Notes on Panel Discussion at 11th International Meshing Roundtable
September 17, 2002
Statler Hotel, Ithaca, New York
Notes recorded and transcribed by
Vavasis, panel host & moderator
This panel was supported in part by
NSF grant ACI 0085969 on Adaptive
Software. Any opinions, findings, and conclusions or recommendations
expressed in this material are those of the author(s) and do not
necessarily reflect the views of the
National Science Foundation.
Questions to be addressed by the panelists
What software advances are needed to make adaptivity more useable
and popular among the computational community at large?
Which major computational problems represent the greatest
successes for adaptivity? Which major problems have defied adaptive
Berger (New York University)
In my work I use two kinds of adaptivity: adaptivity for the
geometry and adaptivity for the solution. Adaptivity should never
hurt a computation, and hopefully will help. But even with
adaptivity, a good enough initial mesh is still needed. Adaptivity
today still requires human intervention. Some issues holding back
adaptivity are: How can you adapt to a curved surface in the middle of
a flow solve? Error estimation is still a bottleneck. There are also
user-interface issues, for example, determining which fields to adapt
to. There is nothing like a PetSC for adaptivity. Some of the main
successes for adaptivity are shock dynamics, astrophysics and
Hoppe (Microsoft Research)
A major distinction in computer graphics is structured versus
unstructured. Local adaptivity not always needed in graphics. In
graphics, one prefers adaptivity that is incremental and amortizable.
Hardware implementations of adaptivity are preferable. Adaptivity in
graphics is view dependent, which is different from scientific
computing. A binary tree is a useful data structure that could go
into hardware. Shape deformation problems require adaptivity.
Turbulence presents a particular challenge for adaptivity because high
order methods are needed, and there is no dominant feature. The
problem with turbulence is that there are many scales. Turbulence is
very geometry-dependent, and the success of adaptivity depends on
geometry. For instance, geometry locally changes the Reynolds number.
Some bottlenecks to parallelism and adaptivity in MPI and OpenMP are:
p-refinement, thread creation, and barrier operations. Some other
issues for handling adaptivity are: Hybrid or nonconforming mesh
generation for adaptivity; lack of mathematical error estimators
means physical criteria must be used; ALE solvers are still slow.
Rob Leland (Sandia National Laboratories)
An issue with adaptivity is the integration of error estimators. Also,
adaptivity in hex meshes is still very difficult. Shock hydro problems
on hex meshes are not solvable adaptively yet. Priorities at the lab
are mainly on meshing and geometry, so adaptivity does not get as much
attention. Intuition among the users of simulation software
concerning errors is lacking. Also lacking is a good coupling between
error estimates and mesh generation on subsequent steps.
Adaptivity is not widely used because: Code structure does not support
changing the underlying mesh; there are insufficient computational
resources; adaptive simulation is not integrated into the design
process. Good error estimators for elliptic problems are well known.
Estimates for higher order elements and goal oriented estimators such
as vorticity-tracking are underdeveloped (although see work by Oden).
AOMD software for adaptive meshes makes the coding easier. Anisotropic
error estimates also need developing. Use of adaptivity in design and
in management of simulation is the main gap.
Discussion between the panel and audience
Question from Tim Tautges
How difficult is it to put adaptivity into existing codes?
Shephard: RPI writes their own adaptive codes from scratch.
Industry wants adaptive software integrated with existing nonadaptive
code. Industry doesn't need optimized software. Leland: SIERRA
provides an adaptive infrastructure for Sandia codes, but it is still
tough to retrofit adaptivity. Berger: If adaptivity is not more
widely integrated, then it won't be widely used. Karniadakis:
Adaptivity is not always the bottleneck. Correct physical modeling
may be the main problem. Shephard: Viseon airflow computations
were carefully validated, so at least in that instance physics
was not the problem.
Question from Scott Mitchell
How important is PetSC-like software?
Berger: There is CCA-AMR plus about six efforts to develop
a good infrastructure for adaptivity. The problem is that it is
very diffcult to provide useful interfaces to complex analysis
codes. Shephard: Keep in mind that PetSC is just a library
of parallel solvers.
Question from Reza Taghavi
Adaptivity is like design optimization: previously, design optimization
was available only in closed codes, but now is implemented as standalone
software. Perhaps there could be standalone adaptivity implemented via
Berger: The real question is how automatable a standalone
program could be. Shephard: Industry demands standalone
Question from Tom Peters
Adaptivity often localizes on geometric details. Why not have a better
CAD-solver interface? Also, how can we get estimates of errors due to
geometric flaws? Can adaptivity give feedback to the designer?
Shephard: If the geometry is piecewise linear, then the solver
will adapt to the wrong model. It's better to ask the CAD system to find
a point. Licensing expense for the
CAD system may or may not be an issue. We can also
have adaptivity ask the designer about length parameters. But
for that we need a higher level infrastructure that is missing.
Question from Dimitri Mavriplis
Currently, error estimators are lacking for many problems. There
is no point in adaptivity if there is no good error estimator.
Karniadakis: Errors in the physical modeling must also be
as well as discretization errors. Shephard: OK to use
mathematically sound but nonrigorous estimators for adaptivity.
Hoppe: Only geometric errors matter for graphics. Perceptual
errors are current hot research topic. Leland: Adaptivity
should also lead to greater efficiency.
Question from Todd Munson
How about adaptivity in video games, and in the interaction between
geometry and graphics?
Hoppe: Physically realistic modeling is a hot topic
in graphics. Adaptivity in graphics is not done so much at run time
as at authoring time.
Question from Lori Freitag
In addition to computing the
results, there is also a bottleneck in visualization.
Hoppe: There is much progress in model simplification. See
IEEE Visualization Proceedings. There is work even on out-of-core
Question from Carl Diegert
Are there simplification tools available from Microsoft?
Hoppe: Yes, in Direct3D, and they're supposed to be fast.
Question from Scott Mitchell
Why is adaptivity more successful in graphics?
Hoppe: Simulation is not an issue, and geometry is
a simpler problem.
Berger & Leland: Graphics should use more simulation.
Hoppe: Agreed. There are papers at the SIGGRAPH conference
in which the graphics processing unit is used for simulation.
Berger: Going back to the issue of CAD. It's not so easy
to integrate a solver with CAD. We'd like to have a lightweight
automatic geometric engine for efficiency.
Leland: Sandia spends months on geometry and meshing.
Question from Tim Tautges
Which geometric model should be used for adaptivity?
What happens when the mesh size is smaller than the geometric
feature size? The CAD engine must be part of the adaptivity.
Question from Karl Merkley
What about adaptivity for multiphysics?
Shephard: The simplest solution is for the finest
mesh to be selected, although this is not optimal.
Karniadakis: It's best to find a norm that
combines the user's criteria from each physical application.
Leland: If multiple grids are used, then parallel
grid transfer is very difficult.
Question from Ted Blacker
Using the CAD engine for adaptivity creates problems of expense
Shephard: We must convince the CAD companies that
this application of CAD is important and useful so they
make multiple licenses affordable.