Correlating near infrared data for improved polyolefin recycling

Polyolefins, the most ubiquitous polymers, continue to be difficult to sort. While the state-of-the-art technique for sorting polymers on an industrial scale, near infrared (NIR) spectroscopy, is fast, it is challenging to differentiate polyolefins by type (e.g., polypropylene, high density polyethylene, low density polyethylene, linear low density polyethylene). Since polyolefins are chemically similar but have diverse topologies leading to properties such as semi-crystallinity that can vary dramatically, the properties post recycling can be significantly degraded if polyolefins are not sorted. Here, we use slow, but precise measurement techniques to build a dataset of polyolefins and their corresponding properties. Using machine learning on the NIR data, we first demonstrate that we can identify the different types of polyolefins. We then show that the machine learned latent space of the NIR measurements can be correlated with physical quantities such as density, crucially connecting the fast, industrial measurement to the underlying physics. This connection can ultimately lead to enhanced sorting and thus improved recycling of polyolefins.