Small Data Machine Learning: Classification and Prediction of Poly(ethylene terephthalate) Stabilizers Using Molecular Descriptors
Aaron L. Liu, Rahul Venkatesh, Michael McBride, Elsa Reichmanis, J. Carson Meredith, Martha A. Grover
Abstract
Experimental data from a patent were analyzed to learn about the small molecule additives that were most effective in mitigating the degradation of polyethylene terephthalate. Two sets of molecular descriptors were calculated for a dataset of 39 additive candidates; unsupervised and supervised analyses were performed to determine the most influential structural features that led to reduced degradation. A clustering approach revealed evidence that performance differences had some structural pattern dependence on the molecular descriptors that were employed. To pinpoint the features responsible for those physical differences, a reduced design region approach was applied to analyze descriptors both individually and in multiple dimensions to determine the effectiveness in a binary classification of high and low performances. For each molecular descriptor type, two or three influential descriptors were identified and justified with respect to the additive performance and physicochemical ability to mitigate degradation. Random forest models were constructed with relatively high predictability for both MACCS-166 (AUC = 0.86) and alvaDesc molecular descriptors (AUC = 0.93). We compare molecular descriptor methods for their ability to construct classifiers and to prioritize experimental work toward building a rich dataset. We find that, in small materials datasets, understanding the underlying physicochemical behavior is indispensable for validating the effectiveness of machine learning models.