Multi-Objective Prediction and Discovery of High-Performance Polymer Dielectrics for Advanced Packaging

Abstract

As integrated circuits advance toward heterogeneous integration, higher operating frequency, and greater power density, polymer dielectrics used in advanced packaging must simultaneously provide efficient heat dissipation and low relative dielectric constant. However, the joint optimization of thermal conductivity and dielectric performance remains challenging because these properties are governed by different and often competing molecular structural factors. In addition, the scarcity of reliable experimental data limits the rapid discovery of high-performance polymer materials. In this work, we propose a transfer-learning-based hybrid graph neural network framework for data-efficient joint prediction and multi-objective screening of polymer dielectrics. The model is pretrained on a million-scale polymer dataset to learn transferable molecular representations and is then adapted to limited experimental datasets through a shared encoder and dual prediction heads. By integrating local message passing and global attention mechanisms, the framework captures both local bonding environments and long-range backbone features. The model achieves accurate prediction for both thermal conductivity and dielectric constant and maintains robust performance under limited data conditions. Large-scale virtual screening identifies a set of Pareto-optimal polymer candidates with improved thermal-dielectric balance. Several candidates show enhanced thermal conductivity at comparable dielectric levels, and synthetic accessibility analysis indicates that most of them are easy to synthesize. Furthermore, representative candidates are validated by molecular dynamics and density functional theory calculations, confirming that the predicted trends are physically meaningful. This work provides a scalable and interpretable strategy for discovering advanced polymer dielectrics for heterogeneous integration and chiplet packaging.