Machine Learning identifies crosslinkers that increase polymer toughness
Researchers at MIT and Duke University have outlined a new way to make polymer materials more resistant to tearing, using machine learning to guide molecular design. The approach could help extend the service life of plastics and synthetic fibres, with clear implications for reducing material waste.
The focus of the study is on mechanophores molecules that respond to mechanical force by changing their structure. Instead of allowing a material to crack under stress, certain mechanophores can absorb and redirect force in a way that improves durability.
“When you apply stress to these materials, they don’t fail immediately,” explains Heather Kulik, Professor of Chemical Engineering at MIT. “They show higher resilience under load.”
Why ferrocenes matter
The team concentrated on a specific group of iron-containing molecules known as ferrocenes. These compounds have existed in chemical libraries for decades, but their potential role as mechanophores has not been widely studied.
Testing even one mechanophore in the lab can take weeks. To avoid this bottleneck, the researchers turned to data and computation. They started with the Cambridge Structural Database, which includes around 5,000 ferrocenes that have already been synthesised.
“This meant we could focus on performance rather than synthesis feasibility,” says MIT postdoctoral researcher Ilia Kevlishvili. “We had access to a chemically diverse and realistic design space from the start.”
From thousands of molecules to clear patterns
The team first ran detailed simulations on about 400 ferrocene compounds. These calculations estimated how much force would be needed to pull key atoms apart. For tear-resistant polymers, weaker molecular links can actually be beneficial, as they dissipate stress before damage spreads through the material.
Data from these simulations, combined with structural information, were used to train a machine-learning model. Once trained, the model predicted the force response of the remaining 4,500 ferrocenes in the database, along with another 7,000 related structures generated by rearranging atoms.
Two structural features stood out. One was the interaction between chemical groups attached to the ferrocene rings. The other was less intuitive: large, bulky groups attached to both rings made the molecule more likely to break in a controlled way under stress. According to the researchers, this second effect would have been difficult to identify without machine learning.
Putting the idea to the test
From the predictions, around 100 promising candidates were shortlisted. One of them, called m-TMS-Fc, was incorporated into a polyacrylate plastic by the lab of Stephen Craig at Duke University. In this material, the ferrocene compound acts as a crosslinker, tying polymer chains together.
Mechanical testing showed a clear result. Polymers using m-TMS-Fc were about four times tougher than similar materials made with standard ferrocene crosslinkers.
“If materials last longer, they need to be replaced less often,” says Kevlishvili. “That directly affects how much plastic waste accumulates over time.”
Beyond strength alone
The researchers see this as a starting point rather than an endpoint. The same machine-learning framework could be used to search for mechanophores that respond to force in other useful ways by changing colour, triggering chemical reactions, or activating catalysts. Such behaviour could support applications ranging from stress-sensing materials to controlled drug delivery.
“Transition metal mechanophores are still relatively underexplored,” notes Kulik. “Computational screening allows us to study a much wider range of candidates than traditional lab methods.”
The work was supported by the National Science Foundation’s Center for the Chemistry of Molecularly Optimized Networks (MONET). While practical deployment will take time, the study shows how data-driven chemistry can guide smarter material design, one molecule at a time.
