Seeking Ways to Make Better Computer Crashes

New models and simulations help improve safety as software and hardware provide more realism to iterate designs more quickly.

September 1, 2015

Terry Costlow

Honda’s new tools (right image) provide improved accuracy and realism, helping engineers glean more information from simulations.

Although dramatic improvements in models and simulations have made them the primary means for determining the crashworthiness of vehicles, developers are pursuing many avenues to make virtual crashes even more realistic. Algorithms are being tuned to run on multiple processors, adding resolution while shortening the time it takes to full-vehicle simulations.

“As the quality and accuracy of simulation models have increased, the number of repeat tests of actual vehicles has been dramatically reduced and the rate of successful first-time actual tests has increased,” said Eric DeHoff, Principal Engineer/CAE Technical Leader for Vehicle Structures Research at Honda.

As these simulations are growing in size they have become more realistic. Though there’s more data in a today’s simulated crashes, the amount of time needed to run them is declining. Computing horsepower has increased while software techniques help streamline processing.

“When we’re running a full vehicle model for a front-end crash, running a model size of five to 10 million elements for a simulation time of up to 120 milliseconds, a 32-CPU system can run it in 10-12 hours,” said Yangwook Choi, Simulia Technical Solutions Director at Dassault Systemes. “That’s improved a lot, even though model sizes have increased, because hardware and algorithms have both improved.”

Simulia’s tools let users run huge crash simulations in a few hours. (Source: National Crash Analysis Center)

Design-tool providers and automakers alike are finding new ways to make simulations more lifelike, accounting for many different parameters. For example, Toyota recently upgraded its total human model for safety (THUMS) virtual human model software to account for precrash actions taken by drivers. These include sudden braking or steering as well as bracing and stiffening by vehicle occupants.

Going forward, modeling and simulation tools will have to address even more parameters. Increasingly, metal components are being replaced by plastics and composites. Software must account for connecting techniques as it’s adapted for the characteristics of new materials.

“Modeling complexity will continue to increase with the challenges of predicting the performance of vehicles built from many different materials like aluminum and fiber-reinforced plastics joined to steel with new connections like adhesives, rivets, and screws,” DeHoff said.

Many facets

Another huge benefit of simulated crashes is that automakers can test many variations instead of simply crashing a car into a wall. Angles and speeds can be changed, and simulations can account for the many different ways drivers find to crash vehicles.

Toyota recently upgraded software to make simulated drivers more realistic.

Models are also being improved to add realism to human injury results. University researchers in Virginia Tech’s medical school are analyzing actual collisions, working with automakers and suppliers to help them determine ways to enhance safety.

“We have 9000 downloads of real crashes using the data recorder in the airbag controller,” said H. Clay Gabler of Virginia Tech’s Center for Injury Biomechanics. “Our role is to find out how systems work in the real world. People have some strange crashes. Someone will depart the road and hit the curb, then hit a telephone pole. It’s difficult to know when to deploy the airbag in that scenario.”

The Insurance Institute for Highway Safety (IIHS), which relies primarily on performing real-world collisions, is also looking at ways to augment existing techniques for determining how to make vehicles safer. IIHS uses injury records to help understand what happens during crashes. These data are shared with automakers and tool developers.

“We are continually going back to the data to see how occupants are being seriously injured or killed in vehicles that otherwise earn good ratings in the crash tests,” said Raul Arbelaez, Vice President of the IIHS Vehicle Research Center. “This research led to our small overlap frontal offset test that we launched in 2012. Right now, we are conducting research to look at passenger-side protection in small overlap crashes. We are also doing research on side-impact crashes to see whether the IIHS test should be changed to reflect other side-impact crash scenarios.”

Dividing vehicles into sections helps Simulia balance computing loads. (Source: National Crash Analysis Center)

Thinking inside the box

Advances in computing technology play a critical role in the snowballing growth in simulation quality. Programmers are taking advantage of the ability to employ several computers to speed up virtual tests.

“A few years ago, simulations ran on six CPUs,” Choi said. “Now they’re running on 128.”

When they run full vehicle tests, automakers are ganging many computers together, taking advantage of multicore processors. All aspects of computing technology come into play when automakers run a broad range of tests, using multiple models and many crash scenarios.

“Current state-of-the-art high-performance computing involves using hundreds to thousands of processors to solve virtual models,” DeHoff said. “Each model uses hundreds of processors to calculate the result. It is typical to run several hundred models each month, which means thousands of processors in total are needed for throughput. The amount of memory needed has increased dramatically due to the details included in each model. Each model generates gigabytes of data, and overall data storage has reached into the hundreds of terabytes.”

IIHS prefers to perform real-world crashes to ensure that simulations mesh with actual collisions.

Full-vehicle simulations are quite large, so it’s challenging to process all the information at a constant rate. Programmers have devised techniques that balance loads, collapsing some sections and splitting models into smaller segments.

“Some companies divide the car into four sections,” Choi said. “We need to develop a load-balancing scheme. During computations, we check the domains to see which have more loads, making adjustments so all the domains have the same level of computing power.”

Simulations allow vehicle engineers to refine their designs to the point that some automakers rarely run physical tests. However, groups like IIHS feel that virtual crashes will never totally replace real-world tests.

“Ultimately, the vehicle needs to be tested because the models aren’t perfect,” Arbelaez said. “Once the vehicle is built, there can be differences between preproduction models used for development crash tests and the finished product coming off the assembly line. A good example of this is the Suzuki Verona. When we tested the 2004 Verona we found that the 1st and 2nd stages for the driver frontal airbag were wired backward, so the 2nd stage deployed first.”