Deux points de vue sont souvent opposés dans le cadre de la modélisation des systèmes complexes.D’un côté, une modélisation microscopique cherche à reproduire précisément le comportement des nombreuses entités qui composent le système, ce qui impose des temps de calculs prohibitifs pour le passage à l’échelle de système réels. À l’inverse, l’approche phénoménologique consiste à nous concentrer sur le comportement global du système. Ces modèles macroscopiques reposent sur des lois descriptives qui autorisent des simulations plus rapides mais impliquent l’introduction de paramètres qui peuvent être difficilement identifiables dans le contexte. Pour répondre à ce problème, nous proposons de combiner les différents points de vue de modélisation et d’utiliser des simulations microscopiques pour nourrir un modèle macroscopique incomplet.L’objectif est d’obtenir une simulation descriptive rapide tout en profitant de la précision d’un modèle microscopique. Pour cela, nous proposons une architecture logicielle qui s’appuie sur la technique de la co-simulation pour généraliser la démarche de simulation redondante d’échelles de modélisation hétérogènes.Nous distinguons deux stratégies de co-simulation qui permettent de piloter un modèle macroscopique en cours de simulation. La première consiste à estimer dynamiquement, et de manière explicite, de nouvelles valeurs pour un paramètre critique donné à l’aide d’un simulateur microscopique dédié. La seconde stratégie permet de déterminer implicitement un jeu de paramètres interdépendants sur la base d’une sortie commune des différents niveaux de description simulés. Nous appliquons nos travaux au problème concret du design de structures offshore pour des conditions polaires. Nous détaillons dans un premier temps l’implémentation d’un simulateur phénoménologique d’interactions glace-structure. Dans un second temps, nous illustrons l’intérêt et l’intégration de nos stratégies de co-simulation pour, d’une part améliorer la précision des simulations des phénomènes hydrodynamiques, et d’autre part guider un modèle de plus haut niveau à des fins de prototypage rapide. / There are usually two opposite points of view for the modelling of complex systems. First, microscopical models aim at reproducing precisely the behavior of each entity of the system. In general, their great number is a major obstacle both to simulate the model in a reasonable time and to identify global behaviors. By contrast, the phenomenological approach allows the construction of efficient models from a macroscopic point of view as a superposition of phenomena. A drawback is that we often have to set empirical parameters in these descriptive models. To respond to this problem, we want to make joint use of different levels of description and to use microscopical simulations to feed incomplete macroscopical models.We would then obtain enhanced descriptive simulations with the precision of microscopical models in this way. To this end, we propose a redundant multiscale architecture which is based on the co-simulation methodology in order to generalize the redundant multiscale approach. We suggest two specific co-simulation strategies to guide a macroscopical simulation.The first one consists in dynamically and explicitly estimating critical parameters of a macroscopical model thanks to a dedicated microscopical simulator The second one allows to implicitly determine a full set of dependant parameters on the basis of an output shared by the different levels of description. Then we apply our works to the effective problem of the design offshore structures for arctic conditions. We first describe the implementation of an ice-structure simulation tool by means of a phenomenological and multi-model approach. In a second phase, we show the benefits of our co-simulation strategies to improve the precision of hydrodynamics simulations on the one hand, and on the other to pilot a more macroscopical model for the purpose of fast prototyping.

Made available in DSpace on 2015-03-04T18:50:39Z (GMT). No. of bitstreams: 1 Apresentacao.pdf: 200775 bytes, checksum: 317576b779951158daadb5222c59a464 (MD5) Previous issue date: 2005-12-20 / Coordenacao de Aperfeicoamento de Pessoal de Nivel Superior / Several problems in physics and engineering are modeled by reaction-advection-diffusion (RAD) equations. However, when the diffusive terms are small compared with the other ones, these problems can become difficult to solve numerically. Besides, formulating the unsteady version of these models in a semi-discrete fashion, it can be interpreted that the overall diffusivity gets smaller as the time step decreases. To overcome these drawbacks, this thesis considers the development of Galerkin (or Petrov-Galerkin) finite element methods based on approximation spaces enriched by residual-free bubbles (RFB) or multiscale functions. Beginning with the unsteady reaction-diffusion problem, new methods using multiscale functions are presented which improve the solutions in the reaction-dominated regime and/or when small time steps are adopted. They also give rise to a general concept of stabilizing unsteady problems differently along the time. In the following, it is shown that switching RFB by suitable multiscale functions in the elements connected to the outflow boundaries of the domain increases the accuracy of the solutions in this region for RAD problems with advection. Next, this methodology is further studied for systems of RAD equations. In a final contribution, an extension of the RFB method is introduced for the shallow waters equations. All these methods are tested through benchmark problems and compared with stabilized methods presenting stable and accurate results. / A modelagem de vários problemas físicos e de engenharia envolve a solução de problemas de transporte do tipo reação-advecção-difusão (RAD), porém, estes podem tornar-se singularmente perturbados quando os termos difusivos são pequenos comparados aos demais. Além disso, ao adotar formulações semi-discretas em problemas transientes, observa-se que diminuir o passo de tempo tem um efeito de redução da componente difusiva. Para superar estas dificuldades, esta tese considera o desenvolvimento de métodos de elementos finitos de Galerkin (ou Petrov-Galerkin) baseados em espaços de aproximação enriquecidos por funções bolhas livres do resíduo (RFB) ou funções multiescala. Começando pelo problema de reação-difusão transiente, novos métodos utilizando funções multiescala são apresentados, os quais melhoram as soluções no regime reativo-dominante e/ou quando pequenos passos de tempo são adotados. Com estes métodos, discute-se também o conceito de estabilização variável ao longo do tempo para problemas transientes. Na seqüência, verifica-se que utilizar funções multiescala nos elementos conectados às fronteiras de saída de fluxo do domínio e RFB nos demais elementos aumenta a precisão das soluções nesta região em problemas de RAD com advecção dominante. A seguir, esta metodologia é estudada para sistemas de RAD. Como contribuição final, estende-se o método RFB para o modelo de águas rasas. Todos estes métodos são submetidos a testes de robustez e comparados com métodos estabilizados, apresentando resultados estáveis e precisos.