Fingerprint applicable for machine learning tested on LCST behavior of polymers

Since the 1990s, there have been many attempts to efficiently decode molecules; however, for polymers, some descriptions have always been neglected. Details like position and occurrence of molecular structural patterns are hardly possible with such techniques. However, consideration of these details can lead to the ability to predict the properties of polymeric materials in a highly efficient manner. In this study, we have succeeded in taking a further step toward a more concise incorporation of this information. We introduce the polymer fingerprint, a way of representing polymers from common experimental characterization data such as size-exclusion chromatography. In addition, we demonstrate the functionality of the representation for the prediction of cloud-point temperatures. The newly created database for this proof consists of various polymers with different architectures and properties, which were extracted from scientific publications. Finally, we evaluate the limitations and potentialities for such representation-prediction models and their future importance.