Parametrized Kirsch


At the moment I am trying functionality of the framework on the Kirsch problem, specifically parameterize model due to the cutout radius in first experiment and model due to the angle of rotation of a elliptical cutout in second

Links to repo you can find above.
There you can find scripts and config files(I used config_fc.yaml with default parameters for fully connected neural network) for my experiments. But I have very large losses for all time during my training(minimal value ~ 1e4). Could anybody explain my mistake?
I attach my Tensorboard’s graphics.

Hello, have you tried normalizing your equations? I remember when we did the linear elasticity problems we spend a while coming up with the correct normalization. Ideally you want all of the equations/boundary conditions to have around the same scale. This can be somewhat tricky with these kinds of problems though.