(145c) Investigating Polymeric Compositions Via Clustering of Production Analytical Characterization Data

In this work, we have explored how multivariate data analysis can improve understanding of complex polymeric structures. Our investigation into silicone resins used analytical characterization data from 14 property indicators in an unsupervised learning problem. The data was used to further understand silicone resin structures via the inter-relationships of the different test methods and the implications of a feature pattern (test result in 14-dimensional space) on the downstream performance of the material. In this way, the analysis can indicate either a truly stochastic distribution throughout the physical property space or distinct resin structure classes. It is our hypothesis that by using the patterns of the data in all the dimensions to define a structure class for a silicone resin, we may capture a more sophisticated and accurate representation of the actual molecular composition and quality.

Since the underlying structure of the resin property data was unknown, eight clustering algorithms were tested with several distance metrics. These included mixture modeling, density-based methods, centroid-based methods such as k-means, and several hierarchical methods. The clusters were optimized in each case, then compared to each other with a variety of cluster validity indices. The results showed that the resin data was not neatly separated into distinct clusters, but at least one clustering solution was able to explain variation in downstream product performance. Multivariate visualizations of the cluster properties are used to help chemists identify the unique resin composition associated with high-performing clusters.