Refined Theoretical Models for Predicting Chain-Direction Thermal Conductivity in Crystalline Polymers

Abstract

Thermal conductivity (κ) is a pivotal physical property governing the performance of crystalline polymers in advanced technologies ranging from thermoelectrics to electronic thermal management. However, low-cost and rapid prediction of κ remains challenging due to the structural complexity of polymers and the three-order-of-magnitude span of κ values (0.1–100 W m−1 K−1) across different crystalline polymers. Herein, we present two refined theoretical models for fast prediction of chain-direction κ in PCFF-compatible crystalline polymers, built on a physics-informed framework that combines significant dataset expansion and machine-learning (ML)-driven descriptor optimization. First, we expanded the PCFF-compatible dataset from 42 to 157 crystalline polymers. Second, we leveraged Python libraries and ML tools for descriptor analysis: we constructed 16 physically interpretable descriptors tied to phonon transport properties and unit-cell characteristics and then performed exhaustive ML-assisted screening to identify optimal four-descriptor sets for two complementary fitting routes. Collectively, the models incorporate four key descriptors: backbone rotation ratio (P), fraction of carbon atoms in the unit cell (C), unit-cell packing fraction (Φ), and interchain density of noncovalent atom pairs (ρin). Route I preserves the Slack-model-derived crystalline term f1 for enhanced physical interpretability and only optimizes the correction factor f2, while Route II jointly optimizes f1 and f2 to maximize predictive performance. Both models achieve a balance of simplicity, predictive performance, and mechanistic transparency, with mean absolute logarithmic error (MALE) and root-mean-square logarithmic error (RMSLE) that are lower than those of our previous linear-regression-based model and even approach the predictive performance of ML models (random forest (RF), Gaussian process regression (GPR), and support vector regression (SVR)). Mechanistic analysis reveals that smaller P, C, and Φ, combined with larger ρin, synergistically prolong phonon lifetimes, enhance group velocities, and increase volumetric heat capacity, collectively boosting chain-direction κ. These advancements enable more reliable rapid screening of κ and provide actionable guidelines for the rational design of crystalline polymers with target κ values, advancing the development of polymer-based materials for energy conversion and thermal management applications.