$ 3.25 million DOE grant funds UChicago / Argonne research on AI models of physical simulations
The most advanced computer models for studying Earth’s climate, fluid dynamics, or the operation of the national power grid operate on Partial Differential Equations (PDEs), mathematical equations used to describe physical laws. But these equations are difficult to solve, and it can take hundreds of hours on the world’s most powerful supercomputers to run a single simulation.
Surrogate models offer an AI-powered alternative. Instead of running the model, scientists feed the initial conditions and results of past runs of the model as data in a machine learning mode, such as a neural network, that learns to reproduce the predictions of the model. full supercomputer simulation. Once trained, these AI models produce results much faster than the original, allowing scientists to better characterize the range of possible results through repeated testing.
“The use of surrogate models in the context of data-centric problems in scientific applications is important, as most of the tasks involve repeated evaluations of the models,” said Sanz-Alonso. “Anytime you need to evaluate the model, the computation is expensive, and in some applications you have to evaluate the model millions of times, so it’s especially important that you can do it at a lower cost. “
More runs also allows for better evaluations of low probability, high impact events that may appear in only a tiny percentage of simulations. The original model, run only a few times due to computational limitations, may miss these rare occurrences. But a surrogate model, run hundreds or thousands of times, could spot them and accurately characterize their frequency.
With this information, the authorities could better prepare for the “black swan” events which result in considerable costs and losses, Anitescu said. Infrastructure designed to survive “once a century” events will have a more reliable estimate of what this means, while the frequency of extreme weather events could be better understood and anticipated.
“It’s more useful for decision making, especially when you care about properly assessing the risk of extreme events,” Anitescu said. “But we need to be able to model them faster, because things change. If we can’t tell people what to expect from real rare events, like floods and hurricanes, and how much it really costs to build on a coast, someone is going to go broke – the people who build houses or government. “
The UChicago / Argonne team is well placed to take on the multidisciplinary scope of the project, which spans from mathematical foundations to cutting-edge data and computational concepts of artificial intelligence, including the modeling of applications in many fields. different scientists. The research was also originally initiated by an “AI + Science” seed grant from the UChicago Joint Task Force Initiative (JTFI), which supports collaborations between UChicago, Argonne and Fermilab.
“Collaborations between UChicago and national laboratories allow us to leverage broad skill sets across disciplines to address these challenges,” said Willett. “The seed funding provided by JTFI in the field of AI for Science has been essential in encouraging us to form partnerships and define great challenges that can only be addressed through great collaborative efforts. “
Source: Department of Computer Science, University of Chicago