High-fidelity Digital Twins provide valuable insights into the behaviour of physical systems, and rely on accurate numerical methods. Digitalisation and physical simulations are becoming increasingly important for creating digital twins that accurately mirror real-world products. However, industrial CAD designs that undergo numerical simulation are constructed with a manufacturing-centric approach. As a result, these models often include features that hinder simulation objectives, e.g., by increasing simulation time without improving fidelity. Systematic removal of these features, known as defeaturing, can substantially reduce computational time while only minimally affecting simulation accuracy. However, determining which features should be retained and which can be removed relies heavily on domain expertise, as there are no systematic methods or established guidelines to inform this decision. This step in the simulation workflow becomes critical, as the pre-processing of geometries is required many times in a product development lifecycle, making the generation of high-fidelity digital twins more complicated. In our research, we introduce a novel methodology employing Graph Neural Networks to learn the criteria for effective defeaturing. Our method utilises generic CAD models along with automated simulation workflows to generate training data sets from simple models. The trained graph networks can then be used on more complex geometries during testing. Performance of the trained Graph Neural Network is assessed by comparing certain important values of a physical simulation (e.g. the maximum temperature) of the predicted simplified models to those of the original baseline assemblies. In computational experiments with models comprised of (a certain number of) parts, we observe a significant simplification of the geometry and corresponding simulation speed-up, with acceptable loss in fidelity w.r.t. the simulation objectives. The training and prediction processes have already been integrated into commercial CAD software, demonstrating the incorporation of machine learning into industrial simulation workflows, which facilitates the development of high-fidelity digital twins.