Abstract
Computer programs which do any task which requires reasoning about physical systems need to use models of those systems with varying accuracy/complexity traders. This paper describes an approach to model generation in the domain of heat transfer which is capable of producing models that vary greatly along this dimension. This approach is based on the law of conservation of energy, which provides a set of choices for the models in terms of control volumes" and heat flows. These choices are made by using rules of thumb, which can be seen as instances of two reduction operators, delta-iso, and dominance. Various rough models are used to estimate the physical parameters on which these rules depend. That is, the rough models are evaluated in the process of building more accurate ones. The application of these operators is only valid for a specific set of physically meaningful quantities; thus we are really reasoning about physics, not equations. This method has been implemented in a running system, MSG.