Physics-informed transfer learning via frontier orbital pretraining for prediction of polymer electronic properties.

Accurate prediction of electronic properties, including bandgap, ionization energy (IE), and electron affinity (EA), is central to the design of polymer electronic materials but is hindered by the vast chemical space and the high cost of reliable reference data. Here, a frontier orbital-guided learning framework is proposed that integrates low-cost quantum chemical pretraining with transfer learning to enable efficient and physically consistent prediction of polymer electronic properties. The model is pretrained on GFN2-xTB-derived frontier orbital properties of polymer trimers and subsequently fine-tuned using limited highfidelity data to predict chain bandgap (bandgap-chain), bulk bandgap (bandgap-bulk), IE, and EA. The resulting models exhibit consistently high predictive accuracy across all target properties, with test-set mean absolute errors of 0.246 eV for bandgap-chain, 0.269 eV for bandgap-bulk, 0.169 eV for IE, and 0.136 eV for EA, corresponding to RMSE values below 0.360 eV, while maintaining strong correlation with reference data (R2 > 0.90) and preserving key physical behaviors, including chain-length scaling and inter-property consistency. Leveraging this framework, electronic properties of ∼12 × 106 polymer repeat units are predicted, enabling statistically robust fragment-level analysis in which the observed trends remain consistent with established physical intuition and known structure-property relationships. This work provides a scalable and data-efficient framework for machine learning-assisted screening and design of polymer electronic materials.