A large effort in the analysis of a physical system is the development of a model describing its behavior. The non-linear and time variant characteristic of many mechanical systems can be hardly represented by an analytical model without a remarkable increase of its complexity which contrasts with the need to obtain acceptable results in real-time such as in multibody simulations, system control design and Hardware in the loop (HIL) testing. In this context, the use of artificial neural networks are recognized as a powerful modeling tool to produce accurate model with reduced complexity. On the other hand their response to inputs outside the learning range may lead to unrealistic results. This paper presents an hybrid modeling technique, which combines a physical model with a neural network. The physical model describes the gross behavior of the system and the neural network captures the non-linear non-modeled behaviors or the effect of time-varying parameters. It is also proposed a method to limit the outside-range unpredicted responses. A RC car shock absorber is used as test case. Experimental results show that the neural network improves the physical model output capturing nonlinear aspects such as the hysteresis, the fluid leakage and the increase of its temperature. Copyright © 2011 by ASME.