I while back I wrote about how clustering can help build expert knowledge into optimisation. Since then we've thought about how this technology can be made accessible to engineers without requiring programming or data science expertise. This is an integral part of what the Engineering Data Science team focuses on. We research machine learning techniques applied to engineering problems and make them acessible to engineers.
In this post I'd like to show you how a front crash rail with 47 design variables can be optimized to conform to a particular deformation mode using our traditional products HyperView and HyperStudy.
The starting point is a DOE with 200 points created by HyperStudy and each run has been evaluated using RADIOSS. In the video below you can see how the 200 runs are clustered into two groups with distinct deformation modes, visualized in HyperView. Note that is required of the user is to choose the loadcase, result type and the number of clusters to view on screen. After clustering, a classifier is created. This is a machine learning model that can be used to distinguish between the two deformation modes.
Suppose we now look to design a front rail that, apart from other functional requirements, crushes axially. All we need to do is add a constraint requiring the classifier to be certain that this is in fact an axial crush. In the video below you can see how the user adds this constraint (>80% probability that it's an axial crush) and then starts the optimiazation process. The top right window shows the objective function and the bottom middle window shows the new constraint that was added. Note that this constraint stabilizes on 80%, which is what we asked for.
Below you can see a comparison between three designs:
- Unoptimized
- Optimized without expert emulation
- Optimized with expert emulation
Please comment below with your thoughts on how expert emulation can be used for other applications. Also, see how this process was used at BMW.
