua em relação a métodos computacionais. Para a validação são cujos gradientes possam ser calculados de outras formas permitindo comparações para calibrar e aperfeiçoar o cálculo do gradiente de sensibilidade. / Optimization problems are widely present in differents fields of Engineering. Finding optimal configurations in a problem means, for example, improving performance, reducing costs, among other achievements. There are several wellknown ways to tackle an optimization problem, each one has its own advantages and disadvantages. Considering the gradient-based optimization methods, the step of their numerical calculation is extremely important, as it may result in huge computational costs, thus making the chosen method impracticable. This work aims to develop and present a new methodology to compute these sensitivity gradients based on the Adjoint Method. The Adjoint Method is a widely studied method with several applications chiefly in A eronautical Engineering. In the present work, all the previous knowledge will be used to derive the equations of the method in order to apply them to viscous incompressible flows. The calculation of the sensitivity gradient, with respect to both geometric and non-geometric paramatersm will be developed as well. To validate the proposed methodology, numerical simulations of the governing and adjoint equations are carried out, using two computational codes called SEMTEX and FreeFem++, the former is based on the Spectral Element Method and the later, on the Finite Element Method, thus showing that the Adjoint Method, in its continuous formulation, is independent of the particular numerical method that is used. In order to validate the algorithm, simple problems are chosen, for which the gradients can be computed by other methods. This choice admits comparison between numerical values of gradients in order to calibrate and improve the methodology proposed.

O método adjunto tem sido extensivamente utilizado como ferramenta de síntese no projeto de aeronaves por permitir que se obtenham sensibilidades de distintas medidas de mérito com relação a parâmetros que controlam a geometria de superfícies aerodinâmicas. O presente trabalho visa uma ampliação das aplicações da formulação contínua do método, ao utilizar propriedades físicas do escoamento nas fronteiras permeáveis do domínio computacional como parâmetros de controle de uma particular medida de mérito. Desse modo é possível, entre muitas possibilidades, determinar a sensibilidade de integrais como sustentação ou arrasto de uma aeronave com relação às condições de cruzeiro, por exemplo. Mais do que isso, essa informação pode ser obtida com a mesma solução adjunta computada para realizar otimização de forma. Vale destacar, ainda, que para que se consiga obter essa informação a partir das equações adjuntas, é necessário que se implemente condições de contorno baseadas em equações diferenciais características, resolvendo o problema de Riemann completo nas fronteiras do domínio. A implementação das usuais condições de contorno homogêneas, vastamente difundidas na literatura, resultaria em gradientes nulos. Esta nova abordagem do método é então aplicada a escoamentos modelados pelas equações de Euler 2-D compressíveis em estado estacionário. Ambos os problemas, físico e adjunto, são resolvidos numericamente com um código computacional que utiliza o método dos volumes finitos com segunda ordem de precisão no espaço e discretização centrada com dissipação artificial. As soluções estacionárias são obtidas ao se postular um termo tempo-dependente e integra-lo com um esquema Runge-Kutta de 5 passos e 2a ordem de precisão. As simulações são realizadas em malhas não-estruturadas formadas por elementos triangulares em 4 geometrias distintas: um bocal divergente, um perfil diamante, um aerofólio simétrico (NACA 0012) e o outro assimétrico (RAE 2822). Os gradientes adjuntos são então validados por meio da comparação com os obtidos pelo método de diferenças finitas nos regimes de escoamento subsônico, supersônico e transônico. / The adjoint method has been extensively used as an aircraft design tool, since it enables one to obtain sensitivities of many different mesures of merit with respect to parameters that control the aerodynamic surface geometry. This works aims to open up the possibilities of the method\'s applications by using flow physical properties at the permeable boundaries of the computational domain as control parameters of a particular measure of merit. This way it is possible, among many possibilities, to compute lift or drag sensitivities of an aircraft with respect to cruise conditions, for instance. Moreover, this information can be obtained with the same adjoint solution used to perform shape optimization. It is also worth noting that in order to obtain this information from the adjoint equations it is necessary to implement characteristics-based boundary conditions, resolving the complete Riemann problem at the boundaries of the computational domain. The use of the traditional homogeneous boundary conditions, widely spread in the literature, would lead the gradient to vanish. This new approach of the method is, then, applied to flows modeled by the 2-D steady state compressible Euler equations. Both, physical and adjoint problems are numerically solved with a computational code that makes use of a 2nd order finite volume method and central differences with artifficial dissipation. The steady solutions are obtained by postulating a time-dependent term and integrating it with a 5-stage 2nd order Runge-Kutta scheme. The simulations are performed on unstructured triangular meshes to 4 different geometries: a divergent nozzle, a diamond profile, a symmetric airfoil (NACA 0012) and a assymmetric airfoil (RAE 2822). The adjoint gradients are then validated by comparison with those obtained by finite differences method in subsonic, supersonic and transonic flow regimes.

After nearly 50 years of development, Computational Fluid Dynamics (CFD) has become an indispensable component of research, forecasting, design, prototyping and testing for a very broad spectrum of fields including geophysics, and most engineering fields (mechanical, aerospace, biomedical, chemical and civil engineering). The fastest and most affordable CFD approach, called Reynolds-Average-Navier-Stokes (RANS) can predict the drag around a car in just a few minutes of simulation. This feat is possible thanks to simplifying assumptions, semi-empirical models and empirical models that render the flow governing equations solvable at low computational costs. The fidelity of RANS model is good to excellent for the prediction of flow rate in pipes or ducts, drag, and lift of solid objects in Newtonian flows (e.g. air, water). RANS solutions for the prediction of scalar (e.g. temperature, pollutants, combustable chemical species) transport do not generally achieve the same level of fidelity. The main culprit is an assumption, called Reynolds analogy, which assumes analogy between the transport of momentum and scalar. This assumption is found to be somewhat valid in simple flows but fails for flows in complex geometries and/or in complex fluids. This research explores optimization methods to improve upon existing RANS models for scalar transport. Using high fidelity direct numerical simulations (numerical solutions in time and space of the exact transport equations), the most common RANS model is a-priori tested and investigated for the transport of temperature (as a passive scalar) in a turbulent channel flow. This one constant model is then modified to improve the prediction of the temperature distribution profile and the wall heat flux. The resulting modifications provide insights in the model’s missing physics and opens new areas of investigation for the improvement of the modeling of turbulent scalar transport.