The objective of this thesis is to design domain decomposition methods which are robust even for hard problems that arise when simulating industrial or real life objects. For instance one particular challenge which the company Michelin is faced with is the fact that tires are made of rubber and steel which are two materials with very different behavior laws. With classical domain decomposition methods, as soon as the partition into subdomains does not accommodate the discontinuities between the different materials convergence deteriorates. For three popular domain decomposition methods (Ad- ditive Schwarz, FETI and BDD) we have proved that by solving a generalized eigenvalue problem in each of the subdomains we can identify automatically which are the modes responsible for slow convergence. In other words we can divide the original problem into two problems : the first one where we can guarantee that the domain decomposition method will converge quickly and the second where we cannot. The final idea is to apply projections to solve these two problems independently (this is also known as deflation) : on the first we apply the domain decomposition method and on the second (we call it the coarse space) we use a direct solver which we know will be robust. We guarantee theoretically that the resulting two level solver is robust. The other main feature of our algorithms is that they can be implemented as black box solvers meaning that heterogeneous materials is only one type of difficulty that they can identify and circumvent.
