Deep learning is the machine learning technique in artificial intelligence behind the technology in driverless cars and voice control in cell phones and other devices. It teaches computers to do what humans do naturally and automatically.
The challenge in developing deep learning models is that computers need an extensive amount of training data in order to reference each specific situation and “make decisions.” In many cases, the simulations that could help supply this data take a great deal of time, computational power and are cost-prohibitive.
A research team, led by Jian-Xun Wang, assistant professor of aerospace and mechanical engineering at the University of Notre Dame, has developed a deep-learning method that creates reliable surrogate models of fluid flows without using simulation data for training. Their new approach to the predictive modeling of fluid-structure interaction problems aim to enable real-time predictions, which are vital to applications from health care to national security.
“Cardiovascular diagnoses, aircraft design, and active flow control — such as understanding the effect of waves on risers around an off-shore oil platform — all rely on detailed information of complex fluids and their interactions with surrounding structures,” says Wang. He and his team leverage recent advances in machine learning, but they also use existing knowledge of physical principles to achieve reliable surrogate models.
While most deep learning models require volumes of data, Wang and his team take limited data and apply physics equations to extrapolate the rest. “We are not making our ‘best guess,’” he says. “All data, whether we have it in hand or not, has to satisfy basic physical laws. Knowing this allows us to create our models rapidly and accurately.”
Team members include Wang and graduate students Luning Sun and Han Gao. Their work, which is being conducted in conjunction with Notre Dame’s Center for Informatics and Computational Science, is supported by the National Science Foundation.