November 1, 2010
Written by Professor Jeffrey Bennighof
You may have noticed that noise levels in cars have been decreasing over the years. One reason is that car manufacturers around the world have been using vibration analysis software written here in our department.
The software uses an approach that has roots in our department dating from the 1960s. Our own Professor Roy Craig, who retired in 2001, played a key role in developing the “component mode synthesis” (CMS) method. In CMS, a finite element model of an airplane, for example, is divided into several substructures (wings, fuselage, empennage, etc.) so that models of substructures can be reduced before assembling them together to form a more manageable model of the overall structure.
Our software implements “automated multilevel substructuring” (AMLS), which takes the CMS approach to an seemingly ridiculous extreme. In AMLS, a finite element model of a car with about 10 million degrees of freedom is divided into substructures, and substructures of substructures, and so on, until the model has been divided into tens of thousands of substructures on dozens of levels.
Models of individual substructures are reduced greatly to produce a model of the overall structure that has about 1/100 as many degrees of freedom as the model had originally. Then this reduced model is used to approximate about 10,000 natural frequencies and modes of the structure, so that vibration can be analyzed efficiently.
In CMS, model reduction is done manually, but in AMLS, the substructuring process is entirely automated.
The trick is to do it very efficiently and without losing accuracy.
Using AMLS for vibration analysis in cars had its beginnings in a research project focused on submarines. In the 1990s, we were funded by the Navy to develop a substructuring method for predicting submarine vibrations. Submarine funding decreased after the end of the Cold War, but we began to see how a new method, which turned out to be AMLS, might work to handle very large vibration problems. At about this time the auto industry began creating fairly detailed models of entire cars so that they could be analyzed and made quieter while they were still “on the drawing board.”
Using the most efficient solver algorithm available at the time, car companies required extremely expensive liquid-cooled Cray supercomputers for vibration analysis. Ford, for example, was spending over $10 million per year on computer time to analyze vibrations. To explore future capabilities, Ford created a reasonably detailed model of a car body but found that their fastest computer took four weekends, when it was not carrying its normal workload, to get the results they wanted.
When a manager at Ford learned that we might have a faster way to solve their problem, he arranged to get us access to their supercomputers. We were able to run their 4-weekend problem in a few hours on a Tuesday afternoon when their machines were busy with their normal job queues.
That got them interested, and they began to fund us for a few years to get our “research code” ready for production use. They also encouraged us to make the code available to other car companies, because they had learned that if a code became standard across the industry it was likely to be of better quality than if Ford had exclusive rights to it.
We developed AMLS for as broad a market as possible. We began working with a German-based firm that provides world-class consulting to the car industry, and started using them as our distribution channel. A number of auto companies in the US, Europe, and Japan quickly adopted AMLS soon after when it became commercially available in 2001. The number of licensees has numbers continued to increase, and now most carmakers worldwide have become AMLS users.
As licensing revenue began to materialize and as we also began to receive hardware and financial and technical support from computer vendors, we found that we had successfully navigated a transition from federal research funding to a healthy level of industry funding. Car companies value AMLS because it enables them to improve marketability of their products: car buyers associate low interior noise levels with higher quality. Harnessing that value has allowed funds to flow back to UT to support ongoing research and software development.
Initially, we developed our software for the Cray supercomputers used for automobile vibration analysis at the time. However, one advantage of AMLS over the previous approach is that AMLS transfers much less data from disk to memory and from memory to processor. This means the huge bandwidth capabilities of supercomputers are not required, allowing AMLS to run on much less expensive workstation-class machines.
AMLS’s ability to perform vibration analysis quickly and efficiently on inexpensive computers resulted in very rapid adoption throughout the car industry. AMLS allowed car companies to switch from supercomputers that cost tens of millions of dollars to machines costing hundreds of thousands of dollars. That trend has continued as these companies have transitioned again to servers that use cheap PC-type processors. We are reaching a point where software is more valuable than the hardware on which it runs, a notion that would have seemed ridiculous in the early days of computing.
We are currently focused on making the transition from using four to eight processor cores to much larger numbers of cores. This includes both the dozens of cores that can be installed on the same motherboard with current and near-future multicore chips, and the thousands of cores that can be installed in one computer using Graphics Processing Units (GPUs) from NVidia, for example.
Over the years, a number of graduate students have contributed to the AMLS effort. Four PhD dissertations and five MS theses related to AMLS have been written in our group. Outside of UT, at least half a dozen software vendors have developed their own implementations of AMLS, and there are several finite element codes for which an interface with our software exists or is currently being developed.
We have enjoyed our success in the car industry, but we would certainly like to see AMLS adopted in the aerospace industry to a much greater extent. (We’re in an aerospace department, after all!) There are some significant differences between the needs of the auto and aerospace industries. But we hope that improvements in AMLS’s capabilities in coming years, along with advances in computer hardware and a growing awareness of what is being done in the car industry, will soon profoundly change the way that structural dynamics is handled in aerospace.
