MIT Uses AI to Discover Ultra-Strong, Tear-Resistant Plastic Material

Researchers at the Massachusetts Institute of Technology (MIT) and Duke University have used artificial intelligence to identify a new class of molecular additives that dramatically enhance the toughness of plastics, increasing their resistance to tearing by up to four times. The breakthrough centers on a type of molecule known as mechanophores—chemical groups that undergo structural or chemical changes when subjected to mechanical force. These changes often produce detectable signals such as fluorescence, color shifts, or changes in electrical conductivity, making mechanophores valuable for applications like stress sensing. In a study published in ACS Central Science, the team discovered that certain iron-containing compounds called ferrocenes—when used as cross-linkers in polymer networks—can significantly improve material strength. When stress is applied, these materials resist cracking and instead exhibit greater elasticity, delaying failure. Heather Kulik, a professor at MIT and senior author of the study, explained: “These molecules can be used to create polymers that are stronger under mechanical stress. Instead of breaking, they absorb energy and rebound.” The research was led by Ilia Kevlishvili, a postdoctoral researcher at MIT, with contributions from Jafer Vakil, a graduate student at Duke University, MIT graduate students David Kastner and Xiao Huang, and Stephen Craig, a professor of chemistry at Duke. The work builds on a 2023 study by Craig and MIT’s Jeremiah Johnson, which revealed a counterintuitive finding: introducing weak cross-linkers into polymer networks could actually increase overall material strength. When stretched, cracks in such materials preferentially break the weaker bonds rather than stronger ones, requiring more energy to propagate and thus enhancing toughness. To explore this phenomenon further, Craig and Kulik teamed up to identify new mechanophores with high potential. “We gained new mechanistic insights and exciting opportunities, but faced a major challenge: how to sift through the vast universe of possible molecules to find the most promising candidates?” Craig said. Traditional methods for testing and characterizing mechanophores are slow and resource-intensive—each experimental test can take weeks, and computational simulations often require days. With thousands of candidates, this approach is impractical. Recognizing the power of machine learning, the team developed a neural network model to accelerate the discovery process. They started with 5,000 known ferrocene structures from the Cambridge Structural Database, ensuring all molecules were synthetically feasible and covering a broad chemical space. The researchers performed force-field simulations on about 400 of these compounds to measure the mechanical force required to break internal bonds—focusing on those with weak, easily ruptured bonds. Using this data, they trained a machine learning model to predict the mechanical response of the remaining 4,500 original compounds and over 7,000 derivative structures. The model identified two key structural features that enhance toughness: first, interactions between substituents on the ferrocene rings, and second, a surprising finding—when both rings are attached to bulky groups, the molecule becomes more likely to break under stress. Kulik noted, “The first feature made sense, but the second was completely unexpected. Without AI, we would never have found this pattern. It’s truly exciting.” From the initial screening, the team narrowed down to about 100 promising candidates. The Craig lab at Duke then synthesized a polymer using one of the top performers—m-TMS-Fc—as a cross-linker in a polyacrylate network. Stretch tests showed that this material exhibited four times the toughness of a standard ferrocene-based polymer. Kevlishvili emphasized the broader implications: “Improving the toughness of existing plastics could extend their lifespan, reducing the need for new production and cutting down on plastic waste.” The team plans to expand their machine learning platform to discover other functional mechanophores—such as those that change color under stress or trigger catalytic activity—opening doors for smart materials in sensors, adaptive catalysts, and biomedical applications like targeted drug delivery. Future work will focus on underexplored transition metal-based mechanophores, a largely untapped area due to higher synthetic complexity. This computational framework could help overcome those barriers and unlock new frontiers in responsive materials science.