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

Abstract

For the robust and reliable design of products, the influence of tolerances must already be taken into account during the design process. Here it is important to have simulation methods available that not only map the function on the basis of the nominal geometry, but also take the influence of tolerances into account. Statistical tolerancing results in a propability distribution of the resulting variants and thus also a probability of failure if it is possible to simulate each of the possible variants with regard to function (e.g. strength, acoustics, etc.). However, this is where the challenge lies, as several thousand variants have to be considered for statistical tolerancing. With conventional simulation methods (FEM, MBS, etc.), standard computers and average model sizes, the calculation of all tolerance situations would take several months. However, if the tolerances are not taken into account, there is a risk that the product will not be reliable under certain tolerance combinations and that unexpected failures will occur. In order to include the statistical tolerance analysis in the functional analysis and thus ensure the reliability of the product, an alternative simulation method must be used. Machine learning methods can be used here to create prediction models. These prediction models deliver results in real time and are therefore very well suited to carrying out a very large number of analyses in a short time. The training data for machine learning is usually created using a design of experiments in combination with automated simulation of variants. Using the example of a gearbox with two gears, the presentation shows how the simulation models can be prepared for such an automation in order to ensure safe and reliable execution of the automatic process, even with more complex simulation models, like the contact model of the gear pair. The predictive quality of the machine learning model is examined and evaluated. The time advantage resulting from the use of machine learning model is also evaluated. In the example shown, the strength assessment for a statistical tolerance analysis with 10,000 variants can be reduced from 140 days to only 4 days.