QC-Augmented NN/GNN for Few-Shot Prediction of Amorphous Polymer Properties Via Multi-Scale Microstructures

Abstract

Accurate prediction of polymer properties is essential for accelerating materials design. However, deep learning techniques, as well as traditional density functional theory (DFT) and molecular dynamics (MD) methods, often face limitations due to the scarcity of experimental data and insufficient understanding of amorphous polymer microstructures, particularly in few-shot learning scenarios. To address this challenge, we introduce a paradigm based on “local clusters,” structural motifs whose properties can be efficiently computed using low-cost quantum chemical (QC) methods. These clusters, simulated across multiple scales, serve as key descriptors that capture essential microstructural features of amorphous polymers. By integrating QC-derived descriptors with graph convolutional networks (GNN) or neural networks (NN), we developed Locluster, a multiscale, microstructure-informed predictive framework tailored for few-shot learning scenarios in amorphous polymer research. Notably, Locluster eliminates the need for full-scale QC simulations of entire polymers and requires only 2–5 descriptors and as few as two dozen training samples to accurately predict critical properties such as density, refractive index, dielectric constant, and glass transition temperature. The model achieves predictive accuracy comparable to large-data approaches and can generate predictions for 100–200 polymer candidates within 24 h on a single 128-core server, making it well-suited for rapid iteration and design updates in early- to midstage polymer development. This work offers an efficient and practical strategy for the rational design and accelerated discovery of amorphous polymeric materials.