We asked and 201 of you answered. You told us about the types of meshes you’re generating (structured hex, unstructured or hybrid, Cartesian), whether they’re overset or high-order, how big they are, and more. We reveal your answers here in Part 1 and we’re saving replies to some of your long-form comments for Part 2.
Because Question 1 collected email addresses to ensure one reply per respondent, we begin with Question 2.
Q2: What mesh generation software do you use primarily?
We ask this question to be able to ensure the results aren’t biased toward Pointwise users and they’re not. Nearly 60% of the respondents use another mesher. No one skipped this question.
Q3: What types of meshes are you generating?
Here we were seeking the cell topology of the mesh you generate regardless of what type of solver you run it in. For example, you could generate a block structured hex mesh but run it in an unstructured solver. You could generate an unstructured mesh but your flow solver could convert it to polyhedra. We wanted to know what you generated, not how it was used.
We asked for responses that totaled 100% to account for people who generate different types of meshes. No one skipped this question.
Here’s an expansion of the bar chart’s labels to the full text included in the survey.
- Block Str = Block Structured Hexahedra (in other words, structured grid, mapped mesh)
- Prism-Tet = Hybrid Prism-Tetrahedra
- Uns Tet = Unstructured Tetrahedra (and only tetrahedra)
- Prism-Hex-Tet = Hybrid Prism-Hexahedra-Tetrahedra
- Hex-Tet = Hybrid Hexahedra-Tetrahedra
- Poly = Polyhedra (including cells with hanging edges)
- Cart cut-cell = Cartesian (with cut cells at boundaries)
- Cart+layers = Cartesian (with boundary-conforming prism or hex layers)
- Cart stair-step = Cartesian (with stair-stepped boundaries)
- Cart IB = Cartesian (with immersed boundaries)
Who else was surprised by the overwhelming popularity of block structured grids? I love them as much as the next CFD guy but I figured everyone had moved on to some form of hybrid unstructured mesh.
I also notice, without trying to infer any meaning, that polyhedra (both honeycomb polyhedra and hanging edge polyhedra) have a popularity that sits right between your more traditional body-fitted, canonical cell type meshes and Cartesian variants.
Q4: Is your solver structured or unstructured?
Regardless of what type of mesh you generated, the answers to this question tell us how your flow solver sees it. No one skipped this question.
More than two-thirds of you are feeding your grids, regardless of their native type, into an unstructured flow solver. So while you may be getting the benefits of structured hex cells, you’re not gaining the computational efficiency associated with a structured flow solver.
But having started my career using structured grids and structured flow solvers it’s nice to see them still hanging around, at least for about one-third of you.
Q5: What discretization scheme does your solver use?
The purpose of this question should be fairly self-explanatory. Knowing how the solver works with the mesh provides insight into what quality metrics of the mesh will be more beneficial. Sixty-seven respondents skipped this question and I’ll simply guess that they didn’t know the answer. (And therefore we should’ve included I Don’t Know as a choice.)
Here’s an expansion of the bar chart’s labels to the full text included in the survey.
2O FV CC= 2nd order finite volume cell-centered
2O FV VC = 2nd order finite volume vertex-centered
2O FE = 2nd order finite element
Other = other
H-O CG = High-order continuous Galerkin
H-O WENO = High-order WENO/MUSCL/expanded stencil
H-O DG = High-order discontinuous Galerkin
Clearly, the majority of you are using a solver with a second order, cell-centered, finite volume scheme.
The next two questions asked about other mesh treatments beyond their cell topology.
Q6: Are you using overset meshes?
Use of overset meshes used to be confined to a highly-specialized community, mostly in aerospace, who used a relatively small suite of tools. Recently, overset capabilities have been appearing in many more CFD solvers. So who’s actually using them now?
The fact that one quarter of you are using overset meshes is, in my opinion, a vast growth relative to what we would’ve measured five years ago.
And whereas 5-10 years ago overset mesh was nearly synonymous with structured grid, the newer algorithms can handle a wider variety of mesh types. That’s what the next question sought.
Q7: What types of component meshes are included in your overset meshes?
The 75% of respondents (150) who aren’t using overset meshes (see Q6) skipped this question.
You can see that traditional overset structured grids still are most popular but use of hybrid and unstructured meshes in an overset topology don’t trail by too much.
Q8: Are you generating high-order meshes?
Because the use of curved, high-order meshes in high-order accuracy solvers is contemplated as a future staple of production CFD, we wanted to find out how prevalent they are now. Only seven people skipped this question.
There are two ways to look at these results. One is that high-order meshes aren’t ready for prime-time yet. The other way is to say that it’s only taken the last couple of years to penetrate 20% into the solver market.
Q9: To what polynomial degree are your elevating your polynomial meshes?
Here we were not interested in the order of accuracy of the CFD solution (often designated by the letter P). Instead we just wanted to know how you curved and elevated the mesh itself (often known as the Q as in a Q3 mesh is a cubic mesh). Only the people who answered yes to Q8 answered Q9 (162 skipped).
Curving and elevating your mesh above Q4 doesn’t seem too popular. In fact, 61% of respondents don’t go beyond Q3.
Q10: What percent of your meshes are planar 2D?
Here we are asking about 2D meshes as the end result, not 2D meshes generated as components within a 3D mesh. Someone has to be doing 2D CFD, right? Only eight people skipped this question.
Yes, it seems that almost half of all meshes generated are for 2D CFD. Which is counterbalanced by the one-third of you who don’t do any 2D CFD at all.
Update 13 Mar 2020: Thanks to alert reader Christopher for pointing out how horribly I mangled the interpretation of the answers to Q10. That’s what I get for writing blog posts at home while ostensibly out on vacation. His interpretation, quoted here nearly verbatim, is probably what I should’ve written. “81% of respondents make 2D meshes less than 25% of the time.”
Q11: Roughly how many cells are in one of your typical meshes?
I was perhaps most interested in the answer to this question. Unfortunately, I could’ve worded it better. As one respondent asked, is 50 million cells supposed to be counted at 10 million or 100 million? Only 8 people skipped this question.
If you’re not using at least 1 million cells, you’re not trying. That’s one takeaway from these results. But the real sweet spot is in the 10-100 million cell range.
And let’s wish good luck to the 3% of you using 1 trillion or more cells in your exascale meshes.
Q12 was where we collected comments and responses to those will be included in Part 2 of this post. So that leads us to the demographics questions.
Q13: What type of organization do you work for?
This question was included to ensure we captured a variety of users across all organization types. Note that a research organization is not an degree-granting institution: that’s academia. Only 16 people skipped this question.
Q14: For what type(s) of applications are you generating meshes and running CFD solutions?
This question was of the “check all that apply” type and only 15 skipped.
Writing a good survey – meaning one that produces results that clearly provide insights in the subject matter – is difficult. Even though I spent a decent amount of time thinking through this one, hindsight tells me we could’ve asked the questions differently or added a few more clarifying questions. Hopefully I’ll get better with practice.
Thank you to the 201 of you who participated and shared your information. I think that’s the largest response we’ve had to a survey yet and we appreciate it.
Don’t hesitate to suggest ideas for future surveys.
Coming in Part 2, I’ll respond to some of the free-form comments we collected in this survey. Update 13 Mar 2020: Part 2 is now ready for you to read.
And if these results have you wondering about Pointwise’s capabilities for generating structured, unstructured, hybrid, 2D, 3D, overset, high-order, and exascale meshes maybe it’s time for you to request a free trial.