Interpretable substructure-based screening of multi-property polymer dielectrics with prompt-ready rules for rational design

Abstract High-performance polymer dielectrics remain difficult to discover because transparent connections between molecular substructures and macroscopic properties are lacking. Existing machine-learning approaches often rely on black-box representations that obscure the chemical origins of thermal and electrical behavior. Here, we introduce sliding threshold correlation analysis (STCA), a white-box framework that identifies threshold-dependent substructure combinations governing polymer performance. By scanning property thresholds, STCA reveals hierarchical motifs that remain predictive across thermal (glass-transition temperature, melting temperature, decomposition temperature, and thermal conductivity) and electrical (electrical conductivity and alternating-current dielectric constant) domains, enabling unified multi-property screening. STCA shows strong performance relative to common interpretability baselines in discriminating high- and low-performance regimes, while its threshold-continuous signed rule families remain applicable in selected stringent screening conditions. The extracted combinations align with polymer physics, capturing rigidity-driven thermal stability and limited delocalization in insulating systems. To demonstrate practical utility, we convert these combinations into chemically grounded prompts for a commercial large-language model, yielding synthetically accessible monomers with the anticipated characteristics. DFT and molecular-dynamics evaluations support representative thermal and electrical trends. STCA therefore provides a transparent framework for rule-based polymer screening and prompt-ready candidate generation within the linear-homopolymer repeat-unit domain.