Polymer Informatics Atlas: Property-Specific Representations, Chemistry-Aware Generalization, and Reliability-Aware Machine Learning Across Five Polymer Properties
- 1Department of Industrial Chemistry, Faculty of Applied Physical Sciences
- 2Enugu State University of Science and Technology (ESUT
Abstract
Polymer-property machine learning is often evaluated as if representation choice, random-split accuracy, chemical extrapolation, transfer learning, uncertainty, and chemical interpretation were a single problem. Here, these questions are separated in a six-stage Polymer Informatics Atlas built from 7,973 unique RDKit-valid polymer repeat units in the NeurIPS Open Polymer Prediction 2025 dataset. Thirty-six interpretable composition/structure descriptors and radius-4, 2,048-dimensional count Morgan fingerprints were evaluated independently and as a hybrid representation for glass-transition temperature (Tg), fractional free volume (FFV), thermal conductivity (Tc), density, and radius of gyration (Rg). Frozen property-specific models achieved untouched random-holdout R2 values of 0.608, 0.704, 0.825, 0.920, and 0.703, respectively. However, chemistry-separated testing changed this interpretation substantially: Tc fell to R2 = 0.326, whereas density remained comparatively robust at R2 = 0.874. Representation-block permutation showed that density was predominantly captured by interpretable global chemistry, while Tc and Rg depended more strongly on local/topological structure; Tg and FFV required mixed information. Masked multitask learning on the overlapping Tc-Density-Rg block used incomplete labels more efficiently than complete-case training, but gains over matched single-task neural networks were small and seed dependent. Five-member ensembles combined with splitconformal calibration produced 90% nominal intervals with empirical coverages of 95.7% (Tc), 92.3% (density), and 98.5% (Rg). Finally, 28 predictive Morgan indices were decoded to atom-centred environments after exact fingerprint-regeneration audits, and an integrated descriptor/fingerprint manifold was used for chemical-space visualization, archetype selection, and an illustrative high-Tc/low-density Pareto analysis. The results support a property-specific, domain-aware view of polymer informatics in which predictive accuracy is reported together with chemical support, uncertainty, and explicit claim boundaries rather than as a single global score.
Related
- Accept or abstain? A pre-specified test of reliability gates calibrated on held-out chemistry for molecular property and reaction-yield models
- AdaptDelivery: a polymer informatics platform for predicting structural properties of polymers via machine learning
- Systematic evaluation of LLM-based zero-shot model construction in chemistry and materials science: predictive performance and response reliability
- A Mini-Review on Thalidomide: Chemistry, Mechanisms of Action, Therapeutic Potential and Anti-Angiogenic Properties in Multiple Myeloma
- Coordination Chemistry of Boron Subphthalocyanine With BINOL Derivatives: Unexpected Oxidation to Oxa[4]Helicene and Properties of the Resulting Axially Coordinated Complexes Containing Oxa[4]Helicene and/or BINOL Derivatives
BibTeX
@misc{orji2026polymer,
title = {{Polymer Informatics Atlas: Property-Specific Representations, Chemistry-Aware Generalization, and Reliability-Aware Machine Learning Across Five Polymer Properties}},
author = {Dennis Obinna Orji},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008969/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008969/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Polymer Informatics Atlas: Property-Specific Representations, Chemistry-Aware Generalization, and Reliability-Aware Machine Learning Across Five Polymer Properties
AU - Dennis Obinna Orji
PY - 2026
DA - 2026/09/17
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008969/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008969/v1
AB - Polymer-property machine learning is often evaluated as if representation choice, random-split accuracy, chemical extrapolation, transfer learning, uncertainty, and chemical interpretation were a single problem. Here, these questions are separated in a six-stage Polymer Informatics Atlas built from 7,973 unique RDKit-valid polymer repeat units in the NeurIPS Open Polymer Prediction 2025 dataset. Thirty-six interpretable composition/structure descriptors and radius-4, 2,048-dimensional count Morgan fingerprints were evaluated independently and as a hybrid representation for glass-transition temperature (Tg), fractional free volume (FFV), thermal conductivity (Tc), density, and radius of gyration (Rg). Frozen property-specific models achieved untouched random-holdout R2 values of 0.608, 0.704, 0.825, 0.920, and 0.703, respectively. However, chemistry-separated testing changed this interpretation substantially: Tc fell to R2 = 0.326, whereas density remained comparatively robust at R2 = 0.874. Representation-block permutation showed that density was predominantly captured by interpretable global chemistry, while Tc and Rg depended more strongly on local/topological structure; Tg and FFV required mixed information. Masked multitask learning on the overlapping Tc-Density-Rg block used incomplete labels more efficiently than complete-case training, but gains over matched single-task neural networks were small and seed dependent. Five-member ensembles combined with splitconformal calibration produced 90% nominal intervals with empirical coverages of 95.7% (Tc), 92.3% (density), and 98.5% (Rg). Finally, 28 predictive Morgan indices were decoded to atom-centred environments after exact fingerprint-regeneration audits, and an integrated descriptor/fingerprint manifold was used for chemical-space visualization, archetype selection, and an illustrative high-Tc/low-density Pareto analysis. The results support a property-specific, domain-aware view of polymer informatics in which predictive accuracy is reported together with chemical support, uncertainty, and explicit claim boundaries rather than as a single global score.
ER -