PolyBench: A Reproducible Open Benchmark Platform for Repeat-Unit-Based Polymer Property Prediction

Abstract

Polymer informatics currently lacks shared benchmarks that enable reproducible comparison of property prediction methods across datasets, representations, and models. Here, we present PolyBench, a public benchmark platform comprising 39 literature-derived prediction tasks with 42,169 curated structure–property records covering 14,717 distinct repeat-unit structures, seven repeat-unit molecular representation types (five fingerprint families, physicochemical descriptors, and task-trained graph embeddings), 30 supervised baseline models (22 traditional machine-learning and 8 deep-learning methods), fixed train/test splits, and an interactive leaderboard. Across more than 10,000 supervised model evaluations spanning all 39 tasks, physicochemical descriptor-based tree ensembles emerge as strong baselines under the PolyBench protocol, though no single method achieves consistent superiority across all properties; matched repeated-split analyses on three representative tasks show that the winning configuration is stable across five seeds for the two larger tasks but not for the smallest, condition-sensitive task. The current release evaluates repeat-unit-centered polymer property prediction and does not explicitly encode higher-order factors such as molecular-weight distributions, tacticity, morphology, processing history, device architecture, or measurement protocol; tasks are therefore grouped into three provisional interpretation tiers, with condition-sensitive tasks explicitly treated as exploratory repeat-unit baselines. PolyBench is intended as a reproducible infrastructure resource for controlled method comparison, not as a complete model of chain-, sample-, or process-dependent polymer behavior, and is available at http://polybench.ciac.jl.cn (http://202.98.16.11)