Computational scientists have long used math- and physics-based modeling and simulations to analyze big datasets and make predictions in a range of scientific areas. These models can work on large scales, such as to simulate the weather and climate, or to predict earthquake risk. They’re useful at small scales, too, to help identify potential new drugs that can do the most good with the fewest side effects, or predict how an infectious disease might affect an organism at the cellular level.

Machine learning has proven to be a powerful tool with a variety of applications. But complex areas like autonomous vehicle operation, rocket combustion, and monitoring the structural health of urban infrastructure increasingly demand predictions that go beyond the existing data, says SFI External Professor Karen Willcox, a computational engineer at the University of Texas at Austin. These types of applications, says Willcox, need new methods that can make accurate and efficient predictions using sparse — or even changing — datasets. 

“Computing power has increased, and we know that computing can play a major role in making critical societal decisions,” says Willcox. “So how do we think about moving beyond existing approaches to solve problems at scale in complex systems?”

Scientific machine learning offers a way forward. It’s an emerging field at the crossroads of computer science and computational science, and it focuses on harnessing new ideas in machine learning together with predictive physics-based models to solve complex, real-world problems.

October 10–12, a group of mathematicians, statisticians, computational scientists, computer scientists, and experts across a wide range of scientific domains converged at SFI to collaborate on new ideas about using scientific machine learning in complex fields. The workshop was organized by Willcox, together with colleagues at UT-Austin, Sandia National Laboratories, and the University of Michigan. Invitees included program managers from the Air Force Office of Scientific Research, ARPA-E, the Department of Energy, and other governmental organizations that explore and establish national priorities for scientific machine learning. 

The fields of computational science and computer science have always been complementary, says Willcox, though finding a common language can be difficult. “There are multiple examples of techniques emerging in the machine learning community that have close connections to the approaches that have been used in a different way for many years in computational science, but they go by different names,” she says. “We need to break down these barriers and exploit the synergies of the complementary computer science and computational science perspectives.”

Willcox and her colleagues will collect the insights from the workshop for a future issue of the journal IEEE Computing in Science & Engineering. She also hopes it will be the start of an ongoing conversation in a wider, interdisciplinary community that’s focused on the future of computational models.