Category: Internal Seminar

Théophile Chaumont: Monday 18 September at 1 pm, A415 Inria Paris. Time-harmonic wave propagation problems are costly to solve numerically since the corresponding PDE operators are not strongly elliptic, and as a result, discretization methods might become unstable. Specifically, the finite element solution is quasi-optimal (almost as good as the best approximation the finite element space can provide) only under restrictive assumptions on the mesh size. If the mesh size is too large, stability is lost, and the finite element solution can become completely inaccurate, even when the best approximation is. This phenomenon is called the “pollution effect” and becomes more important for larger frequencies. For the case of wave propagation problems in homogeneous media, it is known that high order finite element methods are less sensitive to the pollution effect. For this reason, they are employed in a wide range of applications, as the corresponding linear systems are smaller and easier to solve. In this talk, we investigate the use of high order finite element methods to solve wave propagation problems in highly heterogeneous media. Since the heterogeneities of the medium can exhibit small scale features, we consider “non-fitting” meshes, that are not aligned with the physical interfaces of the medium. Instead, the parameters defining the medium of propagation can be discontinuous inside each element. We propose a convergence analysis and draw two main conclusions: – the asymptotic convergence rate of the proposed finite element method is suboptimal due to the lack of regularity of the solution inside each cell – the pollution effect is greatly reduced by increasing the order of discretization. We illustrate our main conclusions with geophysical application benchmarks. These examples confirm that higher order methods are more efficient than linear finite elements.

Continue reading

Valentine Rey: Thursday 22 June at 3 pm, A415 Inria Paris. Domain decomposition methods are robust and efficient methods to solve mechanical problems with several million degrees of freedom. Taking advantage of increasing performances of computers, they exploit the clusters-parallel architecture and are numerically scalable. Verification has been widely developed since 1980’s and proposes tools to estimate the distance between the unknown exact solution of continuous problem and the computed solution. Among those techniques, estimators based on error in constitutive relation provide constant-free upper bounds and are available for varied range of problems. In this talk, we present techniques for steering parallel computation by objective of accuracy on quantities of interest. It relies on a parallel error estimator that provides strict guaranteed upper bound and separates the algebraic error (due to the use of iterative solver) from the discretization error (due to the finite element method). This estimator enables to adapt the solver’s stopping criterion to the discretization, which avoids over resolution and useless iterations. In [*], the estimator is used for goal-oriented error estimation and classical bounds for quantities of interest are rewritten in order to separate sources of error. Finally, we benefit the information provided by the error estimator and the Krylov subspaces built during the resolution to set an auto-adaptive strategy (adaptive remeshing and recycling search directions). *V. Rey, P. Gosselet, C. Rey, Strict bounding of quantities of interest in computations based on domain decomposition, Computer Methods in Applied Mechanics and Engineering. 2015 Apr 15;287:212-28

Continue reading

Gouranga Mallik: Thursday 29 June at 3pm, A415 Inria Paris. In this work we consider a priori and a posteriori error control for the nonsingular solution of von Karman plate bending problem. Conforming and nonconforming finite element methods are employed. Existence, uniqueness and error estimates for the discrete solution are presented. We discuss an abstract framework for a posteriori error control which includes conforming and nonconforming finite element methods. This allows us to compute reliable and efficient local estimators. The key ingredients in establishing well-posedness of the discrete problem rely on the linearization of the continuous problem and suitable enrichment operator. Numerical experiments are performed to justify the theoretical results.

Continue reading

Patrik Daniel: Thursday 15 June at 3:30pm, A415 Inria Paris. We propose a new practical adaptive refinement strategy for hp-finite element approximations of elliptic problems. Following some recent theoretical developments in polynomial-degree-robust a posteriori error analysis, we solve two complementary classes of discrete local problems on the vertex-based patches. The first class involves the solution on each patch of a mixed finite element problem with homogeneous Neumann boundary conditions, which leads to an H(div,Ω)-conforming equilibrated flux. This in turns yields a guaranteed upper bound on the error and serves to mark elements for refinement via a Dörfler bulk criterion. The second class of local problems involves the solution, on each marked patch only, of two separate primal finite element problems with homogeneous Dirichlet boundary conditions, which serve to decide between h-, p-, or hp-refinement. Altogether, we show that these ingredients lead to a computable error reduction factor; we guarantee that while performing the hp-adaptive refinement as suggested, the error will be reduced at least by this factor on the next hp-mesh. In a series of numerical experiments in two space dimensions, we first study the accuracy of our predicted reduction factor: in particular, we measure the ratio of the predicted reduction factor relative to the true error reduction, and we find that it is very close to the optimal value of one for both smooth and singular exact solutions. Finally, we study the overall performance of the proposed hp-refinement strategy on some test cases, for which we observe effectivity indices very close to one and exponential convergence rates.

Continue reading

Jad Dabaghi: Thursday 15 June at 3pm, A415 Inria Paris. We propose an adaptive inexact version of a class of semi-smooth Newton methods. As a model problem, we consider the system of variational inequalities describing the contact between two membranes and its finite element discretization. Any iterative linearization algorithm like the Newton-min, Newton-Fisher Burmeister is taken into account, as well as any iterative linear algebraic solver. We prove an a posteriori error estimate between the exact solution and the approximate solution which is valid on any step of the linearization and algebraic resolution. This estimate is based on discretization and algebraic flux reconstructions, where the latter one is obtained on a hierarchy of nested meshes. The estimate distinguishes the discretization, linearization, and algebraic components of the error and allows us to formulate adaptive stopping criteria for both solvers. Numerical experiments for the semi-smooth Newton-min algorithm in combination with the GMRES solver confirm the efficiency of the method.

Continue reading

Martin Eigel: Thursday 1 June at 10:45 am, A415 Inria Paris. We consider a class of linear PDEs with stochastic coefficients which depend on a countable (inifinite) number of random parameters. As an alternative to classical Monte Carlo sampling techniques, a functional discretisation of the stochastic space in generalised polynomial chaos may lead to significantly improved (optimal) convergence rates. However, when employed in the context of Galerkin methods, the arising algebraic systems are very high-dimensional and quickly become intractable to computations. As a matter of fact, this is an exemplary example for the curse of dimensionality with exponential growth of complexity which makes model reduction techniques inevitable. In the first part, we discuss two approaches for this: (1) a posteriori adaptivity and exploitation of sparsity of the solution, and (2) low-rank compression in a hierarchical tensor format. In the second part, the low-rank discretisation is used as an efficient representation of the stochastic model for Bayesian inversion. This is an important application in Uncertainty Quantification where one is interested in determining the (statistics of) parameters of the model based on a set of noisy measurements. In contrast to popular sampling techniques such as MCMC, we derive an explicit representation of the posterior densities. The examined sampling-free Bayesian inversion is adaptive in all discretisation parameters. Moreover, convergence of the method is shown.

Continue reading

Joscha Gedicke: Thursday 1 June at 10 am, A415 Inria Paris. We extend the Hodge decomposition approach for the cavity problem of two-dimensional time harmonic Maxwell’s equations to include the impedance boundary condition, with anisotropic electric permittivity and sign changing magnetic permeability. We derive error estimates for a P_1 finite element method based on the Hodge decomposition approach and develop a residual type a posteriori error estimator. We show that adaptive mesh refinement leads empirically to smaller errors than uniform mesh refinement for numerical experiments that involve metamaterials and electromagnetic cloaking. The well-posedness of the cavity problem when both electric permittivity and magnetic permeability can change sign is also discussed and verified for the numerical approximation of a flat lens experiment.

Continue reading

Quang Duc Bui: Thursday 1 June at 11:30am, A415 Inria Paris. Parareal method is a numerical method to solve time-evolution problems in parallel, which uses two propagators: the coarse – fast and inaccurate – and the fine – slow but more accurate. Instead of running the fine propagator on the whole time interval, we divide the time space into small time intervals, where we can run the fine propagator in parallel to obtain the desired solution, with the help of the coarse propagator and through parareal steps. Furthermore, each local subproblem can be solved by an iterative method, and instead of doing this local iterative method until convergence, one may perform only a few iterations of it, during parareal iterations. Propagators then become much cheaper but sharply lose their accuracy, and we hope that the convergence will be achieved across parareal iterations. Here, we propose to couple Parareal with a well-known iterative method – Schwarz Waveform Relaxation (SWR)- with only few SWR iterations in the fine propagator and with a simple coarse propagator deduced from Backward Euler method. We present the analysis of this coupled method for 1-dimensional advection reaction diffusion equation, for this case the convergence is at least linear. We also give some numerical illustrations for 1D and 2D parabolic equations, which shows that the convergence is much faster in practice.

Continue reading

Quanling Deng: Tuesday 16 May at 3 pm, A415 Inria Paris. The isogeometric analysis (IgA) is a powerful numerical tool that unifies the finite element analysis (FEA) and computer-aided design (CAD). Under the framework of FEA, IgA uses as basis functions those employed in CAD, which are capable of exactly represent various complex geometries. These basis functions are called the B-Splines or more generally the Non-Uniform Rational B-Splines (NURBS) and they lead to an approximation which may have global continuity of order up to $p-1$, where $p$ is the order of the underlying polynomial, which in return delivers more robustness and higher accuracy than that of finite elements. We apply IgA to wave propagation and structural vibration problems to study their dispersion and spectrum properties. The dispersion and spectrum analysis are unified in the form of a Taylor expansion for eigenvalue errors. By blending optimally two standard Gaussian quadrature schemes for the integrals corresponding to the stiffness and mass, the dispersion error of IgA is minimized. The blending schemes yield two extra orders of convergence (superconvergence) in the eigenvalue errors, while the eigenfunction errors are of optimal convergence order. To analyze the eigenvalue and eigenfunction errors, the Pythagorean eigenvalue theorem (Strang and Fix, 1973) is generalized to establish an equality among the eigenvalue, eigenfunction (in L2 and energy norms), and quadrature errors.

Continue reading

Sarah Ali Hassan: Thursday 11 May at 3 pm, A415 Inria Paris. In this work we develop a posteriori error estimates and stopping criteria for domain decomposition (DD) methods with optimized Robin transmission conditions on the interface. Steady diffusion equation using the mixed finite element (MFE) discretization as well as in the heat equation using the MFE method in space and the discontinuous Galerkin scheme in time are analysed. For the heat equation, a global-in-time domain decomposition method is used, allowing for different time steps in different subdomains. We bound the error between the exact solution of the PDE and the approximate numerical solution at each iteration of the domain decomposition algorithm. Different error components (domain decomposition, space discretization, time discretization) are distinguished, which allows us to define efficient stopping criteria for the DD algorithm. The estimates are based on the reconstruction techniques for pressures and fluxes. Numerical experiments illustrate the theoretical findings.

Continue reading