Multimodal Machine Learning for Soft High-k Elastomers under Data Scarcity

TL;DR

The paper tackles predicting dielectric and mechanical properties ($k$ and $E$) of soft high-$k$ elastomers under severe data scarcity by curating a compact acrylate-based dataset and introducing a multimodal framework that fuses pretrained sequence-based polymer representations with graph-based encodings. By leveraging large-scale polymer corpora, the model transfers rich chemical and structural knowledge to the small target domain, enabling robust few-shot learning with two fusion strategies and a Gaussian Process regression head. Key findings show that pretrained representations outperform handcrafted descriptors, and that latent-space aligned early fusion yields the strongest predictive performance (mean $R^2$ ≈ $0.758$; RMSE ≈ $13.659$), demonstrating the value of cross-modal representation alignment in low-data regimes. The work provides a foundation for data-efficient discovery of soft high-$k$ dielectric elastomers and can be extended to other backbones and composite formulations, with open data and code to accelerate community adoption.

Abstract

Dielectric materials are critical building blocks for modern electronics such as sensors, actuators, and transistors. With the rapid recent advance in soft and stretchable electronics for emerging human- and robot-interfacing applications, there is a surging need for high-performance dielectric elastomers. However, it remains a grand challenge to develop soft elastomers that simultaneously possess high dielectric constants (k, related to energy storage capacity) and low Young's moduli (E, related to mechanical flexibility). While some new elastomer designs have been reported in individual (mostly one-off) studies, almost no structured dataset is currently available for dielectric elastomers that systematically encompasses their molecular sequence, dielectric, and mechanical properties. Within this context, we curate a compact, high-quality dataset of acrylate-based dielectric elastomers, one of the most widely explored elastomer backbones due to its versatile chemistry and molecular design flexibility, by screening and aggregating experimental results from the literature over the past 10 years. Building on this dataset, we propose a multimodal learning framework that leverages large-scale pretrained polymer representations from graph- and sequence-based encoders. These pretrained embeddings transfer rich chemical and structural knowledge from vast polymer corpora, enabling accurate few-shot prediction of both dielectric and mechanical properties from molecular sequences. Our results represent a new paradigm for transferring knowledge from pretrained multimodal models to overcome severe data scarcity, which can be readily translated to other polymer backbones (e.g., silicones, urethanes) and thus accelerate data-efficient discovery of soft high-k dielectric elastomers. Our source code and dataset are publicly available at https://github.com/HySonLab/Polymers

Figures (3)

• Figure 1: Distributions of dielectric constant ($k$) and Young’s modulus ($E$) across all curated acrylate-based dielectric elastomers. Most samples exhibit moderate-to-low dielectric constant and low modulus, with very few high-$k$ outliers indicating enhanced polarization mechanisms.
• Figure 2: Overview of the proposed multimodal learning framework for elastomer property prediction. (A) Late-fusion: sequence-based and graph-based polymer representations are obtained from pretrained Polymer Language Models and a pretrained GNN encoder, respectively. Each embedding is passed through a Gaussian Process Regressor (GPR), and the final prediction is obtained via a weighted linear combination of the two modality-specific outputs. (B) Latent-space aligned early fusion: pretrained encoders first generate modality-specific embeddings, which are mapped into a shared representation space using lightweight MLP projection heads trained via latent-space alignment. The aligned embeddings are fused and subsequently passed to a shared GPR for joint prediction of dielectric constant and Young’s modulus.
• Figure 3: Correlation between predicted and experimental values for the two target properties using the best-performing multimodal configuration, namely latent-space aligned early fusion with element-wise averaging.