Machine learning characterizes plastics by their flow

The global chemical industry produces hundreds of long, repeating molecular chains called polymers. Only a handful comprise the majority of the planet’s plastic products, but all are difficult to tell apart and recycle efficiently.

Elliott et al. designed and trained a novel neural network to determine the molecular weight distribution of a polymer mix — and therefore the lengths of its constituent molecules — using the polymers’ rheology, or flow. Across nine iterations, the model returned accurate molecular weight distributions for both artificially generated and experimental data.

“Everybody realizes that there are problems with plastics,” said author Daniel J. Read. This work aimed “to try to understand how you might be able to rapidly characterize the materials that are being recycled, so that you can try to make quick and sensible decisions as to what to do with them.”

Rheology is a commonly measured feature of polymers, done so by wobbling a material between two plates to obtain a spectrum of its molecular movements. The team used a computer model to generate such spectra and molecular weights for 800,000 polystyrene polymers, then fed the data to nine neural networks. When the team queried their models about a specific rheology curve, they spat out their best guess of the material’s molecular weight distribution in a fraction of a second.

The team envisions a world where recycling plants equipped with rheometers and their free, consolidated model can determine a plastic’s composition in real time, then match it with similar items for proper recycling. They hope their work “solves one little problem in a much bigger space of how to do recycling wisely,” said Read.

Source: “Using neural networks to deduce polymer molecular weight distributions from linear rheology,” by Robert J. Elliott, Luisa Cutillo, Chinmay Das, Johan Mattsson, and Daniel J. Read, Journal of Rheology (2025). The article can be accessed at https://doi.org/10.1122/8.0001063 .