Polymer property prediction has long suffered from a lack of open, standardized benchmarks, making rigorous comparison of machine‑learning methods difficult. Existing resources cover only a narrow subset of polymer architectures, such as homopolymers. To address this gap we introduce PolyBench26, an open dataset containing roughly 250,000 polymer‑property datapoints across eight physical properties. The data are drawn from experimental measurements, density functional theory (DFT), and molecular dynamics (MD).
The benchmark defines four evaluation tasks covering homopolymers, alternating copolymers, random copolymers, and block copolymers: (1) in‑distribution property prediction; (2) scaling with dataset size; (3) impact of repeat‑unit complexity; and (4) transfer to held‑out polymer architectures.
We compare three families of models—language models, graph‑based models, and descriptor‑based models. Results show that graph‑based models achieve the lowest prediction errors across all tasks, retain their advantage as training‑set size varies, and remain robust when repeat‑unit complexity increases.
PolyBench26 provides a reproducible foundation for developing models in the increasingly complex polymer design space. The benchmark and code are released open‑source at https://github.com/rlearsch/PolymerBenchmark2026.
Review