The arrival of new Machine Learnings (ML) architectures such as Large Language Models (LLMs) and new GPU hardware have reignited the push for data-driven problem solving. In the field of engineering one of the most elusive research domains is the development of fast and accurate digital twins. Development of faithful dynamics simulations is valuable for the acceleration of design optimization processes and sparing of time and material during real-world testing. Neural Networks (NNs) hold the potential to facilitate this by using measurement data to capture the dynamics of a mechanical system and extrapolate beyond the training domain. However, as with any Reduced Order Modelling (ROM) technique, the key metrics are the upfront offline cost of generating the model and the online prediction cost using the model. A technique which requires too much data or time to generate the model may not be worth the achieved reduction. Additionally, experimental hardware limitations may restrict the amount or domain of reference training data which can be collected. Improving the ability of NN-based mechanical system modelling to perform accurately despite restrictions on input data is paramount for the future adoption of ML in engineering. Previous work has shown that NNs can predict non-linear or even chaotic multi-body system dynamics by replacing the Equations of Motion (EoMs) in a classical integration scheme, or by replacing the entire time-stepping scheme [1]. Further research has demonstrated that through dedicated tuning of hyper-parameters and specialized training procedures, extrapolation beyond the training domain by multiple orders of magnitude can be achieved [2]. The work presented here aims to go another step further by utilizing multiple “expert” NNs, each specialized on a unique domain of the input training data. Through this specialization of the NNs they are individually able to perform better on their respective domain of expertise. The challenge then lies in automatically developing an assignment policy which determines when the use of which expert provides the best results to extrapolate beyond the training domains. Using the example of a chained duffing oscillator, this work presents methods for the selection of training domains and assignment policy generators which result in an improvement of prediction accuracy.