Nowadays there is an increasing industrial interest in the use of modelling and simulation tools. To cope with the growing demands for simulation models of ever increasing complex industrial systems, the research community effort has been mainly focused in creating different software tools which simplify the modelling task. The need of modelling tools supporting libraries of non-causal models has been widely recognized. These predefined models can be coupled in the same way as the physical units are assembled in the system.The Object-Oriented Modelling paradigm emerges as a methodology based on the main characteristics of Object Oriented Programming (modularity, inheritance, classes, abstracts interfaces, etc) to support reusability of code. The approach adopted by present OOM languages as Omola, Dymola and, recently, Modelica relies on models defined as classes where behaviour is encapsulated using a-causal equations. Model abstract interfaces are defined by the variables shared by the model and its environment. These techniques did not received much attention from the simulation market. This paper analyses the limitations of these mathematical models as class structures and the constraints which they introduce in model reusability.The paper outlines the need of real modular software structures which are capable to support the system behaviour description independently from its context of reuse. This work presents an OOM language, PML, which makes a separation between the physical behaviour representation (phenomena and laws), the system component representation (models for which appropriate phenomena are specified) and the system component operation object (entities as material, energy or forces). The aim of this modularisation of physical knowledge is to shield the model user, who is not concerned with the underlying computational model required for simulation engines.