A tried-and-trusted method comes to the rescue when it comes to improving vehicle safety and other real world design problems. Tom Shelley reports

Single predictions based on the optimum can leave designers in a fool’s paradise, especially when it comes to vehicle crash behaviour; multiple predictions, on the other hand, which are based on a range of scenarios, cover more possibilities and also lead to better designs.

The method, based on techniques used in academic research for decades, are also applicable to any other kind of engineering computer aided design.

It is well known that designing a car to survive a particular kind of crash test can leave it vulnerable to other kinds of accident encountered in real life. But it is less well known that even individual structural members can collapse in a variety of different ways, according to differences in manufacturing and materials within tolerance and slight differences in loading.

Fortunately, all these problems can be addressed by running multiple simulations, under slightly different conditions, and looking at the pattern of results. The basic technique, known as Monte Carlo, has been used by academic computer modellers for decades.

An average of 35 modelled collapses of a square strut

According to Dr Jacek Marczyk, vice president of advanced technology for EASi Engineering, the current obsession with optimisation is like developing a biological species that thrives extremely well under a narrow range of conditions, and then is liable to become extinct if the climate changes. In the real world, engineering designs have to cope with a wide range of circumstances, and this is especially true for motor accidents.

Crashes can happen with colliding vehicles impacting each other from almost any direction. A good design will cope reasonably well with all of them.

Attendees at a recent summit seminar hosted by SGI (formerly Silicon Graphics) in Switzerland were shown how even a simple square strut, compressed axially, could compress in a wide variety of different ways depending on variations in thickness and material properties, all within manufacturing tolerances. One of the attending engineers, from the Fiat subsidiary, ELASIS, interjected that in reality, such struts did not even normally collapse axially, unless saw cuts were made which forced them to do so.

Marczyk agreed, and explained that the model was more constrained than in reality, since the collapsing force was fixed as axial under all circumstances. He defended himself by saying that he had devised the multiple simulation just before the meeting in order to provide a simple illustration of the point he was trying to make.

A basic Monte Carlo simulation starts with a nominal reference model of the system. Values of each variable within the model are then picked at random within their tolerance bands. The nominal values of each variable are then replaced with the random ones and the analysis executed. Output values are recorded and the operation repeated until the statistical distribution of the output values stabilises.

The method allows the engineer to see which parameters matter, and need to be controlled by tight manufacturing tolerances and which do not. It should also reveal the existence of worst case combinations of parameter values and their effects, which the designer may not have been aware of prior to the analysis.

The only problem is that unless the design problem is very simple, the process requires vast amounts of computing effort. A full crash simulation may easily produce Gigabytes of data. If repeated under different conditions, the magnitude increases to Terabytes. Each simulation is carried out independent of all the others, and can be carried out in any order and on a different computer. The process is thus intrinsically parallel, and provided the team have access to enough machines, can be carried out on tens or hundreds of computers or CPUs, hundreds or even thousands of kilometres apart. SGI said that crash analysis already absorbs 55 per cent of computing power used in all aspects of automotive design and one could see them almost rubbing their hands at the idea that users really needed a lot more. Against this, all speakers agreed that each hardware car prototype costs around £250,000 to make, while each design iteration adds about 12 weeks to time to market.

Apart from computing time and costs, the problem of performing a major Monte Carlo-based multiple simulation is to both to manage the process over a number of machines or processors and then perform the necessary data analysis and present results in a manner the engineer can make sense of.

A package called Storm (Stochastic Optimisation and Robustness Management), developed by the Space Division of Casa, Spain’s leading aerospace manufacturer, manages the Monte Carlo process and collates the results. Visualisation of the statistics of the analysis can be undertaken while the analysis is still under way. The package runs with any analysis package and only requires that the solver runs with an ASCII file and produces an ASCII output file.

For those concerned with validating computer models, the multiple simulations produce a point cloud of output results, which should correlate with distributions of experimental ones. A single computer modelling, like a single experimental result, is of completely unknown precision in terms of real world results. Since computers work in terms of 32-bit or 64-bit values, there is a temptation to believe that any results they output are similarly accurate. Performing multiple analysis under varying conditions produces mean and variation values which should give a much truer measure of how accurately the predicted result is likely to relate to the real world performance of the manufactured product.

Such a process has been applied to crash analysis in collaboration with BMW and ESI. A PAM-CRASH model of a BMW car was used with 60,000 elements. The variable parameters were thicknesses and failure mechanisms of certain structural members in the engine compartments, together with crash overlap, impact angle and velocity. Intrusions of the foot well, firewall and A-pillar were chosen as output variables together with accelerations and internal energies of selected groups of materials. The simulation was performed 128 times on four CPUs each in a 512 CPU T3E Cray at HWW in Stuttgart. The analysis took three days. In 1998 SGI built a special computer for BMW for running ST-ORM with 160 processors and dedicated to running these types of crash analysis. Mercedes now has a similar machine from Compaq.

Marczyk said that the failure of most CAE analyses to address the problem of variation in input parameters means that CAE users and software are still in "The Pre-Columbian Era". With the growth of computing power and its ever-lower price, he foresees a revolution in the use of CAE as variability and uncertainty are routinely addressed. The result, he believes, will be more robust designs, cheaper to manufacture, which can be used in a wider range of different circumstances. This contrasts with highly optimised and very precisely made designs, which may not be suitable for use outside a narrow range. When asked if there might be other means of addressing the problem of uncertainty using less computing power, he replied that this did not matter. Whether the processing methods used Monte Carlo, fuzzy logic or chaos theory was of no importance, it was the concept of addressing uncertainty, which was important.

Other users of Storm include ABB, Audi, Toyota, Nissan, British Steel, Takata, NEC and Italdesign. Joint projects are being run with Ford, GM, Isuzu, Jaguar, NASA and Visteon. ABB uses it for performing thermoelastic analysis of heat shields in power turbines. It has also been used to design a composite antenna reflector for use in space, and to perform launcher/payload analysis for the Ariane 5. Porsche have used it to study the front axle suspensions of sports cars and CASA have used it to investigate the effects of small shape imperfections in hydrofoils.