Hebbian Learning on Small Data Enables Experimental Discovery of High T_gPolyimides

Abstract

We report a study combining computational design and experimental evaluation of polyimides with high glass transition temperatures: Tg between 220 °C and 500 °C. The computational approach is based on the recently introduced competitive learning algorithm, supervised self-organizing maps (SUSI), which we recast as an ensemble method, e-SUSI. We use e-SUSI to solve both unsupervised and supervised/semisupervised learning tasks capturing structure-property relationships of high-Tg polyimides historically studied at Almaden Research Center. Predictors trained on historical data were applied to the combinatorial library of novel polyimides and informed selection of the candidates for synthesis and characterization. In this manner, three new polyimides were prepared with Tg values 281 °C, 282 °C, and 331 °C. The measured values closely agree with the predicted values 273 °C, 311 °C, and 335 °C, respectively. We discuss specific reasons that make the proposed computational design strategy attractive in rapid, deliverable-driven efforts with limited, small-batch data sets.