Assembling over subdomain fails

[Gerd Wachsmuth]

Following setup:

I have a mesh (with FunctionSpace V) and a submesh (with FunctionSpace W). Assembling a linear form on V over this subdomain involving a coefficient from W will fail.

I think the problem is that Dolfin would like to evaluate the function from W at the boundary of the subdomain. Maybe due to rounding errors this point lies outside the subdomain and the evaluation fails.

It does not happen, if W is a CG or DG space.

[Marie Rognes (meg-simula)]

On 14. okt. 2010 09:13, Gerd Wachsmuth wrote:
> Public bug reported:
>
> Following setup:
>
> I have a mesh (with FunctionSpace V) and a submesh (with FunctionSpace
> W). Assembling a linear form on V over this subdomain involving a
> coefficient from W will fail.
>
> I think the problem is that Dolfin would like to evaluate the function
> from W at the boundary of the subdomain. Maybe due to rounding errors
> this point lies outside the subdomain and the evaluation fails.
>
> It does not happen, if W is a CG or DG space.
>
> ** Affects: dolfin
> Importance: Undecided
> Status: New
>
>

Have you tried setting

    parameters["allow_extrapolation"] = True

?

--
Marie

[Gerd Wachsmuth (gerd-wachsmuth)]

This works fine in 2D. But in 3D, extrapolation is not implemented.

[Marie Rognes (meg-simula)]

On 14. okt. 2010 11:13, Gerd Wachsmuth wrote:
> This works fine in 2D. But in 3D, extrapolation is not implemented.
>
>

Ah.

--
Marie

[Anders Logg (logg)]

Marie Rognes skrev 2010-10-14 12.28:
> On 14. okt. 2010 11:13, Gerd Wachsmuth wrote:
>> This works fine in 2D. But in 3D, extrapolation is not implemented.
>>
>>
>
> Ah.
>
> --
> Marie

The problem is that this functionality is missing in CGAL so we have to
implement it ourselves and no one has done that yet.

--
Anders

[Andre Massing (massing)]

Den 20.10.10 21.07, skrev Anders Logg:
> Marie Rognes skrev 2010-10-14 12.28:
>> On 14. okt. 2010 11:13, Gerd Wachsmuth wrote:
>>> This works fine in 2D. But in 3D, extrapolation is not implemented.
>>>
>>>
>>
>> Ah.
>>
>> --
>> Marie
>
> The problem is that this functionality is missing in CGAL so we have
> to implement it ourselves and no one has done that yet.
It is already on my todo list, but not with a highest priority.
--
Andre

[Gerd Wachsmuth (gerd-wachsmuth)]

On Wed, 20 Oct 2010, Anders Logg wrote:
> Marie Rognes skrev 2010-10-14 12.28:
>> On 14. okt. 2010 11:13, Gerd Wachsmuth wrote:
>>> This works fine in 2D. But in 3D, extrapolation is not implemented.
>>>
>>>
>>
>> Ah.
>>
>
> The problem is that this functionality is missing in CGAL so we have to
> implement it ourselves and no one has done that yet.
>

In my opinion the problem is that Dolfin want to evaluate the function at
points which does not belong to the integration domain - or at points
lying to close at the boundary.

In my example the function w is defined only on a subdomain. Hence there
should be no error if this function is integrated over this subdomain.
By extrapolating this bug is only covered up.

[Anders Logg (logg)]

Gerd Wachsmuth skrev 2010-10-21 12.58:
> On Wed, 20 Oct 2010, Anders Logg wrote:
>> Marie Rognes skrev 2010-10-14 12.28:
>>> On 14. okt. 2010 11:13, Gerd Wachsmuth wrote:
>>>> This works fine in 2D. But in 3D, extrapolation is not implemented.
>>>>
>>>>
>>>
>>> Ah.
>>>
>>
>> The problem is that this functionality is missing in CGAL so we have to
>> implement it ourselves and no one has done that yet.
>>
>
> In my opinion the problem is that Dolfin want to evaluate the function at
> points which does not belong to the integration domain - or at points
> lying to close at the boundary.
>
> In my example the function w is defined only on a subdomain. Hence there
> should be no error if this function is integrated over this subdomain.
> By extrapolating this bug is only covered up.

Have you checked which point is thought to be outside of the domain?
Perhaps it's a small round-off error that causes the problem?

--
Anders

[Gerd Wachsmuth (gerd-wachsmuth)]

On Thu, 21 Oct 2010, Anders Logg wrote:

> Gerd Wachsmuth skrev 2010-10-21 12.58:
>> On Wed, 20 Oct 2010, Anders Logg wrote:
>>> Marie Rognes skrev 2010-10-14 12.28:
>>>> On 14. okt. 2010 11:13, Gerd Wachsmuth wrote:
>>>>> This works fine in 2D. But in 3D, extrapolation is not implemented.
>>>>>
>>>>>
>>>>
>>>> Ah.
>>>>
>>>
>>> The problem is that this functionality is missing in CGAL so we have to
>>> implement it ourselves and no one has done that yet.
>>>
>>
>> In my opinion the problem is that Dolfin want to evaluate the function at
>> points which does not belong to the integration domain - or at points
>> lying to close at the boundary.
>>
>> In my example the function w is defined only on a subdomain. Hence there
>> should be no error if this function is integrated over this subdomain.
>> By extrapolating this bug is only covered up.
>
> Have you checked which point is thought to be outside of the domain?
> Perhaps it's a small round-off error that causes the problem?
>

Yep. The problem seems to occur at points lying on the boundary.

[Anders Logg (logg)]

Gerd Wachsmuth skrev 2010-10-21 15.00:
> On Thu, 21 Oct 2010, Anders Logg wrote:
>
>> Gerd Wachsmuth skrev 2010-10-21 12.58:
>>> On Wed, 20 Oct 2010, Anders Logg wrote:
>>>> Marie Rognes skrev 2010-10-14 12.28:
>>>>> On 14. okt. 2010 11:13, Gerd Wachsmuth wrote:
>>>>>> This works fine in 2D. But in 3D, extrapolation is not implemented.
>>>>>>
>>>>>>
>>>>>
>>>>> Ah.
>>>>>
>>>>
>>>> The problem is that this functionality is missing in CGAL so we have to
>>>> implement it ourselves and no one has done that yet.
>>>>
>>>
>>> In my opinion the problem is that Dolfin want to evaluate the function at
>>> points which does not belong to the integration domain - or at points
>>> lying to close at the boundary.
>>>
>>> In my example the function w is defined only on a subdomain. Hence there
>>> should be no error if this function is integrated over this subdomain.
>>> By extrapolating this bug is only covered up.
>>
>> Have you checked which point is thought to be outside of the domain?
>> Perhaps it's a small round-off error that causes the problem?
>>
>
> Yep. The problem seems to occur at points lying on the boundary.

On the boundary between different subdomains or on the actual boundary?

--
Anders

[Gerd Wachsmuth (gerd-wachsmuth)]

On Thu, 21 Oct 2010, Anders Logg wrote:

> Gerd Wachsmuth skrev 2010-10-21 15.00:
>> On Thu, 21 Oct 2010, Anders Logg wrote:
>>
>>> Gerd Wachsmuth skrev 2010-10-21 12.58:
>>>> On Wed, 20 Oct 2010, Anders Logg wrote:
>>>>> Marie Rognes skrev 2010-10-14 12.28:
>>>>>> On 14. okt. 2010 11:13, Gerd Wachsmuth wrote:
>>>>>>> This works fine in 2D. But in 3D, extrapolation is not implemented.
>>>>>>>
>>>>>>>
>>>>>>
>>>>>> Ah.
>>>>>>
>>>>>
>>>>> The problem is that this functionality is missing in CGAL so we have to
>>>>> implement it ourselves and no one has done that yet.
>>>>>
>>>>
>>>> In my opinion the problem is that Dolfin want to evaluate the function at
>>>> points which does not belong to the integration domain - or at points
>>>> lying to close at the boundary.
>>>>
>>>> In my example the function w is defined only on a subdomain. Hence there
>>>> should be no error if this function is integrated over this subdomain.
>>>> By extrapolating this bug is only covered up.
>>>
>>> Have you checked which point is thought to be outside of the domain?
>>> Perhaps it's a small round-off error that causes the problem?
>>>
>>
>> Yep. The problem seems to occur at points lying on the boundary.
>
> On the boundary between different subdomains or on the actual boundary?
>
>
On the boundary of the subdomain (where the function is defined and over
which we integrate) and its complement.

[Garth Wells (garth-wells)]

Could you post the simplest possible piece of code that reproduces the problem?

[Gerd Wachsmuth (gerd-wachsmuth)]

I've already posted a piece of code above.

[Marie Rognes (meg-simula)]

We have added an additional strategy for evaluating functions on
semi-matching meshes. The three-dimensional version of the above
example (with "allow_extrapolation" = True) now runs without error.