Mechanical systems have been traditionally represented using parametric physics-based models. In this work, we introduce a novel concept, in this part of the mechanical system is represented using data-based subsystem models, and the overall mechanical system model is composed of these data-based and other, physics-based subsystems. A core element is the interfacing of the subsystems, which gives rise to interaction forces. The interfacing problem is formulated in a way that makes it possible to give a general representation to the interaction forces. We demonstrate that from the point of view of the physics-based subsystems the important element is that the data-based models can represent the interaction force systems properly. The data-based subsystems are developed using deep recurrent neural networks, and the training data is generated based on simulations using the fully parametric physics-based model of the system. Such training data could also be obtained through physical experimentation.