Novel Representations and Quantitative Structure Property Relationships for Polymers Using Machine Learning

Current developments indicate Machine Learning (ML) can be utilized to identify, design, and optimize polymeric structures for desired properties, vastly reducing the costly and time-consuming experimentation required. However, Polymer Informatics faces the significant challenge of accurately and effectively representing large and complex polymeric structures in a computer-comprehensible manner. To create computationally efficient models, we developed novel representations of polymers that simplify yet retain the molecular structure. From each representation, over 1,800 quantitative structural chemical descriptors were generated, and a novel approach was used to better emulate the structure of long polymers and counteract the compressional effects of the representation used with less computational cost. Pruned subsets of these descriptors were utilized to train over 25 different types of ML Models with Bayesian Hyperparameter Optimization and were evaluated by predicting Glass Transition Temperature (Tg) for 564 polymers and were shown to achieve r-squared values up to 0.74. These promising preliminary results indicate that given the proper dataset, this process could be utilized to create ML models for any other desired property(s).