Abstract: Constructing an appropriate model of a physical device is crucial in reasoning successfully about its behavior for purposes of simulation, diagnosis and design. In the compositional modeling approach a system is provided with a library of composible pieces of knowledge about the physical world called model fragments. The system's task is to select appropriate model fragments to describe the situation, either for static analysis of a single state of the system, or for the more complicated case simulation of dynamic behavior over a sequence of states. The selected model fragments must be adequate to solve the query, consistent with each other and be as simple as possible. This paper presents an algorithm, based on relevance reasoning, for selecting model fragments efficiently for the case of simulation. We show that the algorithm produces an adequate model for a given query and moreover, it is the simplest one given the constraints in the query.