Originating from brain simulations, neural networks is a subset of machine learning. It’s a theoretical model of how the human brain works. Input and output data are fed to the network, and connections of different strengths are formed between the neurons, in a way similar to a physical brain.
Deep learning does not require the programming of human-extracted rules and learns its own representation from the data. This could be game-changing. With continuously growing hardware technology, deep learning could likely change the way of scientific computing.
Physics-based simulations have been widely used to guide product design. But depending on the physics and scale of the problem, simulations could take minutes to hours or even days to run. To verify the performance of many design variants and working conditions could be time consuming.
I train neural networks to overcome this challenge. The neural networks learn from the physics-based simulations, recognizing the 3D shapes and underlying materials. It then learns to predict the physics of the mechanical scenario with varying designs and conditions.
Take the EV design as an example, the design of an EV battery pack has to meet the safety standards upon side impact. With every design change in the battery pack, a physics-based simulation has to be performed to test whether the design meets the safety requirements. When a simulation job is started, the engineer waits for hours to days before he or she can see the simulation results. Now, with a trained machine learning model, the engineer submits a design change and gets to see the results in a few seconds. The results from a machine learning prediction is as rich as a physics-based simulation. We could see how the impact force changes over time and the entire 3D animation of the crash event to understand what would happen to the battery cells upon impact, whether there would be a short circuit due to the battery pack deformation.
As another example, an aircraft landing gear has to bear the weight and acceleration of the entire plane during landing. Depending on the landing speed, angle and weather conditions, the landing gear component could be under stress leading to safety concerns. Trained by physics-based simulations, machine learning models could be used to get landing gear stress state within milliseconds as compared to much longer wait times when running actual simulations.
As you could see, machine learning models trained from physics-based simulations speed up by the thousands to predict 3D results in space and time. This could enable almost an interactive 3D design environment that turns weeks or months of a design cycle into hours, and potentially change the way products are designed and optimized.
In the near future, with almost instantaneous 3D feedbacks from neural networks, car makers could quickly evaluate and optimize their battery pack designs for better safety; doctors could quickly screen and find the matching candidates for a clinical trial; aircraft manufacturers could design landing gears that last for longer; bottle packaging could be made lighter and more sustainable, all at a significantly faster pace.
With machine learning, simulations could become smarter and able to solve new challenging problems; simulations performed by experts could be deployed into the hands of designers and clinicians; simulation representations could be learned and simulation knowledge could be accumulated. Detailed 3D information from simulations could become super valuable for machine learning by enriching the sparse experimental observations critical for the training of balanced and reliable AI.
I can’t wait to see how simulation and machine learning together change the way of scientific computing.
Interested in learning more about this topic? Join us in the SIMULIA Community and visit the Machine Learning wiki.
Interested in the latest in simulation? Looking for advice and best practices? Want to discuss simulation with fellow users and Dassault Systèmes experts? The SIMULIA Community is the place to find the latest resources for SIMULIA software and to collaborate with other users. The key that unlocks the door of innovative thinking and knowledge building, the SIMULIA Community provides you with the tools you need to expand your knowledge, whenever and wherever.
Katie is the Editor of the SIMULIA blog, an expert in engineering and scientific communication, and a strategic digital marketer. Katie has a BA in English and Writing from the University of Rhode Island and a MS in Technical Communication from Northeastern University. She is also a proud SIMULIA advocate, passionate about democratizing simulation for all audiences. Katie is a native Rhode Islander but now lives just over the border in southeast Massachusetts. She is a true New Englander and loves sharing all the cool things NE has to offer. She enjoys a variety of hobbies including history, astronomy, science/technology, science fiction, true crime, fashion and anything associated with nature and the outdoors. She is also mom to a 4-year old budding engineer and two crazy rescue pups.