Mapping Chemical Structure-Glass Transition Temperature Relationship through Artificial Intelligence

Artificial neural networks (ANNs) have been successfully used in the past to predict different properties of polymers based on their chemical structure and to localize and quantify the intramonomer contributions to these properties. In this work, we propose to move forward in order to use the mathematical framework of the ANN for embedding the chemical structure of monomers into a high-dimensional abstract space. This approach allows us not only to accurately predict the glass transition temperature (T-g) of polymers but, even more important, also to encode their chemical structure as m-dimensional vectors in a mathematical space. For this aim, we employed a fully connected neural network trained with a set of more than 200 atactic acrylates that provide the coordinates of the vectorized chemical structures into the m-dimensional space. These data points were then treated with a hierarchical nonparametric clusterization method in order to automatically group similar chemical structures into clusters with alike properties. These clusters were then projected into a human-readable three-dimensional space using principal component analysis. This approach allows us to deal with chemical structures as if they were mathematical entities and therefore to perform quantitative operations, so far hardly imaginable, being essential for both the design of new materials and the understanding of the structure-property relationships.

Automated summary

This work explores the use of artificial neural networks to predict polymer properties and encode chemical structures. By embedding monomer chemical structures in a high-dimensional abstract space and employing neural network training and clustering methods, accurate prediction of polymer properties and structure encoding were successfully achieved.