Artificial Intelligence for CAE

Services, Consulting and Software Solutions

Why using Artificial Intelligence for Postprocessing?

In general, there are two core questions when postprocessing simulations:

 1. What notable effects are occurring?
 2. How to avoid or trigger specific effects?

Humans can answer these questions for one or two big models, but if being confronted with five or six models, this usually exceeds our capabilities. Unfortunately as a general trend of digitalization, the amount of simulations performed steadily increases. For many companies it is already not possible anymore to analyze every simulation result thoroughly. In consequence, important issues get overlooked everyday. 

AI has the potential to automate human tasks such as postprocessing, enabling algorithms to compare results on their own. The key aspect is hereby not only to deal with the increasing amount of data, but also attain deeper insights into our existing data. Therefore for future Postprocessing it will be inevitable to incorporate AI technology in order not to loose sight of what's important in our simulation data. To demonstrate a fraction of our capabilities you can find an example project below.

Case Study of a Full Frontal Crash

As an example the simulation model of a 2007 Chevrolet Silverado will be used. The simulation model is freely available on the website of the National Highway and Traffic Safety Administration (NHTSA).

Figure 1: Full frontal crash of a 2007 Chevrolet Silverado (left). For data analysis, the crash absorbing structure (right) was divided into four sub-components: the bumper (cyan), the crossbeam (yellow), the left rail (red) and the right rail (green).


In order to trigger different types of deformation, 24 sheet thicknesses, the impact velocity and the impact angle were varied heavily in a Design of Experiments (DOE) study. In order to have a statistically significant amount of simulations for certainty, 1000 simulation runs were performed. Be aware that a much lower amount of runs is also possible. For explanatory purposes we will focus in the rest of the article only on the left rail (red) shown in figure 1.


What notable effects are occurring?

Manually analyzing and categorizing 1000 simulation results is not feasible and thus requires automation. Using simple scripting to extract some system responses would be an incomplete choice, since we might neglect important effects at locations where we do not measure. Our algorithms uses the entire mesh data to compare simulation results.

Visualizing the similarity of a thousand runs

We use a common method called low-dimensional embeddings in order to visualize the similarity of thousands of simulations.

Figure 3: Similarity cloud of 1000 left rail simulations. Every marker in the plot represents a simulation result. Close markers are runs with very similar results. In order to understand the overall behavior, one can cluster the data and view the most centric samples of the clusters. The samples are colored according to effective plastic strain.



Figure 3 shows the similarity cloud of the left rail from figure 1. The cloud has four obvious clusters, which are the major deformation modes. In order to understand the cluster differences it is helpful to view one representative sample for every cluster in a postprocessor. Cluster 1 reveals a good behavior for keeping the rail relatively straight during impact, whereas the samples of cluster 2 and 3 bend more upwards, which is not perfect but still acceptable. Cluster 4 contains only ~1% (=11) samples and deforms very inefficiently in terms of energy absorption, since the rail is bending upwards rather than being crushed. At this point of the analysis the question arises, how to avoid this rare but still very dangerous type of deformation. This question will be investigated later on.


What is happening in a cluster?

The similarity cloud of figure 3 can also be analyzed on a much finer scale. If samples are for example chosen along a path in cluster 2, one can view the transitional behavior inside the cluster itself on a much finer scale.

Figure 4: Choosing samples along a path in the similarity cloud reveals the transitional behavior of the simulation results.


How to avoid or trigger specific effects?

From the previous analysis, one is able to determine not only the overall deformation behavior of the structure, but also detect undesired types of deformation. We found the behavior of cluster 4 to be especially undesired, since the rail is bending upwards (figure 3, right), thus does not absorb the kinetic energy sufficiently well.

In order to investigate the issue, we will use Rule Mining. Rule Mining is a technique from the field of Knowledge Discovery in Databases (KDD) and has the goal to identify rules (if ... then ...)  for the system. An example rule is:

$if \; thickness \; <  \; 4 \; then \; intrusion \; < \; 100$,

which simply states, that if we would keep the design variable $thickness$ smaller than four, that we would always achieve an intrusion smaller 100. Before performing Rule Mining, the target condition must be stated first, which is an intrusion below 100mm in our case (legal regulation). The Algorithm then searches for multiple if-conditions leading to the specified behavior.

How to avoid the bending behavior of the left rail?

In order to avoid the bending of the left rail (figure 2, cluster 4), we define the target as:

$T := not \; cluster4$.

For this target the Rule Mining algorithm returns two design recommendations:

  1. $t_{rail} \; < \; 2.65$
  2. $t_{rail} \; > \; 2.65 \; and \; t_{crossbeam}  \;> \; 2.30$

The first recommendation is to reduce the thickness of the rail itself, which indirectly lowers the stiffness. This makes sense from a mechanical point of view, since lowering the stiffness makes the rail soft enough to be crushed instead of bending upwards.

The second recommendation simply states, that if the left rail needs to have a high stiffness, then one could also reinforce the crossbeam by increasing the thickness to keep the rail in place.

These design recommendations can also be visualized as zones in the design space. Click on the respective message in the rule mining app below.

As note: the app underneath is a "snapshot" of our rule miner. Within this app we can besides other functions mine rules for arbitrary expressions, as long as the expressions split the data in at least two groups.


What to do against cluster4?


You have 2 reliable options:

  1. t_rail_left < 2.65 (100%)
  2. t_rail_left > 2.65
    t_crossbeam_lower > 2.30 (99.6%)


What to do for intrusion < 100?


You have 2 reliable options:

  1. impact_velocity < 16037 (99.5%)
  2. 16037 < impact_velocity < 16475
    2.56 < t_front0_left
    2.45 < t_front0_right
    3.19 < t_motor_mount_right (94.9%)


Constantin Diez


Do you have questions regarding Artificial Intelligence?

We are happy to help you out:

Feel free to contact us


This website uses cookies

We want to improve your user experience by recording your surfing history on our website. We store information about your device and browser to recognize returning visitors. To do this, we use so-called cookies to make our website as user-friendly and secure as possible. Please refer to our privacy policy for details.