Submission Exercise 3

Deadline: Please see the time table in Ilias.

As submission for this exercise upload your files (named below) in a zip-folder on Ilias.

Solution hand-in

Please hand in a zip-directory named NameOne_NameTwo_NameThree_Abgabe3 containing:

  • your trained model as model.pt

  • the min_max.yaml file containing minimum and maximum values of all input and output channels in previously described format

  • a python file (mymodel.py) containing a function (init_my_model()) to initialize your model

  • a results-file containing

    • the measured evaluation metrics of your best model for train, val, test RMSE, MAE on the un-normed predicted u, v fields

    • number of trainable parameters of your best model

    • train vs. validation loss graphs

    • 1-2 sample data point predictions and comparisons to their labels (optional: visualization of vector fields via plt.quiver on the un-normed velocity fields)

    • visualizations of the three out-of-distribution test samples (have an explanation of what you can see at hand for the oral exam)

Rules

The same rules as for exercise 1 apply regarding commented code, correct results, etc. No third-party libraries are allowed except for numpy, torch, matplotlib, pathlib, yaml, PIL.Image, typing, pyvista.

Questions for the interview (“Abnahme”)

  • What happens if you apply your model to an out-of-distribution Reynolds number and boundary condition?

  • What happens if you apply your model to a larger spatial domain?

  • What happens if you apply your model to an input field where the passing-by-velocity is not at the top but at the bottom (or side wall)?

  • What can you learn from having a high MSE loss but a low MAE loss and vice versa?