Polymer property prediction models, evaluation code and derived data for "Separating Model Skill from Chemical Similarity: A Controlled Audit of Generalization Under Chemical Shift in Polymer Property Prediction"

This repository contains the replication package for “Separating Model Skill from Chemical Similarity: A Controlled Audit of Generalization Under Chemical Shift in Polymer Property Prediction.” The study examines how model skill, reference-label dispersion, chemical similarity, and model class affect reported accuracy in polymer property prediction. The analysis uses 43,095 computed property values for 12,271 polymers across 20 properties from the public ADEPT-PolyGraphMT dataset, with independent validation using 7,367 experimental glass transition measurements from PolyMetriX. The package includes processed data, analysis scripts, result tables, trained models. These materials support comparisons of random-split and grouped evaluation, label-dispersion estimates, learning curves, paired model-class comparisons, grouping-robustness tests, and experimental validation. The source datasets are publicly available and are not redistributed. The supplied scripts regenerate the derived files from these sources. Please consult the accompanying documentation for setup instructions and the workflow used to reproduce the manuscript results.