Helmholtz pytorch predicted problem

[FemLearner]

I try the tensorflow version of the Helmholtz problem and it was working correctly:

But when I try the version of pytorch , and whatever i did(changing the layers or iter) ,the result was bad , if there is some parameters make this problem happen?I checked out the code and I think there is nothing wrong?

[aho38]

I fixed it if you wanna check out mine (https://github.com/aho38/Pytorch-PINN-Marine-Lake/tree/main/Data_Discovery). The issue was in autograd of the parameters. It's computing the wrong input.

[lrk612]

Just change the code for calculating the second derivative in loss_PDE to

u_x1 = u_x[:, [0]]
u_x2 = u_x[:, [1]]
u_xx1 = autograd.grad(u_x1, g, grad_outputs=torch.ones_like(u_x1), create_graph=True)[0][:, [0]]
u_xx2 = autograd.grad(u_x2, g, grad_outputs=torch.ones_like(u_x2), create_graph=True)[0][:, [1]]

The original code will mix the partial derivative into the Laplace term.