The dielectric constant is a fundamental property of materials, which governs their efficiency in various applications like energy storage, microelectronics, and high-voltage insulation. However, predicting the dielectric permittivity remains a challenge due to the complex interplay between molecular structure, processing conditions, and external factors such as temperature and frequency. In this work, we describe a machine learning-based approach for estimating the dielectric constant of polymers by using their chemical structure. We employed a curated dataset of nearly 1000 polymeric materials, from which we extracted unit cell parameters, atomic features, and tokenized atom-wise descriptors. These features were used to train different predictive models, which integrate global structural attributes with local atomic embeddings to establish structure–property relationships with strong accuracy. We show that with this codification, a simple Random Forest approach can outperform a more computationally expensive neural network (ANN). Additionally, we implemented an extension of this approach to also handle SMILES-based polymer representations, allowing approximated predictions for molecular structures without available crystallographic data. This study highlights the potential of data-driven approaches for accelerating the discovery of novel dielectric polymers, providing a computational tool that can complement experimental efforts in materials design.

Keywords