QC-Augmented GNNs for Few-Shot Prediction of Amorphous Polymer Properties via Multi-Scale Microstructures

Accurate prediction of polymer properties is essential for material design. However, traditional methods often struggle with efficiency due to the limited availability of experimental data and insufficient understanding of polymer microstructures, particularly in the case of small datasets for amorphous polymers. To address this hallenge, we introduce a novel paradigm based on "local clusters", structural motifs whose properties are efficiently computed using quantum chemistry (QC) methods. These clusters, simulated across multiple scales, serve as key descriptors that encapsulate essential microstructural features of polymers. By integrating these QC-derived descriptors with graph convolutional networks (GNNs) or neural networks (NNs), we have developed Locluster, a multiscale, microstructure-driven predictive model tailored for data-scarce environments in amorphous polymer research. Notably, Locluster requires only 2–5 descriptors and as few as two dozen training samples to accurately predict critical polymer properties, including density, refractive index, dielectric constant, and glass transition temperature. The model achieves predictive performance comparable to approaches trained on large datasets, thereby effectively mitigating the data scarcity challenge in polymer informatics. This work provides an effective option for the rational design and accelerated discovery of novel polymeric materials.