Abstract

Computer simulations are one of the most important tools in the product development cycle. The generation of simulation models for these, however, can be a complicated and tedious task. Especially in complex assemblies the need for model and boundary condition simplifications rises. These simplifications can reduce the computational effort on the one hand but increase the effort for the model creation on the other hand. Besides involved tedious manual steps, engineering knowledge and experience are important keys for choosing modelling techniques to reduce model complexity while maintaining a sufficient result quality. This aspect of drawing profit from experience for modelling decisions is difficult to capture and to integrate digitally in an automated process. This scientific work is presenting an approach towards a knowledge and experience base for building a digital model understanding. This experience is gained from data which associates geometric information to simulation results. After learning from this data, the algorithm is able to select appropriate Finite-Element (FE) modelling representations while considering the resulting simulation accuracy. The developed knowledge representation is then guiding simplification decisions in the automated CAD-to-FE process. In the scope of this work, the focus is set on the converting and simplifying aero engine casing structures for structural simulations. The starting point is an automated process to transfer quasi-axisymmetric components to analysis models of various level of detail. This segments the geometry into substructures and provides a feature vector containing information about the structure along. Afterwards, the geometric information is coupled with data regarding the associated FE entities and the final simulation result. The simulation model quality is evaluated by composing basic dynamic properties and simulation model properties, e.g. number of degrees of freedom, as conflicting objective. These processes are combined with parametric CAD models and implemented in design-of-experiments studies to build the training data. Finally, self-learning algorithms take this information as input to build the digital knowledge representation.