(33l) Predictive Modeling of Aligned Polymer Thermal Conductivity: Integrating Simulation and Machine Learning

Ultra-aligned polymers are promising candidates for anisotropic thermal management and offer a lightweight, chemically tunable route to discovering novel high-TC materials. Building on atomistic MD studies that revealed significant chain-length enhancements in polymer thermal conductivity (TC)—notably in single-chain polyethylene—earlier work shows evidence of ballistic-like phonon transport at the nanoscale. However, many efforts either focus on single polymer chains—where non-bonded interactions are neglected and long-wavelength modes remain largely undamped—or consider bulk systems only at relatively short lengths (still in the ballistic regime) or through expensive ab initio/equilibrium calculations. Consequently, the onset of diffusive behavior, marked by plateauing TC and driven by anharmonic phonon–phonon scattering, remains underexplored in non-equilibrium molecular dynamics (NEMD) simulations for bulk polymers at micrometer scales.

To address this gap, we employ shorter length-scale (~40 nm) screening to identify potential high-TC polymers. Using the GAFF2 force field, we first simulate bulk systems from the Pl1M polymer dataset comprising approximately 40,000 candidates (down-selected from an initial one million by excluding large sidechains known to suppress TC). For instance, polyethylene at 4–50 nm chain lengths exhibits TC values of 4–50 W/mK, driven primarily by length-dependent effects. We then deploy an active learning workflow: a low-fidelity model based on Mordred descriptors identifies promising monomers, which are validated via simulation and iteratively retrained until meeting an uncertainty criterion; a subsequent higher-fidelity model incorporating MD features achieves an average R² of ~0.86.

Next, top performers are simulated up to 1 μm with roughly nine aligned chains (~200,000 atoms) to capture bulk phonon–phonon scattering. For polyethylene, TC saturates near 200 W/mK with length, aligning with ab initio benchmarks and exceeding typical experimental values (70–100 W/mK) due to real defects caused by polydispersity, defects, and processing conditions. To explain the length dependence quantitatively, we modified the thermal conductivity model originally developed for carbon nanotubes by introducing a polymer-specific ballistic scaling exponent, which highlights that the length-dependent TC of polymers cannot be explained by previous models. Additionally, ring-based and nitrogen-containing backbones emerge as strong candidates, likely owing to higher group velocities from rigid bonds and reduced internal scattering. SHAP analyses informed by Slack theory confirm group velocity as a principal driver of TC, although other structural information content features also modulate phonon dispersion relation and ultimately phonon-phonon scattering. Although these estimates represent upper bounds for homopolymers in ideal morphologies, they guide future work on blends, composites, and tailored structural modifications. By explicitly capturing the ballistic-to-diffusive transition in aligned polymers at large length scales, our study demonstrates that combining active learning with large-scale NEMD enables a balance between rapid exploration of chemical space and the detailed thermal transport mechanisms at large length scales -- ultimately guiding efficient materials screening and design.