September 17, 2002

Statler Hotel, Ithaca, New York

Notes recorded and transcribed by Steve 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.

*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.

*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.

*Berger:* The real question is how automatable a standalone
program could be. *Shephard:* Industry demands standalone
implementations.

*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.

*Karniadakis:* Errors in the physical modeling must also be
considered,
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.

*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.

*Hoppe:* There is much progress in model simplification. See
IEEE Visualization Proceedings. There is work even on out-of-core
model simplification.

*Hoppe:* Yes, in Direct3D, and they're supposed to be fast.

*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.

*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.

*Shephard:* We must convince the CAD companies that
this application of CAD is important and useful so they
make multiple licenses affordable.