AI platform ‘evolves’ metamaterials

With just a couple of "pieces of matter"—representations of one basic unit of a material—the new platform can create thousands of previously unknown morphologies, or structures, with the properties Amir Alavi specified.(Credit: Amir Alavi/U. Pittsburgh)

Evolution has led to some creative, counterintuitive, and downright weird properties in the natural world. Evolution-inspired algorithms produced synthetic materials with equally surprising traits, a study shows.

In a paper published in the journal Advanced Intelligent Systems, Amir Alavi, assistant professor of civil and environmental engineering in the University of Pittsburgh’s Swanson School of Engineering, outlines a platform for the evolution of metamaterials, synthetic materials purposefully engineered to have specific properties.

The platform used generative artificial intelligence (AI), similar to the technology underpinning ChatGPT, to create these new materials using a process resembling evolution by natural selection.

As the process repeated, the algorithm continued to create new materials. They ranged from basic structures like those commonly found in materials science labs to intricate shapes reminiscent of ancient scripts found etched on a clay tablet. “Some of these structures are just so complex and inconceivable by the human mind,” Alavi says. “Yet, they provide excellent mechanical performance, better than all the other solutions we’ve come up with before.”

Beyond generating materials with the properties he wanted, the platform also evolved materials with properties that were unique, and potentially useful, in unexpected ways. After running the program for less than a week, the team could have found nearly 100,000 structures with new modalities. “We’ve accelerated the process of evolution. We can find new materials in a couple of days—materials that could have taken 10 million years to form and evolve in nature,” he says.

Alavi’s long-term vision is to harness the power of generative AI tools for reshaping America’s civil infrastructure. And in his opinion, metamaterials are a perfect fit for large-scale infrastructure projects because small improvements in weight or strength add up quickly over industrial-sized structures.

In two first-of-their-kind projects, his team is building megastructures from metamaterials.

With a grant from the Pennsylvania Turnpike Commission, he is developing sound-absorbing metamaterial to create a new kind of sound barrier—an open one. “You will be able to see through it,” Alavi says, “but we hypothesize noise will be reduced by 90%.”

With a grant from the Impactful Resilient Infrastructure Science and Engineering consortium, Alavi has produced the first prototype of an ultralight, ultra-strong concrete with potential uses in pavement and bridges for the Pennsylvania Department of Transportation. He predicts it will cost around 20% less than traditional concrete.

But the platform isn’t limited to creating stronger concrete or better building materials. It could model a molecule or a neuron or, who knows, maybe the platform could be used to model the evolution of an entire living organism.

That’s part of its beauty, Alavi says.

“We can fit anything into this system and see how it morphologically evolves. We may find solutions never seen before. All we need is a single piece of a system to launch the evolutionary process.”

Unlike ChatGPT and other generative algorithms that are trained on massive amounts of data to make something new, Alavi’s platform doesn’t need thousands of examples. It doesn’t need any examples. With just a couple of “pieces of matter”—representations of one basic unit of a material—it can create thousands of previously unknown morphologies, or structures, with the properties he specified.

Each piece of matter acts as a parent with its own physical properties, like a particular shape or a certain hardness, which are represented in Alavi’s algorithm as genes. Then, to create a new offspring, “the two merge and exchange genes,” Alavi says.

In the natural world, the offspring’s success depends on having traits that help it survive and reproduce. Maybe it was born with a mutation that gave it extra hair that would come in handy during unseasonably cold weather, giving it an edge over its bald brethren.

And like evolution in the wild, Alavi’s algorithm requires randomness to mix things up. In the same way mutated genes can lead to new traits, the evolving metamaterial algorithm randomly changes a property of a parent cell by, for instance, straightening a curve or changing the tensile strength, before incorporating it into the offspring.

But in the lab, it isn’t the environment that determines whether offspring persists, but parameters Alavi sets in advance. Say he wants a material that can resist being crushed by a heavy load: “If the child of the two parents meets the criteria that we have defined for the maximum strength, it will be kept in the population,” he says.

If it doesn’t meet that threshold? “It’s thrown out. Like survival of the fittest. It’s a brutal process.”

Source: Brandie Jefferson for University of Pittsburgh