Extracting knowledge from polymeric data via AI

Artificial intelligence (AI) has successfully identified complex relationships, predicted polymer properties and generated polymer new chemistries. However, most AI models are black boxes that provide predictions but not interpretability, which can impede scientific progress. As an alternative, we demonstrate the utility of symbolic regression, an AI technique that discovers analytical equations, in the context of polymer science. The resulting equations provide both interpretability and a path toward the generation of knowledge as equations can be rationalized in a way that most AI models cannot. Specifically, we demonstrate how symbolic regression is robust to measurement uncertainty, how it can be used to learn fundamental physics across different chemistries and how it can provide interpretability in the form of simple design rules. Finally, we comment on how to identify scenarios in which symbolic regression is more likely to yield simple equations enabling the extraction of knowledge.