More Bang for Your Bond: Small-Molecule Kinetics as Predictors of Polymer Stability

Thermally stable polymers are essential for many high-performance applications, yet traditional experimental methods for studying stability are material-intensive, while machine learning models remain limited by the scarcity of reliable degradation data. We hypothesized that computational small-molecule kinetics could bridge this gap owing to the local character of degradation initiation and the ease of collecting data for molecular systems. To test this hypothesis, we curated computational kinetic data and developed models trained exclusively on small molecules to predict the relative thermal stability rankings of polymers containing C, H, O, N, F, and Cl. Experimental degradation data from 41 common polymers and ∼7k small molecules were used to benchmark the ability of small-molecule kinetics to predict polymer behavior without additional training data. All of the resulting models show remarkable transferability between the small-molecule results and polymer degradation metrics. We find that, while chemically diverse training data generally improve predictions, the accuracy of the underlying small-molecule kinetics is also crucial for maximizing transferability to polymer predictions. Notably, models trained on higher-quality alkane data with consistent kinetic parameters perform comparably or better than those trained on more diverse but incomplete data sets for certain polymer predictions. This suggests that accurately capturing fundamental degradation mechanisms in simple systems may be more valuable than incorporating diverse but potentially inconsistent chemical data. These results underscore significant opportunities for training accurate stability models that can be used for early-stage material design.