Taking the Expert out of CFD
Digital Manufacturing Report buried the essence of their article CFD for the Rest of Us in the very last paragraph where Intelligent Light’s Steve Legensky is quoted “[CFD] is no longer just an expert’s game.”
Velocity field around a racing bicycle with Zipp Firecrest 808 wheels. CFD solution by CD-adapco’s STAR-CCM+, visualization by Intelligent Light’s FieldView. Image from Digital Manufacturing Report.
Dell also misses the boat a little bit in their take on this article, HPC in the Cloud Can Bring CFD to the Masses. While cloud-based high performance computing certainly improves access to resources they walk a fine line in saying “the Cloud” makes CFD more accessible to organizations “not interested in investing in staff with CFD expertise.”
In case you missed the news, Intelligent Light was recognized at SC12, the supercomputing conference, with the HPC Innovation Excellence Award for their efforts on improving CFD workflow as proven by the direct benefits realized by Zipp Speed Weaponry, a manufacture of bicycle wheels. Zipp’s new product line has been so successful the company added 120 jobs.
Intelligent Light’s multi-year study of CFD workflow and the application of that study to their FieldView product showed how advances in automatic meshing, on-demand CFD solvers, and efficient, remote extraction of engineering results from big data files can deliver the engineering insights Zipp’s engineers needed while minimizing the overhead typically associated with CFD.
I’ll nitpick Dell’s headline by pointing out that the masses don’t need CFD. And DMR’s headline gives me too much of an Oliver Twist vibe. I’ve said before that CFD is about process not processors, a sentiment that Legensky expresses much better. CFD used to be like a computer from the 1950s, a huge contraption tended to by legions of dedicated followers. CFD needs to be like a [dare I say] iPad, that you just pick up and use reliably.
This topic is also covered by Concept to Reality magazine.
Caveat: You still must understand and appreciate fluid dynamics before using CFD in the same way that you must understand (and maybe fear) electricity before rewiring your breaker box.
Events
- The first Femap Symposium will be held 26-27 June 2013 in Cincinnati.
- Related, watch this video about geometry and mesh associativity in Femap.
- The SIAM Conference on Geometric and Physical Modeling will be held 11-14 November 2013 in Denver. The call for papers should be posted soon.
- CD-adapco announced the agenda for the STAR Global Conference.
News in Brief
- NASA’s Dryden Flight Research Center is looking to hire an engineer to do CFD.
- FLOW-3D News for Winter 2013 is now available online.
- From TED-Ed, learn the difference between explicit and implicit numerical schemes from Lorena Barba.
- ANSYS and the Pittsburgh Supercomputer Center have partnered to provide simulation software and HPC resources to academia.
- Compatibility and Accuracy of Mesh Generation in HyperMesh and CFD Simulation with AcuSolve for Torque Converter (PDF)
- Typing on a tablet can be done, but not well. Doing CAD work on a tablet is even more problematic. Enter Orange Juice Studios and their patent-pending OMouse which has the potential to make CAD drawing on tablets [and maybe mesh generation?] easier.
Meshing and Geometry
- The Roots of CAD, part 1 from 3D CAD World.
- Here’s a nice overview of CFD grid terminology from Innovative CFD.
- Remember the announcement of Geomagic Spark, their collaborative effort with SpaceClaim on Scan-CAD integration? DEVELOP3D writes about their beta test of Geomagic Spark.
- Read more about precise organic modeling in Dezignstuff’s article about Tsplines, mesh modeling and NURBs.
- Cheng, Dey, and Shewchuk wrote a new text about Delaunay Mesh Generation.
Have you purchased your copy of the new Delaunay Mesh Generation book by Cheng, Dey, and Shewchuk?
Wind Tunnel, Computer, Wind Tunnel
In the early, brash days of CFD it was going to replace the wind tunnel. [Depending on who you are, you are either making an embarrassed smile or a knowing chuckle right now.]
Turnabout is fair play. When you’ve got a souped-up computer (this one for cancer research) and a need for some serious cooling, why not build a wind tunnel around your computer?
Now, if you used that computer to do CFD on the flow through the wind tunnel and around your computer…
Mike Schropp’s wind tunnel cooled computer.
RSS - Posts
I am a little surprised Altair published the “Compatibility and Accuracy of Mesh Generation in HyperMesh and CFD Simulation with AcuSolve for Torque Converter” report because, at least in my view, the results don’t show AcuConsole and AcuSolve in a good light. I was under the impression that Acusim’s technology would be used and marketed as Altair’s flagship CFD offerings.
Maybe I should re-read the report because I didn’t get as much of a negative vibe as you did. I saw it as a snapshot in time as AcuSolve is more fully integrated into the HyperWorks environment.
Maybe not necessarily a negative vibe but the HyperMesh/Unknown FVM solver results matched the test dated much better than the AcuConsole/AcuSolve results.
In regards to “CFD for the Rest of Us”, I think the thing to take away from that piece is “marketing!” For example, regarding bicycle wheels, one could market their standard wheels by say “our wheels have been specially designed to induce turbulence so that the drag of the entire bicycle-rider system is reduced, much like dimples help reduce drag on a golf ball or a vortex generator helps aerodynamics for separated flow.” And, I’m not sure if CFD can really prove this one way or another. 🙂
A quick calculation. My leg is 5 inches in diameter. At a speed of 25mph at sea level, the Re is 97K. It would be good not to have laminar separation on my legs. I should add dimples to my legs!! 🙂
Sadly, there already is clothing with dimples (however it is against the rules for racing). And, Zipp even makes a disc wheel with dimples. Oh well, I guess I’m not going to get rich this way.
Martin: I suspect you’re being a bit tongue-in-cheek but it’s worth saying that CFD can provide sufficient insight during the design phase that with other engineering improvements and some experimental work you can actually come up with an improved design. There’s a difference between trying to nail down the drag on this configuration to the nth degree (which would be really difficult) in any one run and running 100 cases and seeing the trends. As we’ve discussed before, validation and verification needs to be part of the overall process too.
Martin: The key to riches is designing a unique dimple shape (pentagonal frustrums?), trademarking it, and getting a celebrity endorsement. Or selling to the military.
I’m not sure how tongue-in-cheek I am. First, without a doubt, CFD is great. And I’m serious there. I definitely agree one can learn a lot from it and gain many insights. And it is wonderful to see more individuals have access to this tool. However, first and foremost, I do agree with your caveat. It was estimated (a back of the envelope calc. I believe) that the new wheel design would reduce the overall (bike+person) drag by 2-3%. However, if the drag on ones legs increases, (and turbulence in front of a laminar cylinder will reduce drag), then some of that advantage goes away. I have no idea how much. That’s the bummer about aerodynamics, sometimes it giveth and taketh. So I’m hoping Steve is not implying that aerodynamics does not take expertise. Having a little understanding of aerodynamic theory (the ideas of which are easy to grasp [one does not need to go overboard with the math]) sure does help to reduce the amount of trial and error. (BTW, I’ll admit, when I read some of this marketing stuff I just can’t help but get a little tongue-in-check! :p )
Hahaha, OK, I’ve got to share this. Sometimes I feel like such a novice. Given this discussion I decided to try running a cylinder. Anytime I do these “fun” things I use my own code. It’s an implicit structured (Chimera) central difference code with 2nd order spatial stencil and up to 2nd order time. It does not have things like dual time stepping, higher order differencing, nor flux splitting. So, the code isn’t bad, but it’s not state of the art either. Well, at least not yet. So I ran a 2D cylinder at Re=10K, M=0.1, and 2nd order time. I did not use a turbulence model. I also didn’t run 3D since I don’t have the resources to allocate to such “fun” things. The farfield boundary is far away, so that is OK. But, in hindsight, I did make one gridding mistake. My fine wake grid only goes back 20 cylinder diameters. After that it expands (coarsens). Overall I have 650K grid points. So here is what I goofed. Once the vorticity from the cylinder tracked out of the fine grid area it started to dissipate. OK, that was expected. But, I did not realize how sensitive the solution was to that. When the vorticity dissipates, the wake has an “end” and will curl up to some degree. This then feeds back to the cylinder as lift. My lift curve oscillated sort of sinusoidally between 2.0 and -1.5. Thus my average lift was .25. So about 2.3 deg. alpha for a 2D airfoil. Every so often a hiccup would occur (for whatever reason) and a vortex would delay separation thus creating a gap in the wake. This gap would create a new end once it tracked into the grid region which transitioned to the coarse grid. The lift would then oscillate between 2.0 and -1.5, or 1.5 and -2.0. All dependent on the flip of a coin. So, I need to make my fine wake region longer if I’m to get a time averaged lift closer to zero. And, yes, it is all about V&V. LOL, I sure am looking forward to the day when I don’t need to do it! That takes brain cells. I just want to hit the run button and go grab a beer!
Oh, I should add this. My average drag is about 1.65 and experiment is about 1.2. I know I have too few grid points. I doubt I’m close to grid convergence. 2D grid convergence should bring me closer to 1.2. But not all the way since this is 2D. It probably will require 3D to get that extra distence. But, I’m not really sure what to do about a turbulence model at Re 10K. I think the real life RMS of cl is about 0.6, but I’m not really sure. As Re goes up I think the RMS goes down… (?) Such are the questions… And this is just a cylinder! Arg.
Do you think the feedback from the fine-coarse grid transition in the wake was exacerbated by it being 2D? Or do you think a turbulence model is a bigger culprit?
Hmm, trying to get a handle on the question. Would a similar question be “Does one need such a long wake grid for 3D motion” The answer is, I believe, no and yes. No, because the dissipation may kill off the vorticity before it gets to the end. 2D has less dissipation.
Yes because… (I think) The 3 dimensionality of the vortex is dependent upon vorticty interacting and creating dissipation. Without dissipation they will remain 2D. However, for dissipation to occur the vorticity must be close enough to interact. If the 2D vortex tracking ends early because of grid issues and/or low spatial discretization, then the vortices are spread out too far. (The wake end whips them around) They will not form a tight wake. (Or do vortices need a chance to pair up?) So the 3D process will not occur, or will be handicapped. Also, the longer wake grid probably needs to be kept even in 3D mode in case there is a “hiccup” and the vorticity momentarily goes back to 2D. That’s the pain about this. I think. The flow is either 2D with “little” dissipation or 3D with “much” dissipation (“burst”). I’m not sure there is much middle ground, except maybe where in the wake things jump to 3D. I’m not sure about where the jump will occur either since the pressure does feed forward. So maybe once it pops it feeds forward to the aft end of the cylinder.
By turbulence model, do you mean dissipation model? At these Res there are two aspects to the turbulence model, 1) vortex modeling and 2) dissipation. I gather the dissipation modeling will only enter the picture when the vorticity start interacting closely. Sadly, I’m not there yet. 🙁
Did that answer the question?
BTW, I have no real idea about how errors in this translate into errors in integrated pressure loads. That’s kind of what I am (was?) trying to figure out. 🙂
Oh a caveat. If a dissipation model is used which creates a lot of dissipation, regardless or whether it is close to another vortex or not, the vortex will break down into 3D mode. But, that’s not really physics. It may be OK for Re of 100K or even 10K, but 1K? I’m picking on the model. Just knowledge must be had about its performance envelope.
Oops, I meant to say “I’m not pick on the model” Arg.
Sigh, I’m typing to fast. 🙁
OK, an update. My fine grid now goes back 50 diameters. Again, this is a 2D laminar (i.e. no special dissipation (turbulence) model). A noticeable pattern emerged and it possibly does not bode well. To lead into this one must know that the force on a vortex is equal to V X Gamma. For low Re cases this forces the vortices shed from the top to track up and the vortices shed from the bottom to track down. However, at this Re the wake comes off tight from the cylinder (and the vortices have more spin) and the top vortex is able to grab the bottom one (and vice verse) and swing it around. When this happens the forces push the vortices together. And then they seem to stick together. Unfortunately the overall load on this pair is somewhat zero (the vortices that make up the pair are counterrotating). The pair then just float off in some direction. It’s kind of funny looking. 🙂 They do not form a straightish line behind the cylinder. And, there is no reason they should as long as they have not dissipated (i.e. gamma goes to zero) or there is a delta V on them.
So what’s the problem? Sigh, the turbulence model. If SA-DES is used a choice needs to be made whether rotation curvature correction (RC) is used. If SA is used then the dancing pair will have a lot of dissipation. If SA-RC is used they will not. This of course assumes the pair can get away from the cylinder. The back side of the cylinder will have a lot of dissipation in either case since the vortices are rubbing against each other (shear layers exist). I searched the web and apparently SA-RC does give better cylinder loads. Also, SA-RC is the “better” model. However, how well the wake is modeled is unknown. It’s also hard to determine what people are doing in regards to whether they are choosing a turbulence model with rotation curvature or not. And, in general, it seems like they are not using RC. But, I’m not sure. I need to dig around more and look into some manuals. I did post a question about the selection of the turbulence model at cfd-online and the Computational Fluid Dynamics Group at LinkedIn. And this is important because it affects the width and velocity profile of the wake. Which affects the loads on an object behind it.
Of course all this was just poking around, my intent was not to turn it into a research project. And 3D will be different and expensive. I don’t want to go there until I know what is going on for simpler cases. And, 3D is complex. The backside of the cylinder has 2nd, 3rd, 4th, etc. vortices and they will get ripped off with the primary one and spun/twisted around. However, as ugly as it looks (i.e. bizarre 3D shapes), if there is minimal dissipation, they will not go away, in the sense of dispersing the angular momentum properly.
So the next step is to do more homework and then go on to wake comparison of 2D SA-DES with and without RC. Unfortunately 2D SA-DES will need to go on to the back burner. 🙁
I’m open to any thoughts or suggestions from anyone. 🙂
You’ve gone well beyond my knowledge of turbulence models so I’ll leave that topic for others. But you point out the importance of studying model problems. You could probably do a version of the Drag Prediction Workshop just on 2D cylinders.