These slides were presented at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

With the rapid development of AI applications in the recent years the ever growing amount of simulation runs being performed has even more increased, especially with respect to provide simulations to train the AI models with. While current Simulation Data Management systems and the IT infrastructure already allow storing and accessing huge datasets and would facilitate putting this into action for analysis, the users usually only have tools and the time to make rather straight forward model to model comparisons, between current model versions and their immediate predecessors. Making use of the Principal Component Analysis, a dimension reduction algorithm out of the unsupervised learning techniques, a new database was developed and presented in the last NAFEMS World Conference. Continuously being fed with new simulation runs, this database enables us to automatically detect unknown behaviour in the most recent simulation runs compared to all predecessors at a time. To achieve this, the database does not only need to store and detect every new deformation pattern, but in addition several obstacles like a mapping between different simulation models, a performance efficient database format and a technique to also detect local effects had to be addressed. Taking a look at the special needs from engineers to also being able to include history curves into the analysis, the database now had to be extended to also being able to not only store and compress the curve data, but also use the outlier detection on the history data. Furthermore in case a deeper analysis of the curve anomaly is being needed it is shown how structural part deformation behaviour can be correlated against the curve information to derive not only curve to curve relationships, but also being able to compute part to curve correlations. Thus engineers become able to derive design suggestions which lead to an improvement in curve behaviour. In addition the search for deformation patterns had to be extended to also being able to search for similarities among time history data, to being able to identify simulations with a similar behaviour.