Use inheritance for check_vertical_flux #104
Conversation
daubners
commented
Nov 7, 2024
daubners
commented
- Removed some artefacts from previous merge
- adjusted calc_vertical_flux for each solver such that
- check_vertical_flux can be inherited from base solver
| def check_vertical_flux(self, conv_crit): | ||
| vert_flux = self.calc_vertical_flux() | ||
| fl = torch.sum(vert_flux, (0, 2, 3))[1:-1] | ||
| fl = torch.sum(vert_flux, (0, 2, 3)) |
There was a problem hiding this comment.
All the check_vertical_flux functions only differed in the way the field was cropped (see deleted code below) but all the other code lines where duplicated. If the calc_vertical_flux already gives back the right size, there should be no need for cropping and, thus the function can be inherited to all other solvers.
| self.conc[:, :-2, 1:-1, 1:-1] | ||
| vert_flux[self.conc[:, :-2, 1:-1, 1:-1] == 0] = 0 | ||
| vert_flux[self.conc[:, 1:-1, 1:-1, 1:-1] == 0] = 0 | ||
| # Indexing removes boundary layers (1 layer at every boundary) |
There was a problem hiding this comment.
Please write another sentence here
There was a problem hiding this comment.
The input image has the size Nx, Ny, Nz. The field for solving the PDE is Nx+2, Ny+2 and Nz+2 to add cells for boundary conditions. The flux in x-direction should only be computed in the inner of the domain (excluding boundary cells). Furthermore, the flux field is computed from the concentration gradient and, thus, should have dimensions Nx-1, Ny, Nz. This was inconsistent before between all the different solver. Now every solver, no matter how many boundary cells gives back a flux field with size [Nx-1, Ny, Nz]. This is also the reason for the simpler code in the next comment.
| vert_flux[self.conc == 0] = 0 | ||
| vert_flux[torch.roll(self.conc, 1, 1) == 0] = 0 | ||
| # Indexing removes 2 boundary layers at top and bottom | ||
| vert_flux = self.conc[:, 3:-2] - self.conc[:, 2:-3] |
There was a problem hiding this comment.
aren't you neglecting the .roll element?
There was a problem hiding this comment.
I replaced the roll with indexing as it was easier to control the size of the vert_flux output field. The flux in x should always be right minus left neighbour, i.e. here 3:-2 minus 2:-3. Before the flux direction was the wrong way and taking the abs value distorts the actual convergence. In the converged state there should only be flux in positive x direction but before, there can be back-flux in dead-end features which should be accounted for with the correct sign and direction
