Machine learning techniques to assess mechanical similarity in fluorinated thermoplastics | Poster Board #S02

Fluorinated thermoplastics are commonly used in many industries due to their extraordinary mechanical properties and high chemical resistance. However, some health and environmental concerns have surfaced regarding the use and production of per- and polyfluoroalkyl substances (PFAS), which has led to restrictions and bans on the use of these chemicals. PFAS have been vital to fluoropolymer synthesis, which makes it crucial to examine application-specific properties of fluoropolymers synthesized with and without PFAS. To assist in finding adequate polymer substitutes, various machine-learning techniques have been applied to mechanical property data collected via rheology for various polymer lots at various temperatures. The rheology data consists of viscosity measurements, storage modulus, and loss modulus over a range of shearing frequencies. We used principal component analysis (PCA) and linear discriminate analysis (LDA) to reduce the dimension of the data to two dimensions for visual clustering of the polymers. PCA is an unsupervised dimension reduction algorithm that finds the directions of maximal variance in the data. LDA is a supervised dimension reduction algorithm that minimizes the distance between the data points from the same polymer and maximizes the distance between data points from different polymers. We found that LDA is a more efficient and accurate technique for classifying polymer lots using rheological data than PCA. The principal components from LDA found that viscosity measurements taken at higher shear rates provide the LDA model with the most useful information to maximize class separation. These findings enable more efficient and accurate means to characterize fluorinated polymers and could aid in identifying mechanically similar polymers.