New AI Model Discovers Optimally Stiff and Tough Composites
Hitchless car suspension, indestructible planes, and bone-like prosthetics. All three require materials that are both stiff and tough — a new AI program developed by researchers at MIT might be able to deliver just that.
Courtesy of Beichen Li et al., Computational discovery of microstructured composites with optimal stiffness-toughness trade-offs.
Courtesy of Beichen Li et al., Computational discovery of microstructured composites with optimal stiffness-toughness trade-offs.
Courtesy of Beichen Li et al., Computational discovery of microstructured composites with optimal stiffness-toughness trade-offs.
Bend a rubber rod, and you’ll observe what solid mechanicians call deformation, or the change in a solid’s geometry under force. Try this with a bar of concrete, however, and you’ll get . . . nothing, at least at first. But ramp up the force on the bar, and it will begin to stretch, by a thousandth of a millimeter at first, until rapidly, and with an audible crack, it fractures.
The contrasting behavior of the two bars illustrates the stiffness-toughness paradox. It is a principle in material science that to be tough or resistant to fracture, a material must be able to absorb a force (in other words, easily deformable). Exceptions to this rule exist and are highly desirable in fields like aerospace and professional sports, where parts need to be stiff, lightweight, and damage resistant. However, they are few, and finding them is more art than science.
Beichen Li, a Ph.D. student in MIT’s Computer Science and Artificial Intelligence Laboratory, wanted to systematize that process. In an interview with The Tech, Li spoke about his research developing an AI program capable of discovering microstructured composites—millimeter-wide 3D-printed matrixes of various materials—that display optimal stiffness-toughness properties.
The first thing to note about this particular problem, Li says, is its enormity. Because of the stiffness-toughness trade-off, there is no “optimal” matrix arrangement for a given set of constitutive base materials. Instead, imagine stiffness and toughness as two axes of a graph. Every possible microstructured composite for a given set of base materials can be plotted as a point lying somewhere in the plane defined by those two axes. Layer on enough points, and eventually, a curved boundary connecting the two axes, known as the “Pareto-optimal front,” emerges; composites lying on this boundary are “Pareto-optimal,” meaning they can grow no stiffer without sacrificing some toughness, or vice versa. If you’re familiar with the production possibilities curve from macroeconomics, it's the same concept. Li set out to discover not one optimal structure but a whole series of them — the Pareto-optimal front.
The enormity of the Pareto-optimal problem forecloses traditional solutions. For instance, it is almost impossible, Li explains, to attempt to discover all Pareto-optimal structures through the labor-intensive but highly accurate process of physical experimentation. Meanwhile, computational costs and hard-to-model physics make computer simulations equally impractical. Machine Learning, then, seems like the best hope. The only problem? “[Existing] strategies have a common weakness—low sample efficiency,” Li says. In other words, even the most advanced conventional ML algorithms, which function by making small guess-and-check changes to a base composite design, may have to go through an impossible number of iterations before discovering any significant number of Pareto-optimal structures. At first glance, the problem seems intractable.
“[Any] effective computational pipeline for our task must bridge the gap between simulation and reality while minimizing the number of simulations required,” Li says. The solution, he realized, was to combine the three methods — physical testing, simulation, and machine learning — into a single pipeline that he’s since dubbed “the three-part nested-loop workflow.” The pipeline begins with a machine learning (ML) algorithm capable of quickly generating and assessing virtual samples but with limited initial accuracy. The algorithm then feeds the promising samples to the slower simulator, which tests the samples, provides feedback to train the algorithm, and down-selects for the most optimal structures. This last additional step occurs because simulators are particularly bad at modeling fracturing on the Pareto-optimal front. Those are then sent for physical testing, where researchers fabricate the most promising designs, test them, and use the results to train the algorithm and simulator. By allocating the least amount of labor to the slowest component of the pipeline, Li frees the fastest component, the algorithm, to iterate rapidly without losing accuracy. The result is an algorithm capable of creating optimally stiff and tough micro-structured composites from first principles; no flashes of human insight or expert input are needed.
Li called the moment when he started seeing the different components of the computational pipeline as an integral whole an epiphany. And indeed, it is, not just for Li, but the field. While previous works often combined two of the three components — physical modeling and ML, or simulation and ML — the holistic approach of the three-part nested-loop workflow is a new technique that promises to help in other hard-to-simulate fields like fluid dynamics and polymer chemistry. “At the end of the paper, we mostly want to emphasize potential applications of our method in other areas,” Li says.
This broad view of his field characterizes much of Li’s hope for the future and his time at MIT. Since completing the microstructured composites paper, Li’s interest has strayed towards the automated generation of 3D assets for video games, movies, and VR. "This is a significant departure from 3D-printed composites,” Li admits, “but I think they still share the same foundational thoughts — using the power of computational approaches to automate traditional design tasks.”
“[Completing a] PhD is ultimately a long journey shrouded in mystery,” Li says. But if his diverse research interests are to be taken as evidence of anything, perhaps they’d be best taken as evidence of his confidence in the promise of computational methods: "the gist is to develop systematic and efficient approaches to solve design challenges better than a human ever could." That might sound like a big ask, but so was solving the stiffness-toughness problem in microstructured composites. And, as Li points out, the ripple effects of that particular achievement aren't fully apparent yet. After all, it's the computational method, not the materials, that will force a paradigm shift in how people design.