Abstract : This thesis comes within the scope of mechatronic systems design and deals especially with the problem of their sizing. To that aim, it focuses attention on methods allowing the selection of the components constituting their actuating chains with respect to given requirements. Since current competitive laws require a frequent product renewal in spite of an ever increasing product complexity, the adopted approaches for handling this kind of problem have to be less expensive in financial terms as well as in the duration of the design phase. So as to decrease the number of iterations during the design process, one possible solution is to reformulate the problem as an inverse problem where the unknowns of the sizing problem are directly determined from the requirements expressed on the system outputs. With this in mind, the Ampère laboratory suggests a sizing methodology based on the use of inverse bond graph models. The objective of this thesis is to contribute to the development of this methodology in its structural analysis phase on one hand the phase allowing to check before any numerical simulation if the inverse problem is well-posed) and in its sizing phase on the other hand the phase where the model inversion is applied and exploited). Concerning the structural analysis phase, the dissertation aims at describing the mechanisms of such a analysis in the bond graph language and also at specifying its validity domain. To do that, a comparison between the bond graph, the Modelica, the structured system and the state-space approaches allows to highlight the existence of several information and description levels on the system. Depending on the non exploitation or the exploitation of these levels, several levels of analysis are then suggested: the structured level, the BG-structural level and the behavioral level. It is also shown how these different levels of analysis can be useful in a design process and how they permit to reformulate some bond graph properties according to which design phase we place ourselves the properties expressed at a BG-structural level allows to invalidate or validate the architecture of the system). Concerning the sizing phase, the dissertation tackles the case where the sizing problem can not be entirely reformulated as an inverse problem and so where the methodology can not be directly applied. To that aim, the problem of representing an optimization problem in the bond graph language is dealt with so as to handle requirements which can not be expressed as functions of time. A bond graph-based optimization procedure is thus studied and extended to a class of non linear systems. Finally, an example of a coupling between the sizing methodology and dynamic optimization is implemented till the obtaining of numerical results so as to illustrate the feasibility of the methodology all along the design process.