Question about loss2 about harmonic-oscillator-pinn

[gersonjferreira]

The initial conditions are hidden in the first loss. You can reduce the number of points there to

x_data = x[0:2]
y_data = y[0:2]

and the code still works.

I wrote (with my student) a version of this code reducing the 2nd order ODE into two 1st order ODEs, then it gets clear where the initial conditions enter. In your code the data is simply the initial position and velocity at t=0.

Regarding loos2, indeed the 1e-4 factor is weird and I was also checking that. I use:

    physics = (dx2 + mu*dx + k*yhp) / k
    loss2 = torch.mean(physics**2)

Dividing physics by k helps because the loss1 and loss2 error are now written in the same units. The 1e-4 term seems to be an implicit workaround to fix the units.

I'll put my code on gitlab soon.

[Sanchez-web]

Thanks, your explanation was great!
Looking forward to your code!

[gersonjferreira]

Thanks. My version is nearly identical to the one here. The only differences are: (i) this normalization of the loss to fix the units/dimensions; (ii) a second version in terms of two 1st order ODEs (reduction of order); (iii) many comments on each step to make it as pedagogical as possible (since this is a project for students entering this research field).

My version is ready, but I want my student to develop her version on her own. So, I'll only upload to gitlab once she finishes her version. It may take a while (couple of weeks maybe), she is still at the beginning of the project.