In a hybrid simulation, a structure is divided into numerical and physical substructures in order to obtain more accurate responses in comparison to a full computational analysis. It is more computationally efficient to test complicated parts physically and model the remaining parts using a standard finite-element program. However, due to a lack of test facilities (for example due to specimen size, capacity of the hydraulic power pack, limitations of reaction frames and walls, and the number and capacity of actuators) and budget limitations, only a few substructures can be tested experimentally and the others have to be modelled numerically. Here, a new framework for hybrid simulation is introduced, which uses well-trained neural networks instead of physical substructures. This new framework is called a neuro-hybrid simulation (NHS). With the aim of overcoming the limitations of hybrid simulations (the need for numerous substructures, errors arising from the test setup, data acquisition, simultaneous interaction of experimental and numerical parts and so on), an initially trained Prandtl neural network is used to act as a virtual physical substructure. The proposed NHS was verified with some numerical examples, based on which the capability and accuracy of the framework was proven.