Polymer property prediction models and training data for "Provenance and Uncertainty Limit Multi-Objective Screening of High-Temperature Dielectric Polymers"

Seven trained models that predict polymer properties from a repeat-unit structure, together with the labels they were trained on and a script for applying them to new structures. The models cover glass transition temperature, chain band gap, molecular frontier-orbital gap, thermal conductivity, Young's modulus, dielectric constant, and refractive index. Each is an ensemble of ten independently seeded gradient-boosted trees operating on a folded Morgan fingerprint concatenated with 59 molecular descriptors. Input is PSMILES carrying exactly two connection points that mark the ends of the repeat unit. Training data comprises 18,712 structure-property labels across 12,498 unique repeat units. Only the glass transition model is trained on experimental values; the remaining six are trained on molecular dynamics or density functional theory labels, whose agreement with measurement ranges from 0.28 for modules to 0.46 for refractive index. Those six should be treated as ranking instruments rather than absolute predictors. Spread across the ten seeds is not an uncertainty estimate, since it understates chemistry-aware error by roughly an order of magnitude. Supporting data and models for the article "Provenance and Uncertainty Limit Multi-Objective Screening of High-Temperature Dielectric Polymers."