(1). A much simpler question,: Given a problem, a group of users each run the same commercial CFD code, but each with a different set of turbulence model and numerical method, do you think that they will get the same results? if not, will the results qualitatively the same? (2). If next day, each repeats the same case with the same options and boundary conditions, except the mesh arrangement (keeping the total number of ponts or cells the same), do you think that the results will be the same as their previous resluts? or qualitatively the same? (3). If they refine the mesh and each obtains the mesh independent solution, will this solution the same as his first solution? or qualitatively the same? (4). Will all the mesh independent solutions converge to the same solution? or qualitatively the same ? (5). If a client is looking for the solution of this problem, which solution is better for him? My feeling is without validation, we will have to invent the "equivalent Uncertainty Principle in CFD", which states that quantitatively correct solution is practically impossible , thus, in reality, multiple soultions exist for any CFD problem. (6). I think, I will be the first one to use the "Equivalent Uncertainty Principle in CFD" to characterize the current state of the art of CFD.

Everyone talks as if validation by an experiment is the gold standard.

If you gave n groups of experimenters the same physical data to collect in different rigs using different instrumentation/ methods of detection how much variation would you see. If everyone achieved the same answer heat transfer textbooks, for example, would be somewhat thinner. To do good experimentation is at least as hard as doing good CFD.

At least in CFD you should be aware of the many limitations in the numerical techniques, even if you cannot necessarily quantify them in a particular situation. 'I measured it so it must be right' is a common result of experimentation.

Many experiments of real life engineering type problems are made with many assumptions in them, i.e. not reynolds number matched, wrong flow boundary conditions, geometry simplified,... This is often required as it is not possible to run the 'correct' experiment within a reasonable budget/timescale. Care therefore has to be exercised in interpreting the results and drawing conclusions. Generally subscale experiments have to be validated on the real device which, depending on the industry, may not be possible until the first production unit is in place. Not ideal but often reality.

In these cases CFD can provide a useful second avenue of investigation. Often it points out the blindingly obvious that no-one saw until the CFD guy put all the pieces together and put some flow through it. The real advantage CFD has is great instrumentaion and the static nature of the data set. The answer may or may not be right at any point in space but at least you can query the solution whenever and wherever in space you want. If you run a $100,000 test and three weeks later you spot something unexpected in the data or that you needed more instrumentation it is often very difficult to re-run the test.

Additionally, you can return to the analysis possibly years later as the result of a field failure and often detect at least some probable causes, not spotted or appreciated at the time, that can at least help jump start the investigation. This is typically less true of testing as the sparseness of instrumentaion generally means that only items of interest at the time were measured.

Overall engineering is a matter of getting together the best available, but usually inadequate, data you can and making the best decision possible. If you are wrong and something unexpected happens, get over it and fix the problem by trying to understand what went wrong and how to do better the next time.

John, you make it sound as if wind tunnel testing, prototyping and the like are the only ways of subjecting your results to a critical evaluation. For sure, they aren't, and I very much agree with Greg in that such extensive validation is not always needed. Admittedly, I would certainly not fly with a plane knowing that it hasn't had its final design thoroughly validated by the means you are referring to. But on the other hand, I wouldn't leave a plane in panic if I learnt that a number of preliminary designs were rejected after CFD analyses without such extensive result evaluation. After all, a very important task for CFD must be to reduce (but not eliminate!) the experimental work required to arrive at an optimal design. So much for applications of hypercritical importance for safety. These must of course be subject to the most stringent quality demands, and that's why the aerospace industry, which has used CFD longer than any other civilian industry, still rely on wind tunnel tests and exhaustive test flight programmes to ensure the safety of their products.

However, there are other industries than aerospace and defense, and most of these do not have access to anywhere near the resources of the former ones. Does this mean that they should refrain from applying CFD altogether? Hardly, because we have to see the use of CFD in these industries (AND how it is used) in connection with what methods they have relied on earlier. For instance, if the flow distribution in a cooling water network with complex 3D geometrical features has previously been calculated by means of a 1D software, spreadsheet calculations with guessed flow coefficients, comparisons of cross sectional areas, or even pure guesswork (experience), there is little reason why CFD should do significantly worse, EVEN if the reattachment length after a sharp corner is not right. And if the corresponding heat transfer coefficients have previously been estimated by assuming fully developed pipe flow and applying for instance the von Karman analogy, CFD would definitely offer more accurate data. Of course, in the latter case, a near grid independent solution is required for a significant improvement to be achieved.

Looking at other areas, such as burning fuel sprays, the models and methods available all have serious weaknesses and extensive result validation and model tuning is required to obtain accurate quantitative results. However, if you know next to nothing in advance about what really happens inside your engine cylinder, less than perfect results may still be extremely valuable in gaining increased understanding of the processes in question, and this understanding can always be put into practical use in directing experimental efforts and thus aiding the technology development of your product. It is anyway never a question of basing the development of a product solely on CFD analyses. Knowing the weaknesses of CFD, as of course every CFD analyst should do, no responsible engineer or company would allow that.

The bottom line is: while we all strive for perfection, we do it within a framework. Some of us do it in an aerospace company and may have the possibility and requirement to do so solely within our own discipline. Others have more narrow frames, working for a company with less resources and/or being at a less mature stage in the application of CFD. In such a situation, the task is to take whatever CFD can give and put it to practical use, naturally in a responsible and careful manner, to add value to the final product, not for the sake of CFD itself.

If more companies start using CFD, the discipline will mature faster than otherwise. Some companies will fail to use it wisely and may die, but no company should be irresponsible enough to allow hazardeous designs to enter the market. In that case, the quality control of the entire product has failed as much as the CFD analyses that may or may not lie behind such disasters. It is our responsibility as employees and professionals to contribute to avoid this. However, we have no responsibility whatsoever to discourage the use of CFD in new industries or areas just because they posess less resources or experience than the already established ones. That will potentially deprive, not only individual companys, but eventually the entire society (as well as the environment of that society) of significant benefits.

Thanks for your time.

Lars Ola

All of this discussion should be underpinned with the question every engineer using CFD should ask themselves prior to conducting a CFD simulation:

"Why am I about to perform this simulation?"

If the above strikes you as obvious then you'd be surprised by the number of people who get so involved in the 'how' they forget the 'why'.

The above question may be answered by any of the following:

"To make a better design"

"To ensure the design conforms to no more than 4.21 bananas"

"To make the best design"

Each question will determine the type of numerics/approach to be taken when performing the CFD. In applications where the BCs are only known to 20ish% and the above first answer is required then CFD will be a more than useful tool NOW. Working in the aerospace industry and striving for the 3rd answer...well is CFD there yet (on industrial time scales?).

[A litre of virtual brandy for anyone who can figure an application where the 2nd answer is required!]

Fred.

But there's one part of the process which you didn't mention directly and is most often the Achilles heel of all CFD results: the grid induced error. All good discretization schemes should converge to the exact solution as the grid is refined, but for a coarse mesh, and most meshes out there are very coarse in my opinion, there can be significant errors induced by the grid. It is up to the user to assess how big the grid error is.

One way of doing this is by computing the error on different grid levels and using the Richardson extrapolation to obtain the relative error on the finest mesh. But it seems that few pay attention to this, or simply do not wish to pay attention to this study as they do not want to find out just how wrong their answer is.

I'm amazed at how little importance is given to such issues in the scientific litterature. It seems that for experimental results, it is a requirement to estimate the error in data gathering, but such error estimation is not judged as critical for CFD results. Would anyone really believe that for a complex 3D problem, it is possible to achieve grid independancy on modest hardware?

How big do you think is your grid induced error? 5%? 50%? 500%? I have seen some mixing problems where the error on the mixing efficiency went up to 500% for a mesh which was judged at first adequate. Afterall, 1 million nodes seems like an awful lot of nodes.. But for certain problems, it wouldn't even qualify as an extra-coarse mesh.

Hi all,

Following on from the previous discussions, I really suggest that you folk visit: <www.featflow.de>

for some very interesting perspectives from the mathematicians viewpoints.

I personally have been fascinated by the studies their team have undertaken into algorithmic design, influence of viscosity, mesh details etc.

It is good reading.

Best regards,

Des Aubery...

(1). I am begining to see the light from the other end of the tunnel. (2). On one side, as long as there are computers around (the computer will be faster in the future), people will conduct "CFD simulation" because of curiosity alone. (3). On the other hand, if one is running a 3-D problem using an old and small workstation, then more problems will be created. (4). And in most cases, the industrial problems are complex in physics and complicated in geometry. Given a problem, has anyone tried to estimate the required mesh density and the corresponding computer hardware resources? (5). Have you seen a single tutorial example ,from any commercial CFD code vendor, illustrates the mesh independent solution and the process? For most of my 3-D calculations in the last several years, the mesh independent solution status was not achieved. (naturally, because of the lack of support, and the hardware is just one element.) (6). Validation in this case, means that the mesh independent solution issue is addressed. If the engineer can not obtain a reliable and repeatable solution because of the coarse meshes used, then something must be done in the year 2001, which means the hardware has to be upgraded first. (with many companies, this is not easy, but not impossible. in some cases, it is going to be very hard, because they used to look at the computer requirement from 1-D analysis point of view. And suddenly, we are back to the super-computer era.) (7). For a simple 3-D flow through a turbine blade passage, (not a stage with a stator and a rotor), it will take 2,000,000 mesh points to get a reliable solution.(it is about two GigaRAM when a commercial code is used) This may not be adequate, because the loss is evaluated at the exit plane and the secondary flow tends to move the high gradient regions of wall boundary layer away from the wall. And therefore, higher mesh density will be required also away from the wall. (8). In testing, the accuracy of test data is not a function of the total number of data points taken. In CFD, the accuracy is directly linked to the total number of mesh points (or cells) used in a calculation. (9). And if the performance of the system promised to the client is based on a coarse mesh calculation, the failure to meet the contract requirement later is likely to cost the company millions of dollars.

(1). Thank you very much for the information. I have visited the website several times already. (2). I think, they are working on finite element method in incompressible, laminar flows. It is probably easier to identify the Reynolds number used in their sample output, instead of using the terms "low Reynolds number", "medium Reynolds number" results. (3). And also there was no information about the mesh size used. So, it is hard to make any comment at this stage.

John,

I seem to remember Boeing making a big thing out of designing the 777 aircraft completely on a computer. The aerodynamics were done using CFD. The final configuration was wind tunnel tested, but only that configuration.

Last year, I managed to get from London to Atlanta on one of those birds. Of course, it's as cramped and packed and uncomfortable as its predecessors (at least with the seating package that Delta flies), but it did finish the flight successfully.

(1). First of all, Boeing commercial aircrafts company has been leading the CFD applications in aircraft design for a long time. There is no question about their experience and knowledge in aircraft design, or CFD applications. And Naturally, they also have been using super-computers and had a dedicated CFD laboratory to support the activities. By the way, they also have their own wind tunnels. Their safety record has been excellant. They have been using CFD long before 90's. (2). Here, we are mainly talking about the commercial CFD code user's community. We are not talking about the defense industries either, because they have their own source of codes. (3). By the way, I have seen a story related to the Boeing's CFD design application in one of the PBS TV station. Although the original idea was consistent with what you have mentioned, in the end, they did not abandon the flight test and was lucky to identify the problem in time. That's the only portion of the story I still remember. I did not pay attention to the particular model involved in the PBS story. (4). On the military aircrafts side, it is a completely different story such as a new fighter development or X-series experimental aircraft design. There, high performance requirement tends to push the technology to the extreme edge, where the test pilot is the only one to identify the potential problem. (5). But, as in the recent case of tire problem on a SUV, the problem can happen even in a hundred-year old company. ( this is because the company is no longer the same as the original company.) So, part of the responsibilty to safeguard the quality of a product is on the consumer side. And hopefully in the case of safety issue, the government should also actively involved, as in the commercial aircraft business. (6). The impact of CFD on product quality is so broad, that we will have to address it on the individual cases. That is, the qualitative CFD solution of thermal-fluid flow simulation in a PC case or board design may be acceptable because of the short product-to-market time involved, and because it is not a safety issue. But even in this case, I had identified the cooling and overheating problem of my son's PC and actually had helped him to remove the portion of the front lower casing and installed a fan to add the cooling capacity. With this, it solved the overheating problem of the PC. (7). If you have the time to visit the computer store, you can take a look at the PC casing. You will be surprised to see that most PC casing do not have adequate inlet area for cooling. And the mass flow through the PC is severely limited.

Hi John,

Thanks for your reply.

FeatFlow is indeed using Finite Element on Incompressible Laminar Flows.

The mesh studies are more interested in element aspect ratios, but actual element sizes can rather simply be determined from their mesh files.

There are also a huge number of articles available for download - lots of reading... I have been at this for some 6 - 8 months already... )

Best regards,

Des Aubery...

(1). Turbomachinery CFD is an extremely complex field, relative to the steady external flow over wings and fuselage. (2). The flow through a turbine blade passage is three dimensional, and the inlet condition is usually non-uniform. In addition, at off-design conditions, flow separations can move forward into the inlet station, thus making the inlet condition hard to specify. (it must be moved even further ahead of the inlet of the radial pump.) (3). In the radial turbine case,the flow in the scroll and through the nozzles can be fairly complex and non-uniform. This becomes another issue, when calculating the flow field of the radial turbine balde passage. (4). And then the coupling of the separated axisymmetric diffuser flow and the turbine flow field under certain conditions force the calculation of the turbine and the diffuser flows coupled and together. (5). similar situation exists in the axial turbine flow field, where the secondary flow field is so complex that it is very hard to develop consistent design model for the efficient turbine blades. (6). The effort to design efficient 3-D advanced blades is also hampered by the inability to predict accurately the total pressure loss across the blade row. As a result, the performance must be obtained from the testing, which can not identify the exact source of the loss mechanism. As a result, the test result is useless in creating a more efficient blade design. (7). So, in the turbomachinery applications, grid generation must be improved to give the adequate resolution, the turbulence model must be developed to provide the accurate total pressure loss, and the inlet, exit boundary conditions must be able to take into consideration of the non-uniform conditions, if we are going to make any progress in the future. (8). Without these improvements and understanding of the physical total pressure loss mechanisms, the new blade design results will likely be inconsistent with the test data.

Firstly ,well done for concentrating minds in this overall scenario covering all future aspects of cfd.

Your comments on the turbomachinery field ring very true to technical papers that I have read and referenced. As you know I am reviewing rotodynamic pumps but find that manufacturers do not have much material available for public consumption. There is dated stuff on some aspects notably-rotor/stator interaction to much more powerful parallel computation ,leakages,circulations to forces on blades and gains from noise reduction in flow -that some users claim successful cfd contributions.

Fed97-3374,Fedsm97-3328,Fedsm-3373 ARE CERTAINLY WORTH A VIEW .Another paper which you highlighted to another turbo question-5th european sgi/cray mpp workshop is another [turbines here].Also epfl-scr n09 -capa dealing with unsteady simulation-diffuser interaction in cent pump yet another.

My educational site on pumps is now at http://fortunecity.com/meltingpot/broadway/1211/

Anyone wishing to offer references ,papers etc are most welcome to email.

Please keep up the good work and hopefully cfd workers not hit too hard from GM recent news.

(1). Thank you for the information. You seem to have changed your webpages and website. (2). I am getting the runtime errors when visiting the website. There are only text, but no figures. (3). I am using MS Internet Explorer.

I have the same problem at present but normally both explorer and netscape put out all information-possibly sunday maintenace by the servers.

Sorry for inconvience -hope it resolves itself.

My old site still exists but with scrolling [about 1500 visitors] was not too friendly I`m told.

I can`t retrieve my email either presently. Hopefully the gremlins go away.

(1). The Boeing aircraft you just mentioned is making big news on TV. It is making the passengers sick in the long distance travel, as reported by the TV news in LA. (2). The other Boeing military aircraft V-22 is also making big news in the capital hill. The aircraft had crashed and killed military personnel, but the news was related to some other issues. (3). So, I think, there are problems in both the commercial side as well as the military side. I think, both the commercial market and the military budget will have long term impact on aerospace companies world wide. It is the kind of business, which depends strongly on the government programs, whether it is Boeing, Lockheed, or Airbus. (4). CFD can make the product safer and perform better, but in reality, CFD is always the first to cut in the merger or budget problem. It is something they think they can do without affecting the profit. I can tell you that there were already a lot of real problems in turbomachinery business, and I expect to see more problems in many other areas, commercial and military. It is simply because many companies today do not have good experienced engineers working for them anymore. (5). The technology accumulated through 60',70's and 80's are all but disappeared in 90's in US. So, I would expect more real problems appearing in the next 5 to 10 years. As a result, there will be real wars in both economic and military areas.

Just a note on your comment about the large number of DOF needed for accurate calculations of seemingly simple flows.A lot research here at Duke at many other research institutions are looking reduced order models of such flows which give very accurate results and reduce DOF by an order 10^2-10^3. While this may be passe in the future when computer time is not at a premium it certainly makes life easier(while providing me with research ) at the moment and provides an interesting look into the behavior of fluid flows.