Machines Rise To The Grand Challenges of the 21st Century

What role can machine learning play in advancing modeling and simulation capabilities for physical systems as seemingly disparate as aircraft and infectious disease spread? Researchers from The University of Texas at Austin are leading two new multi-institution projects that will address this question.

RISE of the Machines is one of six projects selected as part of the Department of Energy’s (DOE) Advanced Scientific Computing Research program on Artificial Intelligence (AI) and Decision Support for Complex Systems.

Led by researchers from the Center for Scientific Machine Learning at the Oden Institute for Computational Engineering and Sciences, this project will develop intelligent automation and decision support methodologies that enable a predictive digital twin.

A digital twin – which is an evolving virtual model of a specific system or physical asset, assimilating asset lifecycle data so that it becomes an asset-specific model – is recognized as a key technology that could revolutionize fundamental decision-making processes for complex systems. The project will target the development of digital twins in two high-impact application areas: structural condition monitoring of engineering assets and infectious disease spread.

“AI methods offer great promise for transforming decision support for complex engineered and natural systems through the concept of a digital twin,” said project lead, Dr. Karen Willcox, Director of the Oden Institute and professor of aerospace engineering at UT’s Cockrell School of Engineering.

“However, the central challenges in applying AI methods to scientific domains remain: how to ensure robustness, interpretability, scalability, and efficiency? These challenges are particularly heightened when it comes to the societally-critical systems under the purview of the Department of Energy.”

The robustness, interpretability, scalability, and efficiency (RISE) challenges will be tackled by the team through new methods grounded in large-scale optimization and optimal control theory. In particular, the team – which includes researchers from Sandia National Laboratories and Emory University – will be developing surrogate models that can be quickly adapted with dynamic data using interpretable machine learning; Bayesian inference methods that tackle the challenge of inferring unknown parameters from sparse sensor data; active learning and optimal experimental design methods to determine the most informative data; and robust model-based reinforcement learning for decision-making.

Willcox will also lead a new Multidisciplinary University Research Initiative (MURI) project funded by the Air Force Office of Scientific Research to advance machine learning for physics-based systems. One way to develop AI with greater safeguards is through the emerging field of Scientific Machine Learning – where autonomous decision-making is tethered by predictive physics-based models – a key research priority at the Oden Institute.

The MURI project, which also includes researchers from Carnegie Mellon University, Northwestern University, and the University of Wisconsin-Madison, will target the challenges of training for machine learning models. “Training is a huge computational bottleneck for applying machine learning to systems governed by partial differential equations,” said Omar Ghattas, co-PI on the MURI project, Director of the Center for Computational Geosciences and Optimization in the Oden Institute, and a professor of mechanical engineering and geological sciences. “It is just not feasible to generate the massive data sets that are needed to adequately train state-of-the-art machine learning models.”

The team will tackle this challenge by developing faster optimization methods and optimized training data selection strategies that will enable new scalable machine learning approaches that require much less training data. The MURI project will also be investigating optimal neural network architectures that are tailored to the structure of the target physical problems, theoretical analysis of the approximation and generalization properties of the resulting machine learning surrogates, and rigorous uncertainty quantification to support certified predictions.

“It is timely and exciting to be kicking off these two new projects,” Willcox said. “There is a great deal of enthusiasm for applying machine learning to complex problems in engineering, but a growing recognition of the need to draw on the long history of mathematical modeling and scalable algorithms from the field of computational science in order to successfully realize the potential.”