Civil Engineering Prof. Steve WaiChing Sun Wins NSF CAREER Award

As part of his NSF project, Sun has invented a tool that incorporates augmented intelligence and deep learning to help scientists make better predictions of materials failure; will improve complex analyses and designs for infrastructure

Feb 14 2019 | By Holly Evarts | Diagram Credit: Steve WaiChing Sun/Columbia Engineering | Photo Credit: Jane Nisselson

Are there similarities between playing chess, Go, video games, and conducting research? If these similarities exist, how can we exploit them to make us re-examine our thought processes so that we can be better scientists? These are the questions that inspired Steve WaiChing Sun, assistant professor of civil engineering and engineering mechanics, to explore a new idea that has won him a National Science Foundation CAREER award for his work in predicting material failures for engineering applications. The five-year, $500,000 award, which is one of the NSF’s most prestigious honors for faculty at or near the beginning of their careers, will fund his project, “Deep-reinforcement-learning-enhanced computational failure mechanics across multiple scales.”

Sun’s idea is to create a new augmented intelligence approach to improve the prediction of localized failures of material, such as shear banding and fractures. “I see augmented intelligence as a slightly different design concept from artificial intelligence,” he says. “Augmented intelligence assists in advancing human capabilities, while artificial intelligence oftentimes replaces them.” 

While there have been many attempts to use various approaches, such as support vector machine and neural networks, to predict material responses, these earlier efforts resulted in predictions for which the rationale is impossible for humans to understand. The inability to interpret these “black-box approaches” renders them inapplicable to engineering mechanics problems—such as the design of aircraft—where understanding the rationales behind these predictions are critical for ensuring safety.

Sun takes a completely different research path. He is interested in making augmented intelligence emulate the thought process that a scientist goes through while deducing material laws. He calls this new approach a meta-modeling game—a reference to the fact that his approach is about modeling the thought process of a scientist establishing a material law, rather than modeling the material responses for a particular substance.

His new approach essentially idealizes the scientific process as a class of games like chess and Go, where a given set of universal principles is provided, with a sequence of actions completely compatible with the stated rules (i.e., the law of physics). The AI then repeatedly proposes different causality relationships among data (e.g. displacement, crack density, damage, force, stress) and test the plausibility of these hypothesized relationships. While both the augmented intelligence and humans can both enter the game, the best idea, regardless of whether it is generated by human or AI, will always win at the end of the gameplay.

I’m very honored to have won the NSF CAREER award...It’s a great recognition of the reach of my research and gives me the ability to continue exploring how to harness the power of augmented intelligence to complex tools that will revolutionize how we study mechanics, materials, structures, and infrastructures.

Steve WaiChing Sun

“Automating the time-consuming trial-and-error process will accelerate the progress of science, and help us make more efficient, robust, and precise analyses and designs for hand-held devices, structural components, civil infrastructure, and geological disposal sites of nuclear waste and carbon dioxide,” says Sun, who is also a member of the Data Science Institute.

While Sun’s research interests are focused on the mechanics and physics of geological and porous materials such as soil, rocks, concrete, and salt, his research has pioneered many new applications of mathematical concepts to the field of mechanics of materials, such as level set, phase field, variational principle, group theory, and, in his NSF work, graph theory and combinatorial optimization. Conventional mechanics models are often products of “continuum mathematics” (e.g. calculus), but the exponential growth of applications of discrete mathematics, partially due to the growth of demands in analyzing big data, has created new perspectives for many century-old problems. Sun’s research can be applied to a wide range of these problems—from field-scale techniques (geological storage of carbon dioxide, hydraulic fracture, geological disposal of nuclear waste, and vehicle-soil-water interaction) to micro-scale simulations (3D printing processes and fragmentation and fracture of a single crystal). His research promises to bring new order to the chaos of big data and provide elegant ways to explain classes of physical phenomena in the simplest and most efficient way possible.

Sun has already launched his NSF project by creating a meta-modeling tool that can autonomously generate models that capture the effects of changes in microstructures from micro-cracks, plastic slip and wear at sub-scales, and then can recursively upscale responses to the scale of interest. The tool uses deep reinforcement learning to generate material laws for interfaces and can efficiently self-improve the models it generates through continuously self-deriving and self-validating material laws and, most important, learning from past failures, similar to the way AlphaGo Zero improves its gameplay.

In a paper to be published in Computer Methods in Applied Mechanics and Engineering in April 2019, Sun’s meta-modeling game essentially frees modelers from repetitive trial-and-error processes. Instead, future improvements on the models can be made by expanding the action space or by simply leveraging computational power to improve the models over time.

“This freedom allows us the unprecedented luxury of placing our focuses on things that really matter to a modeler—finding the best cause of actions that lead to the most predictive model,” he notes.

In addition to this NSF CAREER award, Sun has won the Air Force Office of Scientific Research Young Investigator Award in 2017 and the Army Research Office Young Investigator Award in 2015. His contributions in computational poromechanics were recently recognized by the EMI Leonardo Da Vinci Award from the American Society of Civil Engineers.

“I’m very honored to have won the NSF CAREER award in the Mechanics of Materials and Structures Program at NSF,” Sun adds. “It’s a great recognition of the reach of my research and gives me another kind of freedom – the ability to continue exploring how to harness the power of augmented intelligence to complex tools that will revolutionize how we study mechanics, materials, structures, and infrastructures.”