Numerical models of PDE systems can involve very large matrix equations, but feedback controllers for these systems must be computable in real time to be implemented on physical systems. Classical control design methods produce controllers of the same order as the numerical models. Therefore, reduced order control design is vital for practical controllers. The main contribution of this research is a method of control order reduction that uses a newly developed low order basis. The low order basis is obtained by applying Proper Orthogonal Decomposition (POD) to a set of functional gains, and is referred to as the functional gain POD basis. Low order controllers resulting from the functional gain POD basis are compared with low order controllers resulting from more commonly used time snapshot POD bases, with the two dimensional heat equation as a test problem. The functional gain POD basis avoids subjective criteria associated with the time snapshot POD basis and provides an equally effective low order controller with larger stability radii. An efficient and effective methodology is introduced for using a low order basis in reduced order compensator design. This method combines "design-then-reduce" and "reduce-then-design" philosophies. The desirable qualities of the resulting reduced order compensator are verified by application to Burgers' equation in numerical experiments. / Ph. D.

In this dissertation we discuss bilinear immersed finite elements (IFE) for solving interface problems. The related research works can be categorized into three aspects: (1) the construction of the bilinear immersed finite element spaces; (2) numerical methods based on these IFE spaces for solving interface problems; and (3) the corresponding error analysis. All of these together form a solid foundation for the bilinear IFEs. The research on immersed finite elements is motivated by many real world applications, in which a simulation domain is often formed by several materials separated from each other by curves or surfaces while a mesh independent of interface instead of a body-fitting mesh is preferred. The bilinear IFE spaces are nonconforming finite element spaces and the mesh can be independent of interface. The error estimates for the interpolation of a Sobolev function in a bilinear IFE space indicate that this space has the usual approximation capability expected from bilinear polynomials, which is <i>O</i>(<i>h</i>²) in <i>L</i>² norm and <i>O</i>(<i>h</i>) in <i>H</i>¹ norm. Then the immersed spaces are applied in Galerkin, finite volume element (FVE) and discontinuous Galerkin (DG) methods for solving interface problems. Numerical examples show that these methods based on the bilinear IFE spaces have the same optimal convergence rates as those based on the standard bilinear finite element for solutions with certain smoothness. For the symmetric selective immersed discontinuous Galerkin method based on bilinear IFE, we have established its optimal convergence rate. For the Galerkin method based on bilinear IFE, we have also established its convergence. One of the important advantages of the discontinuous Galerkin method is its flexibility for both <i>p</i> and <i>h</i> mesh refinement. Because IFEs can use a mesh independent of interface, such as a structured mesh, the combination of a DG method and IFEs allows a flexible adaptive mesh independent of interface to be used for solving interface problems. That is, a mesh independent of interface can be refined wherever needed, such as around the interface and the singular source. We also develop an efficient selective immersed discontinuous Galerkin method. It uses the sophisticated discontinuous Galerkin formulation only around the locations needed, but uses the simpler Galerkin formulation everywhere else. This selective formulation leads to an algebraic system with far less unknowns than the immersed DG method without scarifying the accuracy; hence it is far more efficient than the conventional discontinuous Galerkin formulations. / Ph. D.

Protection from the potentially damaging effects of shock loading is a common design requirement for diverse mechanical structures ranging from shock accelerometers to spacecraft. High-damping viscoelastic materials are employed in the design of geometrically complex impact absorbent components. Since shock transients have a broadband frequency spectrum, it is imperative to properly model frequency dependence of material parameters. The Anelastic Displacement Fields (ADF) method is employed to develop new axisymmetric and plane stress finite elements that are capable of modeling frequency dependent material behavior of linear viscoelastic materials. The new finite elements are used to model and analyze behavior of viscoelastic structures subjected to shock loads. The development of such ADF-based finite element models offers an attractive analytical tool to aid in the design of shock absorbent mechanical filters. This work will also show that it is possible to determine material properties’ frequency dependence by iteratively fitting ADF model predictions to experimental results. A series of experiments designed to validate the axisymmetric and plane stress finite element models are performed. These experiments involve the propagation of longitudinal waves through elastic and viscoelastic rods, and behavior of elastomeric mechanical filters subjected to shock. Comparison of model predictions to theory and experiments confirm that ADF-based finite element models are capable of capturing phenomena such as geometric dispersion and viscoelastic attenuation of longitudinal waves in rods as well as modeling the behavior of mechanical filters subjected to shock. / Ph. D.

A class of immersed finite element (IFE) spaces is developed for solving elliptic boundary value problems that have interfaces. IFE spaces are finite element approximation spaces which are based upon meshes that can be independent of interfaces in the domain. Three different quadratic IFE spaces and their related biquadratic IFE spaces are introduced here for the purposes of solving both forward and inverse elliptic interface problems in 1D and 2D. These different spaces are constructed by (i) using a hierarchical approach, (ii) imposing extra continuity requirements or (iii) using a local refinement technique. The interpolation properties of each space are tested against appropriate testing functions in 1D and 2D. The IFE spaces are also used to approximate the solution of a forward elliptic interface problem using the Galerkin finite element method and the mixed least squares finite element method. Finally, one appropriate space is selected to solve an inverse interface problem using either an output least squares approach or the least squares with mixed equation error method. / Ph. D.

We study the behavior of numerical stabilization schemes in the context of linear quadratic regulator (LQR) control problems for convection diffusion equations. The motivation for this effort comes from the observation that when linearization is applied to fluid flow control problems the resulting equations have the form of a convection diffusion equation. This effort is focused on the specific problem of computing the feedback functional gains that are the kernels of the feedback operators defined by solutions of operator Riccati equations. We develop a stabilization scheme based on the Galerkin Least Squares (GLS) method and compare this scheme to the standard Galerkin finite element method. We use cubic B-splines in order to keep the higher order terms that occur in GLS formulation. We conduct a careful numerical investigation into the convergence and accuracy of the functional gains computed using stabilization. We also conduct numerical studies of the role that the stabilization parameter plays in this convergence. Overall, we discovered that stabilization produces much better approximations to the functional gains on coarse meshes than the unstabilized method and that adjustments in the stabilization parameter greatly effects the accuracy and convergence rates. We discovered that the optimal stabilization parameter for simulation and steady state analysis is not necessarily optimal for solving the Riccati equation that defines the functional gains. Finally, we suggest that the stabilized GLS method might provide good initial values for iterative schemes on coarse meshes. / Ph. D.

Nine independent and four dependent variables are used to build a response surface to calculate strip crown using the difference in the industry standard strip height measurements. The single element response surface in use provides the advantages of continuous derivatives and decouples rolling load from the determination of exit height. The data points to build the response surface are the product of a calibrated finite element model. The rolling dynamics in the finite element model creates a transient that requires nonlinear regression to find the system steady-state values. Weighted-least squares is used to build a response surface using isoparametric interpolation with the non-rectangular domain of the mill stands represented as a single element. The regression statistics, the 1-D projections, comparisons against other response surface models and the comparisons against an existing strip crown model are part the validation of the response surface generated. A four-high mill stand is modeled as a quarter-symmetry 3-D finite element model with an elastic-plastic material model. A comparison of the pressure distribution under the arc of contact with existing research supports the pressure distribution found with experiments conducted by Siebel and Lueg [16] and it also suggests the need for one improvement in the initial velocity for the strip in the finite element model. The strip exit heights show more sensitivity to change than strip exit crown in seven out of the nine independent variables, so a response surface built with the strip exit height is statistically superior to using the derived dependent variable strip exit crown. Sensitivity of strip exit crown and the strip exit heights to changes in work-roll crown are about equal. Backup-roll diameter sensitivity is small enough that oversampling for the mean trend has to be considered or ignore backup-roll altogether. Strip entry velocity is a new independent variable, unless the response surface is built from the derived variable, strip exit crown. A problem found is that the sensitivity of strip entry crown and work-roll crown requires a larger than typical incremental change to get a reliable measure of the change strip exit crown. A narrow choice of high and low strip entry crowns limits the usefulness of the final response surface. A recommendation is to consider the use of the strip cross-section as an exit crown predictor. / Master of Science

An integrated procedure, combining finite element modeling and fatigue analysis methods, is developed and applied to the fatigue optimization of a notched, induction hardened, steel shaft subjected to combined bending and torsional loading. Finite element analysis is used first to develop unit-load factors for generating stress-time histories, and then, employing thermo-elastic techniques, to determine the residual stresses resulting from induction hardening. These stress fields are combined using elastic superposition, and incorporated in a fatigue analysis procedure to predict failure location and lifetime. Through systematic variation of geometry, processing, and loading parameters, performance surfaces are generated from which optimum case depths for maximizing shaft fatigue performance are determined. General implications of such procedures to the product development process are discussed. / Master of Science

High-fidelity computer simulations of childbirth are time consuming, making them impractical for guiding decision-making during obstetric emergencies. The complex geometry, micro-structure, and large finite deformations undergone by the vagina during childbirth result in material and geometric nonlinearities, complicated boundary conditions, and nonhomogeneities within finite element (FE) simulations. Such nonlinearities pose a significant challenge for numerical solvers, increasing the computational time. Simplifying assumptions can reduce the computational time significantly, but this usually comes at the expense of simulation accuracy. The work herein proposed the use of reduced order modeling (ROM) techniques to create surrogate models that capture experimentally-measured displacement fields of rat vaginal tissue during inflation testing in order to attain both the accuracy of higher-fidelity models and the speed of lower-fidelity simulations. The proper orthogonal decomposition (POD) method was used to extract the significant information from FE simulations generated by varying the luminal pressure and the parameters that introduce the anisotropy in the selected constitutive model. In our first study, a new data-driven (DD) variational multiscale (VMS) ROM framework was extended to obtain the displacement fields of rat vaginal tissue subjected to ramping luminal pressure. For comparison purposes, we also investigated the classical Galerkin ROM (G-ROM). In our numerical study, both the G-ROM and the DD-VMS-ROM decreased the FE computational cost by orders of magnitude without a significant decrease in numerical accuracy. Furthermore, the DD-VMS-ROM improved the G-ROM accuracy at a modest computational overhead. Our numerical investigation showed that ROM had the potential to provide efficient and accurate computational tools to describe vaginal deformations, with the ultimate goal of improving maternal health. Our second study compared two common computational strategies for surrogate modeling, physics-based G-ROM and data-driven machine learning (ML), for decreasing the cost of FE simulations of the ex vivo deformations of rat vaginal tissue subjected to inflation testing to study the effect of a pre-imposed tear. Since there are many methods associated with each modeling approach, to provide a fair and natural comparison, we selected a basic model from each category. From the ROM strategies, we considered a simplified G-ROM that is based on the linearization of the underlying nonlinear FE equations. From the ML strategies, we selected a feed-forward dense neural network (DNN) to create mappings from constitutive model parameters and luminal pressure values to either the FE displacement history (in which case we denote the resulting model ML) or the POD coefficients of the displacement history (in which case we denote the resulting model POD-ML). The numerical comparisons of G-ROM, ML, and POD-ML took place in the reconstructive regime. The numerical results showed that the G-ROM outperformed the ML model in terms of offline central processing unit (CPU) time for model training, online CPU time required to generate approximations, and relative error with respect to the FE models. The POD-ML model improved on the speed performance of the ML, having online CPU times comparable to those of the G-ROM given the same size of POD bases. However, the POD-ML model did not improve on the error performance of the ML. In our last study, we expanded our investigation of ML methods for surrogate modeling by comparing the performance of a DNN similar to what was used previously to that of a convolutional neural network (CNN) using 1-D convolution on the input parameters from FE simulations of active vaginal tearing. The new FE simulations utilized a custom continuum damage model that provided material damage and failure properties to an existing anisotropic hyperelastic constitutive model to replicate experimentally-observed tear propagation behaviors. We employed our DNN and CNN models to create mappings from constitutive model parameters, geometric properties of the propagating tear, and luminal pressure values to either the full FE displacement history or the POD coefficients of the displacement history. The root-mean-square error (RMSE) with respect to the FE displacement history achieved by full order output ML predictions was reproducible with POD-ML using a basis of only dimension l=10. Additionally, an order of magnitude reduction in offline time was observed using POD-ML over full-order ML with minimal difference between DNN and CNN architectures. Differences in online computational costs between ML and POD-ML were found to be negligible, but the DNNs produced predictions slightly faster than the CNNs, though both online times were on the same order of magnitude. While convolution did not significantly aid the regression task at hand, POD-ML was demonstrated to be an efficient and effective approach for surrogate modeling of the FE tear propagation model, approximating the displacement history with RMSE less than 0.1 mm and generating results 7 orders of magnitude faster than the FE model. This set of baseline numerical investigations serves as a starting point for future computer simulations that consider state-of-the-art G-ROM and ML strategies, and the in vivo geometry, boundary conditions, material properties, and tissue damage mechanics of the human vagina, as well as their changes during labor. / Doctor of Philosophy / Computer simulations of childbirth are extremely time-consuming, making them impractical for guiding decision-making by obstetricians when a patient is entering labor. The complex geometry, material microstructure, and large deformations undergone by the vagina during childbirth result in material and geometric properties that are challenging to mathematically model. Consequently, numerical solver methods (e.g., finite elements) require large amounts of time to simulate childbirth. Simplifying assumptions can reduce computational time, but this simplification usually comes at the expense of simulation accuracy. The work of this dissertation proposes the use of several techniques to reduce model complexity and create accurate approximations and predictions of results from full-order models (FOMs) with profound reductions in computational time. Our first study used reduced order models (ROMs) to extract the significant information from a FOM of the rat vagina subjected to inflation. We compared a basic ROM and an advanced, data-driven ROM. Our second study compared the basic ROM to a basic machine learning (ML) technique for approximating a FOM that simulated inflation of the rat vagina with a pre-imposed tear. A hybrid technique incorporating elements of both ROM and ML to approximate FOM results was also considered. Our final study made use of ML and hybrid techniques using a more advanced neural network (a convolutional neural network). These ML models were used to predict the results of a FOM simulation of vaginal tear propagation. These numerical investigations serve as a starting point for future development of computer simulations using state-of-the-art ROM and ML strategies as well as more realistic models for the mechanics of the human vagina during childbirth.

In der vorliegenden Arbeit wird ein gemischtes Element gegen Locking-Effekte untersucht. Dazu wird ein Fünf-Feld-Hu-Washizu-Funktional (CoFEM-Element) für lineare und quadratische Hexaeder-Elemente unter einer hyperelastischen, isotropen, polykonvexen sowie einer transversal-isotropen Materialformulierung implementiert. Die resultierenden nichtlinearen Gleichungen werden mithilfe eines Mehrebenen-NEWTON-RAPHSON-Verfahren unter Beachtung einer konsistenten Linearisierung gelöst. Als repräsentatives Beispiel der numerischen Untersuchungen dient der einseitig eingespannte Cook-Balken mit einer quadratischen Druckverteilung am Rand. Zur Beurteilung des CoFEM-Elements wird das räumliche Konvergenzverhalten für unterschiedliche Polynomgrade und für verschiedene Netze unter Beachtung der algorithmischen Effizienz untersucht. / This paper presents a mixed finite element formulation of Hu-Washizu type (CoFEM) designed to reduce locking effects with respect to a linear and quadratic approximation in space. We consider a hyperelastic, isotropic, polyconvex material formulation as well as transverse isotropy. The resulting nonlinear algebraic equations are solved with a multilevel NEWTON-RAPHSON method. As a numerical example serves a cook-like cantilever beam with a quadratic distribution of in-plane load on the Neumann boundary. We analyze the spatial convergence with respect to the polynomial degree of the underlying Lagrange polynomials and with respect to the level of mesh refinement in terms of algorithmic efficiency.

Cette thèse s'intéresse à la génération, à l'utilisation et à l'évaluation des maillages hex-dominants, composés d'hexaèdres et de tétraèdres, dans la cadre de la simulation numérique par la méthode des éléments finis. Les éléments finis hexaédriques sont souvent préférés aux éléments tétraédriques car ils offrent un meilleur ratio entre précision et temps de calcul dans un certain nombre de situations. Cependant, si la génération automatique de maillages tétraédriques est aujourd'hui un domaine bien maîtrisé, ce n'est pas le cas de la génération de maillages hexaédriques alignés avec le bord, qui reste un problème largement ouvert. En l'absence de progrès significatifs, les approches actuelles se contentent de maillages hex-dominants afin de tirer parti des performances supérieures des hexaèdres et de la flexibilité géométrique des tétraèdres, qui rend possible le maillage automatique. Dans une première partie, nous développons des algorithmes robustes pour la génération de maillages hex-dominants à partir de champs de directions, notamment pour l'isolement et le remplissage des régions difficiles à mailler (singularités et autres dégénérescences). Dans la seconde partie, nous essayons de déterminer dans quelles situations et dans quelle mesure les maillages hexaédriques, et hex-dominants générés précédemment, sont plus intéressants que les maillages tétraédriques. Ceci implique spécifiquement d'étudier plusieurs manières d'effectuer des simulations par éléments finis avec les maillages hybrides, dont une approche où nous utilisons des contraintes de continuité pour maillages non-conformes. Pour mesurer l'influence du maillage sur l'approximation des solutions, nous proposons une nouvelle méthode d'échantillonnage pour calculer très efficacement des distances globales entre solutions éléments finis définies sur des domaines compliqués / This thesis focuses on generation, usage and evaluation of hex-dominant meshes, which are made of hexaehedra and tetrahedra, in the context of the finite element method. Hexahedron finite elements are often preferred to tetrahedron elements because they offer a better compromise between accuracy and computation time in certain situations. However, if tetrahedral meshing is a well mastered subject, it is not the case of hexahedral meshing. Generating hexahedral meshes with elements aligned to the borders is still an open and difficult problem. Meanwhile, current automated approaches can use hex-dominant meshes in order to take advantage of both hexahedron accuracy and geometrical flexibility of tetrahedra. In the first part, we develop robust algorithms for the generation of hex-dominant meshes with elements aligned with the borders. Specifically, we propose a method to extract and fill the areas where hexahedral meshing is difficult (singularities and degeneracies). In the second part, we try to identify and to quantify the advantages of hexahedral and hex-dominant meshes over tetrehedral ones. This requires to study various ways to apply the finite element method on hybrid meshes, including one in which we propose to use continuity constraints on hexahedral-tetrahedral non-conforming meshes. To measure the impact of meshes on the finite element accuracy, we develop a new sampling method which allows to compute efficiently global distances between finite element solutions defined on complicated 3D domains