aka Sampling the Parameter space
For advanced material models it may be quite difficult to know how to set a good initial value for the many parameters. We know that in using optimization based tools, the success in finding a good solution may be highly dependent on these initial values. Certain math models may have multiple local minima and this adds another layer of difficulty.
Back when I was learning about Isight, I struggled with this same issue. My colleagues Malik Kayupov and Charles Yuan suggested that in these situations it may be best to run a DOE first, over a fairly broad parameter space (large span of min & max bounds). Recently I was having a similar discussion with Dan Cojocaru, a member of the calibration app team who works on the calibration kernel, including things like optimizers and the emerging NN surrogate technology. Dan pointed out that while the calibration app does not have a dedicated DOE functionality, some of the optimizers begin by sampling the design space. The optimization controls can be used to (mostly) degenerate the process to sampling the design space.
•The following slides illustrate how to use the sampling capabilities available under existing global minimizers to identify a decent initial guess:
•Latin hypercube sampling
•available under the Bayesian minimizer
•Random sampling
•available under the Differential Evolution and Particle Swarm minimizers
Each of the global minimizers lets the user select to automatically run of one of the local minimizers (HM, JC, etc) after the sampling process completes.
My background (Tod) in optimization technology is fairly limited, so if there are errors here, please be kind and helpful ;-)
This opens up the question of whether the calibration app should have a dedicated DOE functionality, what do you think?
As Dan's slides mention, the Abaqus runs as part of the sampling process are not run in parallel today, we have this on our RFE list for future development.
For more background on this Holzapfel-Ogden example, please see: