Polymer Informatics in the Era of Foundation Models: From Representation to Autonomous Design
- 1Department of Mechanical Engineering
- 2University of Wisconsin–Madison„ Madison
Abstract
Foundation models, pretrained on broad datasets and adapted to multiple tasks, offer a way to reuse chemical knowledge across polymer property prediction and inverse design. Realizing this opportunity requires connecting learned representations to materials whose properties also depend on molecular populations, formulation, processing, and measurement conditions. Here, we review advances in polymer foundation models and the data, representations, and experimental workflows that support their use in materials design. We organize these developments around the polymer material state, a description linking chemical identity to molecular distributions, architecture, morphology, and preparation history relevant to a target property. This organization connects data resources and pretraining strategies to questions about what models learn, how that knowledge transfers, and which experimental choices it can support. We compare learning objectives and adaptation methods, examine representations of polymer chemistry and state, and follow inverse design from candidate generation through synthesis and processing to validation. Pretrained models now support prediction across multiple properties and generation for specified targets, with selected polymer candidates reaching synthesis and thermal characterization. Broader polymer design campaigns show how precursor accessibility, polymerization outcomes, and specimen preparation influence whether a predicted advantage becomes a measured material response. Building on this evidence, we discuss opportunities to combine chemical pretraining with richer sample information and to assess transfer across polymer families, preparation conditions, and experimental campaigns. The resulting synthesis connects model development to a practical objective for polymer informatics: reducing the new experimentation needed to obtain materials with reproducible, targeted performance.
Related
- A review on foundation models for polymer informatics
- Scalable Thermoset Informatics Integrated with Autonomous Experimentation for On-Demand Discovery
- AdaptDelivery: a polymer informatics platform for predicting structural properties of polymers via machine learning
- Polymer Informatics Atlas: Property-Specific Representations, Chemistry-Aware Generalization, and Reliability-Aware Machine Learning Across Five Polymer Properties
- PolyNet: A Python Package for Polymer Informatics
BibTeX
@misc{he2026polymer,
title = {{Polymer Informatics in the Era of Foundation Models: From Representation to Autonomous Design}},
author = {Jianxin He and Ying Li},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009334/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009334/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Polymer Informatics in the Era of Foundation Models: From Representation to Autonomous Design
AU - Jianxin He
AU - Ying Li
PY - 2026
DA - 2026/09/23
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009334/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009334/v1
AB - Foundation models, pretrained on broad datasets and adapted to multiple tasks, offer a way to reuse chemical knowledge across polymer property prediction and inverse design. Realizing this opportunity requires connecting learned representations to materials whose properties also depend on molecular populations, formulation, processing, and measurement conditions. Here, we review advances in polymer foundation models and the data, representations, and experimental workflows that support their use in materials design. We organize these developments around the polymer material state, a description linking chemical identity to molecular distributions, architecture, morphology, and preparation history relevant to a target property. This organization connects data resources and pretraining strategies to questions about what models learn, how that knowledge transfers, and which experimental choices it can support. We compare learning objectives and adaptation methods, examine representations of polymer chemistry and state, and follow inverse design from candidate generation through synthesis and processing to validation. Pretrained models now support prediction across multiple properties and generation for specified targets, with selected polymer candidates reaching synthesis and thermal characterization. Broader polymer design campaigns show how precursor accessibility, polymerization outcomes, and specimen preparation influence whether a predicted advantage becomes a measured material response. Building on this evidence, we discuss opportunities to combine chemical pretraining with richer sample information and to assess transfer across polymer families, preparation conditions, and experimental campaigns. The resulting synthesis connects model development to a practical objective for polymer informatics: reducing the new experimentation needed to obtain materials with reproducible, targeted performance.
ER -