search

holl- / PDE-Control

Code for the ICLR 2020 paper "Learning to Control PDEs"
MIT License
29 stars 8 forks source link

Question regarding the CFE chain differentiable loss example for Burger's equation #4

Closed kuwt closed 3 years ago

kuwt commented 3 years ago

In this example,
image

I see that the loss is just force loss. I wonder whether the neural network will just ignore the target state input and imitates the natural flow directly.

holl- commented 3 years ago

That's from burgercfe_diffphys.py, right? In the Burgers example, the last force is computed to fit the target state exactly. This will be very large if the network does not predict any prior forces.

You can see that in line 52: u_reconstructed = self.u_reconstructed = [u_true[i] if i == 0 or i == n else None for i in range(n+1)] which fixes the first and last state to the reference.