How AI is Rewriting the Rules of Materials Discovery

Discoveries of new materials underlie many of the forward leaps in human society throughout the ages. Roads made from Roman concrete that still stand today allowed merchants and armies to travel quickly across long distances. Adding various elements to the production of steel unlocked strength and versatility that continues to form the basis of much of today’s infrastructure. Semiconductors, titanium, carbon fiber, complex composites and biogels ushered in many of the modern conveniences we now take for granted.

All these materials have something in common—their discovery took a painstakingly long time of trial and error or were created completely by accident. The advent of powerful computer processors over the past decades, however, has enabled scientists to begin to predict what properties new material recipes will have.

Now, with the rapid ascent of AI, the trajectory of the field is looking to fly to even greater heights. With a long history of strength in both computational materials and AI, Duke University is the perfect nexus for these fields to fully integrate.

Students enter the program with expertise in either AI or one of many fields encompassing materials science. Through the program’s curriculum, they receive training in scientific concepts in both their home field and the new domain. Starting in the second year, students are grouped into small, diverse teams to pursue a research project harnessing the knowledge and talents of all members and their collective training in AI and materials science.

“I came into the program without any knowledge about AI, and the program provided a really nice roadmap for how to learn about AI and incorporate it into my research,” said Daniel Duke, a PhD student in the laboratory of Gaurav Arya, professor of mechanical engineering and materials science. “The collaboration with my colleague also enabled me to learn about AI much more effectively and pursue my work from a new perspective.”

His cross-disciplinary research project is a classic inverse problem. Predicting how a certain combination of components will behave together in a new material is difficult but doable. Going the other way around—asking a computer to give you the best recipe to create a material with a specific set of properties under a variety of constraints—is a horse of a different color.

In Duke’s case, he worked with Han Zhang in Brinson’s lab to do just that with acoustic metamaterials—structures that use both physical geometry and intrinsic material properties to shape and manipulate sound waves for applications such as acoustic cloaking. The duo is using diffusion models to tackle the inverse design problem for these materials.

The process starts with a dataset in which the solutions are already known. For each datapoint, the researchers add a little bit of noise to make it blurry and train the AI to recognize what adding noise to data looks like. This process repeats over and over again until the AI learns what it looks like to go from clean data to complete noise. Then, the trained AI performs the task in reverse, taking an unknown solution and making it clearer and clearer with each iteration until it emerges with an accurate geometry to achieve the desired property.

“My PhD research is on DNA origami, so the acoustic metamaterials project is more aligned with my colleague’s research focus,” Duke said. “But my advisor is interested in learning how to use diffusion models in our research, so we’ll definitely benefit from the skills that I get out of this project.”

In a similar vein, Jacob Peloquin, a former PhD student in the laboratory of Ken Gall, professor of mechanical engineering and materials science, helped set the stage to build an AI model that can predict the mechanical behavior of complicated hierarchical porous materials.

Porous materials are all around us—even inside of us—every single day. Think of how soil is a random mix of different shapes and sizes of dirt, rock and organic matter, or how bones are complex microstructural networks of interconnected pores and channels. It was long hypothesized that four properties could adequately predict the properties of a porous material: the surface area of the internal structures; the connectivity of the various internal structures; the amount of negative space within the porous structure; and the average size of the constituent particles or grains.

Working with Winston Lindqwister, then a PhD student in the lab of Manolis Veveakis, professor of civil and environmental engineering, Peloquin created a framework for predicting the mechanical response of porous materials based on these features. They used 3D micro-scale scans of the interiors of various porous materials such as rock, bone, concrete, and wood, fabricated 3D-printed replicas, evaluated their mechanical performance both physically and using computational simulations, and ultimately trained an AI model to predict their compressive behavior.

“As I progressed through the program, I recognized how essential AI and data science are becoming in the field of materials design and discovery,” said Peloquin. “The program’s goal of integrating data science and AI with domain expertise in materials science aligned perfectly with my background in mechanical engineering and materials science, providing me with the necessary knowledge and skills to innovate at the intersection of AI and materials science.”

Others at Duke University are now taking the foundations laid by Peloquin and Lindqwister to work on their own inverse design problem. Laura Dalton, assistant professor of civil and environmental engineering, is leading a team in collaboration with Peloquin to predict microstructural geometry needed to achieve specific mechanical performance targets of various porous natural materials. Instead of using structure to predict response, they’re using desired response to inform structural design.