Unlocking Properties of New Polymers Using Machine Learning

Predicting the mechanical properties of new polymers after they have been processed into a product can be tricky and typically involves destructive physical testing. Researchers at the National Institute for Materials Science (NIMS) in Tsukuba, Japan, have developed a technique harnessing machine learning and X-ray diffraction that, they say, offers a nondestructive alternative to conventional polymer testing methods. The study is published in Science and Technology of Advanced Materials.

Led by Dr. Ryo Tamura, Dr. Kenji Nagata, and Dr. Takashi Nakanishi, the research team developed the method on homo-polymer polypropylene groups using X-ray diffraction patterns of the polymers under different preparation conditions to provide detailed information about their complex structure and features.

Importance of representative descriptors

“Machine learning can be applied to data from existing materials to predict the properties of unknown materials,” explained the researchers. “However, to achieve accurate predictions, it’s essential to use descriptors that correctly represent the features of these materials.”

Thermoplastic crystalline polymers, such as polypropylene, have a particularly complex structure that is further altered during the injection molding process. It was, therefore, important for the researchers to adequately capture the details of the polymers’ structure via X-ray diffraction and to ensure that the machine learning algorithm could identify the most important descriptors in that data.

Mechanical properties analyzed

To that end, they analyzed two datasets using a tool called Bayesian spectral deconvolution, which can extract patterns from complex data. The first dataset was X-ray diffraction data from 15 types of homo-polypropylenes subjected to a range of temperatures; the second comprised data from four types of homo-polymer polypropylene that underwent injection molding. The mechanical properties analyzed included stiffness, elasticity, heat-deflection temperature, and tensile strength.

The team found that the machine learning analysis accurately linked features in the X-ray diffraction imagery with specific material properties of the polymers. Some of the mechanical properties were easier to predict from the X-ray diffraction data; other properties such as tensile strength — the point at which a material breaks when it is stretched or pulled — were more challenging, according to the NIMS news release.

The researchers also suggested that their Bayesian spectral deconvolution approach could be applied to other data, such as X-ray photoelectron spectroscopy, and used to understand the properties of other materials, both inorganic and organic.

“It could become a test case for future data-driven approaches to polymer design and science,” they said.