Constructing models of physical systems is a recurring activity in engineering problem solving. This paper presents a generic knowledge-level analysis of the task of engineering modelling. Starting from the premise that modelling is a design-like activity, it proposes the Specify-Construct-Assess (SCA) problem-solving method for decomposition of the modelling task. A second structuring principle is found in the distinction between and separation of different ontological viewpoints. Here, we introduce three engineering ontologies that have their own specific roles and methods in the modelling task: functional components, physical processes, mathematical constraints. The combination of the proposed task and ontology decompositions leads to a particular approach to modelling that we call evolutionary modelling. This approach is supported by a knowledge-based system called QuBA. The implications of evolutionary modelling for structuring the modelling process, the content of produced models, as well as for the organization of reusable model fragment libraries are discussed.