Structure-Informed Reconstruction of Sparse Polymer Property Profiles

Polymer discovery often requires comparing candidates across multiple interdependent properties, but experimentally reported polymer property records are frequently incomplete.Such profile-level sparsity limits multi-property screening by leaving many materials with only partially reported property profiles.Here, we present PolyTab, a structure-informed framework for reconstructing partially observed polymer property profiles.PolyTab combines PSMILES-derived embeddings from a pretrained polymer language model with observed property values, missingness masks and statistical priors.A quartile-based classifier provides distribution-aware initial estimates, which are further refined by stacked adaptive regression layers while preserving observed entries.Across heterogeneous polymer datasets and sparse per-polymer retention settings, PolyTab outperforms conventional statistical, tree-based and tabular-learning baselines, achieving a median 𝑅 2 of approximately 0.8 under the 𝑅 25 setting and 𝑅 2 values above 0.977 for all five gas-permeability properties.PolyTab also preserves inter-property correlation patterns more accurately than baseline methods, remains stable under noise perturbations, and improves downstream prediction when reconstructed tables are used for supervised learning.Case studies involving 6FDA-based gas-separation membranes and Polyolefin-based functional films illustrate how structure-informed reconstruction can retain otherwise incomplete experimental records for preliminary multi-criteria screening.