I am trying to get accurate predictions for the deflection of a beam using the Altair physicsAI. I want to get a field prediction so I am looking into the GCNS and TNS architecture.
Initially I followed the I-beam tutorial. This gave very accurate results, even for smaller data sets. However I noticed that all beams had the same mesh and the variations in dimensions were very small. This did not seem realistic for a real-life application.
I setup my own dataset of 12 rectangular beams. The mesh size is the same for all beams, so the number of elements and nodes differ. The set is shown below. The length is 1000 mm for every beam.
I tried both the GCNS and the TNS to predict the field for these beams. The hyperparameters I used are shown in the table below:
The best prediction I obtained is shown in the figure below:
The results for the TNS were not bad (the total deflection was very accurate). However, the field has a really weird pattern. This pattern does not show up in any of the FEA analyses, so I find it strange that it does show up in the predictions.
What changes could I make to my dataset or my hyperparameters to get better results?
One additional question I have is what do the heads and sections do for the TNS architecture and how should I determine how many I need there?
