Menu
Search Contact

Surrogate Modelling

Applications

  • Typical full vehicle crash modells consists of elements with edge length of 1mm or below. This may be accurate for global strength evaluation but its not if very detailed structures, like coils, capacitors or chips within the components are analyzed. Surrogate Modeling can here be used to map a detailed model into the comparatively coarse model
  • Enables a detailed analysis of safety relevant values without smaller mesh size or reduced time step.
  • The detailed model is not limited to structural mechanics and may also account for other physics like electromagnetics or heat-flux. The comlexity of the coarse model can be reduced in mesh size, geometry or the material model.

Impact Classifier for Crash Dummy Head

  • If only e.g. the time of collision from the Dummy Head to any Vehicle part is of interest, one can remove the Dummy from the model to reduce complextity.
  • The contact evaluation can be shifted to post-processing by using a data driven model

A causal chain analysis enables surrogate modelling

  • availability of graph based algorithms
  • probabilistic Machine Learning algorithms to do causal correlation analysis

 

Challenge

Full vehicle crash models are huge in complexity and calculation time. Surrogate Modelling can be used in 2 ways

  1. Map the deformation of a roughly modeled component on the results of a detailed model to expand the prediction capability of the comparable coarse full vehicle model
  2. Remove submodels with small impact on the overall stiffness and where only specific values or signals are of interest

Solution

FEM Operations Toolbox as well as Geometric Operations Toolbox have several methods to calculate features as 1D or 2D sequences which can further be used for several Machine Learning Models like Neural Networks, Support Vector Machines, Gaussian Process Regression, Convolutional Neural Networks, LSTM and many more (see also the ANDATA Tool Brainer).

An important task is to check the validity of the used machine learning model and case of new data. For this challenge we often use our tool Expectator in combination with Self-Organizing-Maps.