Hi William,
You have found a bug! A fairly heinous one at that. So, it appears
that arithmetic face values were back to front in terms of the
assumption about the face to cell distance ratio. The fix for this
issue is this:
$ svn diff
Index: fipy/variables/arithmeticCellToFaceVariable.py
===================================================================
--- fipy/variables/arithmeticCellToFaceVariable.py (revision 3186)
+++ fipy/variables/arithmeticCellToFaceVariable.py (working copy)
@@ -42,7 +42,7 @@
def _calcValuePy(self, alpha, id1, id2):
cell1 = numerix.take(self.var, id1, axis=-1)
cell2 = numerix.take(self.var, id2, axis=-1)
- return (cell1 - cell2) * alpha + cell2
+ return (cell2 - cell1) * alpha + cell1
def _calcValueIn(self, alpha, id1, id2):
val = self._getArray().copy()
All, the tests appear to pass when I make this change other than two
tests that are manifestly broken due to our assumption about the
correct arithmetic face average. I am not sure how we have gone along
with this particular problem for so long, but evidently we have. Most
of our own work is with regular grids and hence we obviously haven't
noticed it. Anyway, I'll correct this in version 2 and on trunk and do
a release of version 2.0.3 fairly soon. Many thanks for raising this
issue,much appreciated. I'll let you know when I've fixed the new
version and fixed trunk.
PS You might want to update your version of fipy. It appears you are
using a pre 2.0 version. Definitely recommend upgrading, but you might
want to wait for version 2.0.3.
Cheers
On Wed, Jul 29, 2009 at 11:19 PM, William Gathright<[email protected]> wrote:
>
> Hello,
> I was calculating gradients over a nonuniform mesh when I noticed that
> the gradient that Fipy calculates is not what I expect. I have
> constructed the simplest example I could think of below to show this
> behavior. In this example, I calculate the gradient of a
> monotonically increasing variable in 1D. I expected to see a gradient
> of 1 on all interior cells. Understanding that Fipy expects zero-flux
> boundary conditions on all variables, I expected to see a gradient of
> 0.5 for the boundary cells. These are exactly the results from using
> the uniform grid.
>
> When the gradient is calculated using the non-uniform grid spacing it
> does not come back with these results. Instead, the gradient is
> slightly larger than 1 except at the mid-point. In this example, the
> mid-point is special because it is where the grid spacing switches
> from getting smaller to getting larger. At this point, Fipy
> calculates a gradient that is much too large (11% in this case).
>
> I did not see any bug reports about such a problem in the archives.
> This was run on a Windows Vista machine, Intel Centrino Duo2
> processor, Fipy 1.2, and Python 2.4.
>
> Is there an alternate way of specifying a nonuniform grid spacing to
> avoid this unexpected behavior?
>
> Thanks for your help,
> Will Gathright
>
> #------------ Begin Example Code -------------------------
> from fipy import *
>
> #The uniform grid will not cause the strange behavior
> #dx = Variable("1 nm") * numerix.cumproduct(numerix.concatenate(([1]*100,)))
>
> #The variable grid seems to change the way that FiPy calculates the gradient
> #This grid spacing gets smaller and smaller for 10 points, then larger
> for 10 points
> dx = Variable("1 nm") *
> numerix.cumproduct(numerix.concatenate(([0.9]*10, [1/.9]*10)))
> L = numerix.sum(dx)
> dx = dx / L
> mesh = Grid1D(dx=dx)
>
> A=CellVariable(mesh=mesh, name='A', hasOld=1)
>
> x = mesh.getCellCenters()[...,0]
> A.setValue(x)
>
> print A.getGrad().getValue()
>
> tsv=TSVViewer(vars=[A, A.getGrad()])
> tsv.plot('getgrad_test.txt')
>
>
--
Daniel Wheeler
