I have a fundamental understanding problem here. In the linear elasticity example with the bracket, one defines the geometry of the bracket with computational geometry and then learn the physics using the Linear Elasticity module and their inherent equations.
I am confused to what has the model learned? Has it learned the linear elasticity physics for that specific geometry or has it learned to generalize the physics on the bracket for different geometry ie If the length of the bracket is 2x times longer than the one we used for training? Since the physics for different bracket geometry for the same load and material is the same, can the model generalize them all? If not, how can this be done in Modulus?
The way this problem is set up in the UG, it has only learnt for that geometry. In some cases, if the new configuration is not too far off, transfer learning could work (e.g. aneurysm) but in other cases it’ll need to be parameterized (e.g. 3-fin or the industrial heat sinks). There are neural operators that can learn the parameterization implicitly with some data which is the current focus of Modulus in terms of further development
Thank you. That is exactly what I was looking for in the examples.