Interpretable Machine Learning Prediction of the Dielectric Constant and Bandgap of Polymers for Flexible Electronics.

Polymer dielectrics are central to flexible and printed electronics, where a material must combine a sufficiently high dielectric constant with a wide electronic bandgap to suppress leakage. Experimental or first-principles screening is slow, motivating data-driven surrogates. Here we develop an interpretable machine learning workflow that predicts both the total dielectric constant and the HSE bandgap of polymer repeat units directly from a monomer structure, using an open density-functional-theory dataset of 284 four-block polymers. Polymers are encoded with 217 RDKit descriptors and a 1024-bit Morgan fingerprint; four regressors are benchmarked under nested five-fold cross-validation with paired significance testing. The bandgap reaches R2=0.83±0.04 (MAE =0.35 eV) and the total dielectric constant R2=0.64±0.06 (MAE =0.44), but paired tests find the models statistically indistinguishable for the bandgap. Decomposing the permittivity explains its lower ceiling: the ionic component is only 15% of the magnitude yet carries 28% of the squared error, and learning curves confirm a representational rather than a data-quantity limit. Read through the Penn relation, the SHAP descriptors yield an explicit design rule-raise permittivity with polar, non-conjugated motifs rather than extended conjugation. Screening 571 unseen candidates with bootstrap uncertainties, an applicability domain and a threshold sensitivity analysis nominates carbamate/urea-type wide-bandgap high-k repeat units.