Cornell University's chemical researchers have developed a machine learning model that enables manufacturers to customize and improve HDPE materials, thereby reducing the amount of material needed for various applications. It can also be used to enhance the quality of recycled HDPE, making it comparable to new HDPE, thus making recycling a more practical process.

Molecular weight distribution allows the use of less material

"Implementing this method will aid in the design of next-generation commodity materials and achieve more efficient polymer recycling, thereby reducing the overall environmental impact of HDPE," said Robert DiStasio Jr., associate professor of chemistry and chemical biology in the College of Arts and Sciences (A&S).

"Published on March 14 in the Journal of the American Chemical Society, 'Designing Polymers Using Machine Learning Based on Molecular Weight Distribution' is a collaborative effort by DiStasio with polymer experts Geoffrey Coates (Professor in the Department of Chemistry and Chemical Biology (A&S) at Tisch University) and Brett Fors (Frank and Robert Laughlin Professor of Physical Chemistry (A&S))."

PhD student Jenny Hu is the first author. Postdoctoral researcher Zachary Sparrow from the DiStasio group; Brian Ernst, a former postdoctoral researcher; and PhD student Spencer Mattes contributed.

HDPE requires so much energy because it is mass-produced, Fors says, whose lab focuses on sustainable polymers. There are also challenges in recycling it.

"Recycling polyethylene is more expensive than producing virgin plastic," he said. "Another issue is that when you mechanically recycle it, you start to break down the polymer chains, which leads to a decrease in performance."

Coates said that HDPE material loses quality every time it is recycled. "You can't just take these plastics, melt them down. It's not like aluminum which is perfect every time. You have to work hard to make it valuable, to make the plastic useful."

Coates said that recyclers have about 5 cents to value or improve the quality of each pound of recycled plastic.

Currently, recycling facilities improve the quality of recycled products by adding a small amount of virgin plastic. However, the mixture of recycled materials changes daily, making it uncertain how much new plastic needs to be added.

Using less material (and energy) to produce polyethylene, and controlling the quality and physical properties of recycled materials, hinges on understanding how the various lengths of polymer chains in a sample (referred to as its molecular weight distribution) affect its properties. Key factors: its viscosity during the manufacturing process and its strength and toughness as a finished product.

The first step towards smarter, more specific, and sustainable polymer design

DiStasio and his lab members trained their machine learning model, called PEPPr (Polyethylene Properties PRedictor), using a library of over 150 polyethylene samples synthesized and characterized by Coates, Fors, and their lab members.

"We need a library of polymers with different molecular weight distributions," DiStasio said. "We also want the polymers to have a variety of processing and mechanical properties."

The researchers wrote that machine learning capabilities are necessary for the complex task of understanding the relationship between the composition of these materials and their properties.

DiStasio said, PEPPr addresses two problems. If the molecular weight distribution of the HDPE sample is known, the model can predict its properties: melt viscosity, toughness, and strength. It can also be used for the inverse; if a user has a set of target properties in mind, the model can tell them which polymer samples will have these properties.

"If you want to make a plastic bag, the melt properties you need are different from those for making a kayak," Fors said.

The researchers stated that the PEPPr method is the first step towards smarter, more specific polymer design, and more efficient and sustainable recycling processes. They plan to expand the range of predictable properties and add processing methods that could be highly impactful. They also hope to extend the model to include other categories of polymers.

"We should be able to develop these types of models for any kind of commercial polymer," Fors said. "This should also be a general approach for tuning properties and recycling other materials."

This research was supported by the National Science Foundation Center for Sustainable Polymers and the Cornell Center for Materials Research, and was funded by the Undergraduate Research Experience Program.