In the preliminary stage of aircraft development, the fundamental fuselage design is specified as part of an iterative optimisation process. In order to mitigate the need for significant redesign in subse-quent stages, it is essential that all design-driving load cases are incorporated. This is not feasible for the crash load case in current processes, due to the high computational effort of crash simulations. In the project HYFLIP rapid, efficient methods are being developed to enable a crashworthiness evalua-tion in the optimisation loop with the limited available design parameters. In this contribution, the methodology of the accelerated simulation is presented and demonstrated on a model of a fuselage cross-section. As illustrated in Figure 1, the crash model is divided into substructures. This approach allows the mechanical finite element problem to be subdivided into smaller, parameterised problems, for which surrogate models can be developed efficiently. The substructures are interconnected by interfaces, which are represented by circles in Figure 1. To accelerate the solution, the problem is projected onto a latent space using dimensionality reduc-tion. A neural differential equation scheme is implemented in order to solve the mechanical system in time. Therefore, the derivatives of the internal states of the substructure’s individual models, as well as the forces on the interfaces, are predicted by feedforward neural networks. Connectivity is ensured via the interfaces. The time is solely introduced by the time stepping differential equation solver of the workflow shown in Fig. 2, distinguishing it from other contributions, e.g. [1]. Based on the design parameters of the substructures the system can be solved completely data-driven, so no access to a finite element solver is needed in the online stage. Thus, an enormous time saving can be achieved while maintaining adequate accuracy, allowing the crashworthiness analysis to be incorporated in the preliminary design optimisation loop.