Thoughts on Successful Implementation of a Computer Model
A few months ago I posted a discussion thread to the Reactive Flows and Chemical Kinetics Group on Linkedin, a professional networking site. The discussion was titled: A controversial topic – modeling without experiments.
I asked group members if they thought that in the next 30 - 100 years engineers would be able to rely solely on computer modeling without conducting actual lab experiments to verify the computer results. I figured that by looking far enough in advance, several generations into the future, I would surely get some responses along the lines of:
“In about 50 years pretty much everything will be computerized. Computers will be built into our brains and cell phones will be implanted into our ears. If we are able to modify our bodies in such a way, something as comparatively simple as combustion inside of an engine or reaction on a surface of a silicon chip could surely be modeled entirely on a computer. It is only obvious that actual laboratory experiments will become obsolete.”
Boy, was I wrong.
First off all, thanks to all of you who responded to that discussion. It is always great to hear ideas and opinions from people all over the world. That being said, I found it interesting that the responses from the group members to that discussion post could not have been further from my thinking.
Here are some excerpts from the posts:
“Of course we will probably see some periods where [computer modeling] will suffice…but probably just for a while.”
“I do not think that in 100 years anyone will built…a plant without running the experiments first.”
The prevailing theme in the comments to this discussion post was that some laboratory work will always be required when designing engineering systems. Whether it is 30, 50 or even 100 years in the future.
These responses left me rather confused. We managed to figure out how to send humans to the moon in a span of about ten years. Yet, somehow, we will not be able to extract the laboratory from the experimental process when designing and building stuff a hundred years from now?
At the SAE World Congress last week, I attended an interesting panel discussion in the ATX Theater. The topic was: What lies over the horizon – a forecast for the economic/policy climate. The panel was composed of diverse number of speakers. One of the panelists was Richard Goetz from Dykema, who spent 32 years working for Ford. I wondered what the speakers thought about the future of computer modeling in terms of the automotive industry. At the end of the discussion, I asked if the panelists foresaw any growth in computer modeling in the automotive Research in Development groups.
Richard’s response was, and I quote:
“There will be massively more computer modeling. That is the ONLY way that the industry can move forward.”
I try to stay in touch with as many software vendors as time allows. Sometimes that involves showing up at company events, taking training classes or just picking up the phone, calling a vendor up and asking them what’s new with their latest product. A number of computer modeling development companies, including CD-Adapco and Maplesoft told me that they have seen a healthy growth in revenue this past year, which is obviously an inverse trend to the rest of our economy. Representatives from these companies told me that since everyone is trying to save money, computer simulations are replacing more and more of the costly laboratory experiments in the industry. And the software vendors are happily observing increase in their profits.
So if the experts in the field and the software vendors are predicting and seeing the rise of the computer use in research and development, why is there still such skepticism from actual engineers doing the R&D work?
Obviously there are a number of answers to that question. If you think that you might have an answer, please post your response in the comments section. We would love to hear it. Here are my two cents on why so many of us researchers are so reluctant to let go of the idea that one day, perhaps far in the future, a Bunsen burner and thermocouple might no longer be needed in order to design an new engine.
Creating and running computer simulations of large realistic systems is not an easy task. I think that one of the main reasons why this is so, is not the limit in computational power. Nor is it the fact that engineering problems are complicated. Nor is it a problem of not knowing all the parameters, variables and assumptions.
I think that it is mainly because a standard and systematic procedure has not been established for taking a real engineering system and converting that system into a set of computer simulations. Sure, there are rule-of-thumb guidelines that various research communities have put together in order to get some meaningful results out of computer modeling. I would argue that it is simply not enough.
If you are conducting a lab experiment and drill a hole in a wrong place in your combustion cylinder – you’ll find out that you have a problem when you burn something inside that combustion chamber. Obviously this is a very simplified case, but with computer modeling there are just too many answers that make sense, even though they might be physically impossible. In a lab, physically impossible things simply do not occur.
So the next obvious questions becomes:
“What would this approach of converting an engineering system into a set of computer models, look like?”
Well, I promise to address that point in the future posts. Meanwhile feel free to register for this blog and post your thoughts. I would love to hear from you.
Thanks for reading,
Masha
While I acknowledge that computer simulations are increasingly useful in predicting results of experiments, I do not believe that they will ever completely replace the hands-on scientific method. We must remember that computers are only as “smart” as the people who program them. Reality is much “smarter” than any of us can ever hope to be.
I would have to disagree with the majority of the respondents in your earlier poll. In many areas, computer sims have already replaced testing.
I have been using CFD sims for nine years and after gaining confidence and establishing correlation with wind tunnel testing (static and rolling road) and track testing, took a big leap 4 years ago and designed a complete NHRA Funny Car body entirely in CATIA V5 utilizing Exa’s PowerFLOW CFD code. Post production verification testing showed our sim and wind tunnel data correlated within 2% in drag and 4% in lift. That body design holds the world speed record of over 334mph in the 1/4 mile. I have used CFD to design entire brake cooling systems with post production testing showing airflow correlation within 1-3%.
We have also used vehicle dynamics sims in the same manner for lap time predictions and to determine initial vehicle set-up with great success. While we still use wind tunnel testing to develop the ride height maps, the program uses the ride height map force data to iterate the true vehicle attitude during a lap. Our track testing shows very good correlation.
The simulations are only as good as the physics and coding behind them. Often, parameters that we do not understand contribute to the need for testing. For example, professional drag race simulations are one of the hardest to develop, even though it would appear on the surface to be a simple F=MA problem. The problem is that accurate tire models for a drag race tire are not available. The tire does some very strange things right at launch and then grows in circumference during the run and takes on a ‘D’ shape, and then ’slaps’ the ground to increase the normal force, increasing tractive effort. As soon as we have enough test data, we will be able to add these variables to the sims and have a good predictive tool.
Consider the huge skyscrapers now being built. They are designed and developed entirely on the tube utilizing FEA and CFD because they were correlated to experimental data in the past.
I feel that once correlation is developed for particular problems and the full underlying physics are understood, and included in the code, simulations can be used entirely in place of testing. Just like testing, they have to be run under the correct conditions, and experienced judgement must be used to interpret the results.
This is an interesting argument. Do we give into the fact that computers are faster and more efficient than humans or do we resist it?
Does giving up lab work for the sake of conducting experiments in a form of computerized models necessarily mean that we are giving up our creativity and uniqueness as humans, to a computer? I am not sure..
Normally we do testing to understand the physical phenomenon involved. If we understand then we try to represent it mathematically for computer modeling & simulation.
However the real challenge lies in understanding a physical phenomenom and then represent it mathematically. For example to get the material properties of a new material (Be it determination of viscosity, conductivity, melt characteristics, chemical kinetics etc) we have to rely on tests because of the difficulty of representing the properties mathematically from some existing data.
So the actual modeling starts only when we have some physical understanding from the tests. The areas where we have understood the physics from experiment and are able to formulate mathematical models of it, we are making good progreess in those fields using computers.
However there are still so many areas where we do not understand what is happening. It took aerodynamists so many years just to understand how bumblebees fly because mathematical models were not producing lift sufficient to fly. Scientists now understand the phenomenon, but without testing it would have been impossible to formulate the mathematical model of how bumblebees fly. So we have no option but to do more and more tests to understand the physics and then try to formulate a computer model of it (If possible). This is really an enormous task and hence it is quite impossible to do everything using computers.
The thing is that if we dont understand something, then computers also dont understand because its we who have to make the computer models. That is what I think but not sure.
I used to run a CFD / process modeling group in an R&D group in the chemicals industry. I’ve also developed and validated complex kinetic CFD combustion models for ‘non-standard’ applications. From my experience I think the following points are most noteworthy:
* large investment decisions tend not be made based upon an un-validated computer model
* to validate a complex model, you’ll need to pilot / expt. which also proves the concept / process
* if you’ve a good model of a process, you can save $$$’s in development
* the KEY to successfully using modeling is having the individual(s) running the model fully understand how the model works, its limitations, a good idea as to what the results SHOULD look like beforehand and an ability to relate the output to the physical world - all under the umbrella of good engineering experience and common sense.
Modeling is a powerful tool but one that adds to a toolbox of professional knowledge and know-how - it should never replace it especially when the results are being used for business critical decisions (or many other situations for that point). Unfortunately, too many people assume that the results from a model are right just because it came for a computer - sometimes boundary conditions are input INcorrectly to a model, the computer correctly works it out and before you know it, you’ve crashed your satellite into Mars or something…..
Hi Masha,
You can design most of the mechanical parts of the automobile in the computer, and when you actually build it all could fit together right away. Not so when it comes to chemistry. It is still not an exact science, and I doubt it will be in the near future. A 0.1 percent error in one of the kinetics equations will amplify to a large and unpredictable error, if that process is running concurrently or consecutively with others. This results in a chaotic behavior, which can be predictable to some extent. There is also enormous amount of variables, such as as the location of the fuel inlet in your example. The trick to designing a stable process is to have a good model and design your process with such parameters where the small error does not matter.
All this requires experimentation.
That’s my five cents worth.
It is an interesting argument. While computers have replaced a lot of lab and expereimental work, it is difficult to judge if we can really replace the tests. In many fields, the analysis starts with some assumptions and the success of the simulations depend on how valid these assumptions are. Computational analysis is one step farther away from reality, because of the assumptions involved.
It may be easy to reduce the number of tests, but it will be a while before we can completely replace tests. I was at a NASA workshop about 5 years ago and a lot of energy was being spent on getting a good turbulence model for prediction. A member of the audience was quick to point out that the these issues cannot wait for computational models to get better. Instead of waiting for the models to get better a quicker solution can be obtained by running a test.
Another thought, though the idea is what some scientists and physicists have thought of. They think that the universe is not a physical or material world but a computer generated simulation - a kind of virtual reality. The evidence lies in treating physical phenomenon in terms of quantum mechanics.
Stefan Wolfram, the author of “A New kind of Science” says that any system whose behaviour does not look simple to us is as simple as any computational system and all processes can be viewed as computations.
John Barrow says “We now have an image of the universe as a great computer program, whose software consists of the laws of nature and which run on hardware composed of the elementary particles of nature.”
Actually we are limited by the conventional CFD approach of solving the Navier Stokes equation. To replace CFD with testing we have to move from our conventional approach to the broader quantum mechanical concept of probability. Some work is being done based on this approach.
I also understand that it it very difficult to understand all the complex physical phenomenon and put it mathematically and logically for computer modeling. But may be some day we can treat most (if not all) physical phenomen mathematically (As people thinks it is computable) in the form of a computer model as some scientists have thought of. But we need to do a lot of experiments to reach there.
Thank you all for wondeful comments!
wsviii - I would really like to believe that reality is much smarter than any computer program. But who is to say, as Savankar mentioned in the last comment, that universe is not in itself a kind of computer simulated virtual reality?
Terry - great comment that “once correlation is developed for particular problems and the full underlying physics are understood.. simulations can be used entirely in place of testing” - that is sort of the idea behind writing this post.
Arron - I like your comment “Computational analysis is one step farther away from reality, because of the assumptions involved. ” I think that is one thing a lot of experimentalists tend to forget - the importance of assumptions.
Thank you all for responding and I am looking forward to more comments from you on future blogs!
Stay tuned for this week’s post GT-power vs. WAVE software
Masha,
this is an excellent idea - providing a place where professionals in the chemical kinetics and reactive flow communities can exchange ideas about the prospects of chemical reaction simulation computer modeling.
It seems that there are two main schools of thought here:
1) Empiricists - the doctrine that all knowledge is derived from sense experience - those who are compelled to look at science from a historical viewpoint and thereby evaluate what may be possible based on what has been possible in the past. These people will be the last to let go of “real” experiments. Granted, any reasonable scientist should rely on experimental data, since all good work is based on it.
2) Visionaries - those characterized by fanciful, not presently workable, or unpractical ideas, views, or schemes. Groundbreaking experimental design is not possible without foresight and audacious imagination.
Einstein was a true visionary, in that he used his mind as the laboratory and created brave new theories. His theories, however, would never have been proven without the empiricists, so the truth is that the world needs both of these types of scientists. The visionaries present the possibilities, while the empiricists test them.
Though Newton was proven wrong, with regard to his theory of a perfectly predictable universe, as his work was shattered by Einstein’s Theory of Relativity followed by modern quantum mechanics, he may have been on to something by believing that it was possible to understand all of the physical laws of the Universe. Though quantum mechanics, which is perhaps the most precisely validated modern theory, is based in the recognition of inevitable degrees of uncertainties, the study of statistics has certainly corralled the parameters of this so called uncertainty. This is the first evidence that the Universe may ultimately be entirely predictable, yet, one must accept the seemingly contradictory concept that there are very precise levels of tolerance, with regard to uncertainty. This concept is paramount to successfully designing computer simulations of physical phenomena.
Each time a visionary pushes the limits, it takes time for the empiricists to catch up and validate the predictive system. Though there is a lag in this validation time frame, the process is speeding up with the help of supercomputers, which can crunch countless complex calculations in the twinkling of an eye. These new tools have augmented the powers of human perception and imagination to new levels. As a result, we are in the midst of a paradigm shift in human understanding (augmented by computers) of the laws of the universe.
As pointed out by the predictions of the great visionary and futurist, Raymond Kurzweil, as outlined in his book, The Singularity is Near, computer simulations are improving exponentially. Since its inception, Bore’s Law has proven true, with regard to the evolution of computational power. Though software development lags behind the evolution of raw computational power and is thereby the evolutionary bottleneck, Bore’s Law empirically suggests that computer simulations will continue to improve exponentially.
Historically, when one looks at the difference between the virtual reality human-computer integration in the very affordable Nintento Wii, and the old video game, Pong, one is forced to be amazed in that it took only about 30 years to accomplish. Projecting out 30 years from now, virtual reality will likely be close to indistinguishable from “real” reality, and eventually, perhaps in another 50 years from then, it will be completely indistinguishable from “real” reality. At this point, one is forced to ask the philosophical questions, “What is more important? Which reality matters the most?” The answer is a question of scale. If a computer system employs all of the appropriate assumptions, the simulation is simply a microcosm, or a miniature mirror of reality. This could be perceived as a new dimension. Therefore, it is also reality and is thereby no less “real” or important than matter on a larger scale.
Once these computer modeling systems become more widespread and accepted, there will be no reason to do “real” experiments at all. At that point, to continue to rely only on “real” experiments would be similar to insisting on only communicating with “real” people in person and rejecting the viability of communications via cell phones and video conferencing. In the end, it is the information that is important, not the physical shell. This is because with the proper information, any form can be re created with different physical particles. It is thereby the form that matters, not the matter.
The key to successful implementation of a computer modeling system is ensureing that the propor assumptions are made. To a certain degree, it does not matter if certain minute details are ignored. It still requires the human mind to determin what can and cannot be ignored. But, just like one cannot ignore the usefullness of cell phones, one cannot ignore the usefullness of a computer simulation.
I am hopeful that on the other end of this paradigm shift we will have a sufficient understanding of all of the physical laws of the universe to model it perfectly. I believe that it is only a matter of time before this happens.