Slashdot Mirror
Physical Models In an Age of Computers
Harperdog points out this article "about the Bay Model in Sausalito, California, which was built in 1959 to study a (terrible) plan to dam up San Francisco Bay. The model was at the forefront of research and testing on water issues that affected all of California; its research contributions have been rendered obsolete by computer testing, but there are many who think it could contribute still. Now used for education and tourism, the model is over 1 1/2 acres and replicates a 24-hour tidal cycle in just 14 minutes. Good stuff."
There was also the Army Corps of Engineers model of the entire Mississippi/Missouri/Ohio/Arkansas/Red river basins. It was built by POW labor near Clinton, MS. See what's left of it here. http://maps.google.com/maps?ll=32.30606,-90.316173&spn=0.003922,0.006335&sll=36.977452,-121.987122&sspn=0.118622,0.202732
Computer models may be far more efficient for research but physical models still have a very important role. Children, while far more efficient then ever at using technology, do not have the ability/opportunity to learn from these computer models like a researcher would. I think physical models are still very important for teaching the next generation about different science and engineering principles, like the tide cycle from the article. The ability for kids to physically see and experience the science and engineering topics is what gets them interested and engaged in learning about those topics in the future. Models like this are essential for future education for kids. I mean what kid doesn’t start to wonder about the world around them after going to the local science center?
What's interesting is that real life models sometimes still yields a better result than the computer simulated, but both have advantages and disadvantages.
Especially streaming water is a tricky thing to model entirely right and you can get surprises when you have a complex situation and try it in a physical model compared to a computer simulation.
If builders built buildings the way programmers wrote programs, then the first woodpecker would destroy civilization.
The model cover about a 3000 sq mile area. The Golden Gate Bridge is about 6 inches long IIRC.
You have to see the Bay Model if you visit the Bay Area. There is a fully restored Nike missile silo nearby that is also a must-see of Bay Area nerdy sights.
Give a man a fish and you have fed him for today. Teach a man to fish, and he'll say "WHERE'S MY FISH, YOU IDIOT?"
What model railroad scale is the closest? I have no interest in CA, so I don't know if 1.5 acres makes that bigger than G scale or smaller than Z scale or something in between. The live steamers might want to turn it into a live steam park, if allowed. Around here, the live steam parks are not quite as elaborate as this sounds.
You don't have to have an interest in CA to read the first few paragraphs of the article:
its 1.5 acres replicate a 1,600-square-mile area that runs from the Pacific Ocean to the Sacramento Delta
1600 mi^2 is 1024000 acres, so it's a 1.5:1024000 (or 1:682666) scale if you believe the article.
However, the bay model's webpage tells a different story:
http://www.spn.usace.army.mil/bmvc/bmjourney/the_model/facts.html
Model Scales (Model to the Bay)
Horizontal: 1 foot = 1000 feet
Vertical: 1 foot = 100 feet
Velocity: 1 foot/ second = 10 feet/second
So using their numbers, it's a 1:1000 scale.
I have no interest in model trains, but Wikipedia tells me that Z-scale is the smallest commercially available scale, and is 1:220, so this is a much smaller scale than any model train system.
As soon as we get a 3D printer that can properly print bedrock, dirt, forests, and San Francisco.
I thought that was what the magenta ink tank was for.
I imagine testing of scale (including full scale) models is still used in most industries. Computers help you reduce the number of mistakes and shorten the iterative development cycle so you make fewer elementary errors.
In Oil & Gas, almost all of the major manufacturers have been modeling all their new components in 3D for the last decade. You can have whole departments dedicated to running Finite Element Analysis on these 3D models. It's not as simple as putting in your constraints (loads, fixed and pinned reaction points, etc) and hitting "solve". You still have to make certain assumptions, tailor the matrix (the 3D lattice/approximation made of pins and beams) or make simplifications to the model if you want it to converge toward a solution. The computer solves the cases iteratively. Once you think you're close, you perform a test at scale and verify your assumptions. Then your match it to your computer model and hopefully you're not too far off. For example, things like incorrect friction factors between materials matter a lot when they are amplified by huge forces. The manufacturer of a particular coating may claim it's .07 (if they even tell you). You learn it's more like 0.08 through repeated testing. At this point there is no way of calculating things like that.
Of course, this assumes that you have the time, money, and server time to run such experiments. Most of the time you focus on the one or two critical areas of a design for FEA and use past experience or classic formulas on the rest. Customers still want final tests before product ships, so technically you have one last point to catch a failure/issue before it ships.
For what it's worth, this is the same Bay Model the Mythbusters used in their season 1 episode Alcatraz Escape.
The accurate reconstruction of the tidal effects allowed them to convincingly show that a raft would be unable to reach Angel Island, and that a more plausible route would have been toward the Marin Headlands - before confirming the model's result experimentally.
Any model that is inherently chaotic (read: almost all of them) cannot be simulated on a computer accurately. There will be a cumulative error which will grow extremely rapidly (this is known as the "Butterfly Effect" after a well-meaning but ultimately damaging title of a research paper on the subject). Fluid dynamics is a major headache when it comes to chaotic systems, since the equations are violently unstable at most points most of the time. The only reason anyone can get any work done at all with CFD is that most research groups simplify the equations to one very narrow, very specific range of conditions. Even then, CFD is treated as a first approximation. A physical model of some sort is almost always built, somewhere along the line.
Indeed, one of the unsolved Millenium Challenges is to find out just how bad the Navier-Stokes equations really are. I'd have thought a more interesting challenge would be to find an alternative way to model chaotic systems. (Chaos means that the system is deterministic but not predictable. Which is no different from saying that you have a superposition of probablistic outcomes where the actual outcome is only knowable by observing the system. A field that is of interest to all kinds of people, including physicists and sci-fi writers. A standardized way of dealing with such systems that can yield any additional information at all - doesn't matter what or how little - would have a lot of appeal.)
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
Something I've always wondered about physical models is, how can you compare them to real situations at different scales?
IIRC the folks who design America's Cup sailboats have (had) a rule of thumb that a hull model less than 1/3 scale would not accurately predict the drag of the full-size design. Off-topic: if you think water and fluid dynamics are tough to figure out, take a look at what little we know about how ice skating works -- or to be exact, just how the surface of ice behaves to allow gliding motion.
https://app.box.com/WitthoftResume Code: https://github.com/cellocgw
The days of large physical models of tidal hydraulics in large estuarine systems are past because properly calibrated/validated numerical models provide good results at a fraction of the cost.
However, it is paramount the the numerical models are capable of simulating the correct physical processes without over-simplication. For example, the flow hydrodynamics near the Port of Anchorage in the Knik Arm of Cook Inlet are dominated by large gyres that are shed off prominent headlands. A large physical model of the Knik Arm constructed by the Corps of Engineers at their Vicksburg, MS, research facility reproduced the large gyres with a good match to measured field data, and local tug pilots agreed that the flows resembled what they experience daily. Initial attempts at numerical modeling the flow fields produced no gyres, and it was not until a very sophisticated adaptive turbulence closure scheme was added that gyres formed in the numerical model. Both the physical and numerical models required good boundary and initial conditions for success.
Physical models are still useful for simulating processes that are beyond our ability to describe mathematically (required for numerical modeling). Examples in the field of hydraulic engineering include some sediment transport processes, stability of rubble-mound structures such as jetties and breakwaters, erosion of cohesive sediments, wave forces on structures, and resiliency of levee grasses subjected to wave overtopping, just to name a few.
Numerical modeling in hydraulic engineering is making rapid advances, and often any unknown processes can be adequately represented in the model by empirical formulations that have been developed based on physical model tests. Whatever the skill of the numerical model, it is imperative that engineers who apply the numerical model to a problem have a good understanding of what physics are being simulated and what compromises have been made during model development. Failure to understand what the model does will assuredly lead to disaster.
Finally, physical models can be successful provided: (1) The dominant forcing in the real world is correctly represented in the scaled model, (2) any forcing not correctly represented has minor influence, (3) laboratory and scale effects can be minimized or some compensation can be applied, and (4) model results have been validated to the extent possible. A similar set of criteria applies to numerical models.
In the future, physical modeling will continue to be used to validate numerical models, they will provide physical understanding and empirical formulas for use in numerical models, and physical models will continue to address those engineering problems that cannot be formulated mathematically.
Any model that is inherently chaotic (read: almost all of them) cannot be simulated on a computer accurately...Indeed, one of the unsolved Millenium Challenges is to find out just how bad the Navier-Stokes equations really are.
Is that really a chaotic system problem, or an incomplete model problem? Taking the Navier-Stokes equations as an example, they actually are surprisingly accurate at describing turbulence, even though, as you've mention, we haven't proven that they describe turbulent flow properly in all cases. So typically problems with fluid flow models come from making too many simplifications, having incorrect initial conditions, or making bad assumptions, not a problem with the Navier-Stokes equations per se.
In fact, as far as chaotic systems go, wouldn't a physical model have the same problems as well? The challenge there is a very high sensitivity to certain variables, so your physical model of the system may turn out to not scale at all.
The Navier-Stokes equations are definitely chaotic, but turbulence is itself chaotic. It's actually a wonderful example of it.
A physical model would have the same problems, yes, since you can't scale atoms and the sensitivity to initial conditions means that given a long enough run (which is going to depend on the exact nature of the system) the cumulative error will swamp the system. The advantage of physical systems (for now) is that both the step size and the particle size under consideration are considerably smaller and modeled to a far greater level of precision. Physical models are still imperfect and can lead to all kinds of false assumptions if relied upon too heavily, but as the step before the full-scale system, it's the best we currently have.
A 1:1000 model essentially treats a block of 1000 molecules on the full scale as being the same as 1 molecule on the small scale. This means that at 0'C and 1 atmosphere, such a model would consider 2.687 x 10^22 (remember, you're considering 1000 at a time) molecules of an ideal gas per cubic metre of gas under consideration. In comparison, the world's fastest supercomputer can perform around 1 x 10^16 FLOPS and the cartesian form of Navier-Stokes looks to me like you're going to need to perform around 24-25 floating point operations per iteration per molecule. So you're looking at around 600 million seconds per cubic meter of simulated gas flow to get CFD equal to a physical model, if no simplifying assumptions can be made, and that's only true if thee world's fastest supercomputer's FLOPS rating is with a level of floating-point precision great enough not to introduce rounding errors within those 600 million seconds.
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)