The increasing power density and miniaturization of organic electronics necessitate efficient heat dissipation. Conventional polymers, being thermal insulators (0.1-0.5 W m⁻¹K⁻¹), hinder this advancement. Therefore, polymers with significantly higher thermal conductivity (TC) are crucial for organic energy storage and electronic devices. Polymer morphology and topology directly influence TC; enhanced crystallite orientation and crystallinity reduce phonon scattering and boost TC along chain directions. Recent studies have demonstrated exceptionally high TC (up to 62 W m⁻¹K⁻¹) in polyethylene (PE) films through chain disentanglement and alignment, highlighting the potential for overcoming the limitations of traditional amorphous polymers. Strong intra-chain atomic interactions are key to improving TC, and experimental techniques like micro-mechanical stretching, electrostatic spinning, and nanoscale templating enhance crystallinity and chain orientation, increasing TC by orders of magnitude. Molecular dynamics (MD) simulations suggest that ordered chains and large radius of gyration (Rg) are favorable for high TC. Debye's theory, k = vgCl (where k is TC, vg is phonon group velocity, C is volumetric heat capacity, and l is mean free path), highlights the role of repeating unit characteristics and backbone bonding strength in determining vg and C. However, the vast chemical space of polymers makes the current trial-and-error approach for discovering high-TC polymers inefficient and costly. Polymer informatics, leveraging machine learning (ML), offers a data-driven approach to overcome this challenge by predicting and designing polymers with specific properties. While ML has shown promise in predicting various polymer properties, creating accurate polymer representations and selecting relevant features remains a significant challenge. Existing studies often rely on graph descriptors, which, while accessible due to readily available toolkits, lack interpretability and make it difficult to understand the relationships between molecular structure and material properties. The paper addresses these challenges by developing an interpretable ML framework for high-throughput screening of high-TC polymers, combining physical descriptors with high-throughput MD simulations.

The literature extensively covers the relationship between polymer structure and thermal conductivity. Studies show that increasing crystallinity and chain alignment significantly improves thermal transport. Experimental techniques such as micro-mechanical stretching, electrostatic spinning, and nano-templating have been employed to achieve this enhancement. Molecular dynamics (MD) simulations have played a vital role in understanding the underlying mechanisms, highlighting the importance of factors like chain stiffness and radius of gyration. However, most studies focus on simple polymer structures, such as polyethylene. The application of machine learning to polymer design is an emerging field with significant potential. Existing efforts have explored the use of graph descriptors and polymer chemistry fragment statistics, but these approaches often lack interpretability, making it challenging to draw clear relationships between molecular structure and properties. The challenge lies in finding a balance between the ability of machine learning models to accurately predict polymer properties and the need for understanding those predictions through interpretability.

The research employs a four-component framework: (1) polymer library construction; (2) MD simulation for TC calculation; (3) monomer feature representation and hierarchical down-selection; and (4) ML model construction for TC prediction. A benchmark dataset was compiled from literature sources and augmented with data from PoLyInfo and PI1M databases. The TC of polymers in the training dataset was computed using MD simulations with the GAFF2 force field. Initially, 320 physical descriptors were generated using Mordred software and force field parameters. These descriptors were then subjected to a hierarchical down-selection process consisting of three stages: (i) removal of features with low variance, (ii) primary filtering using various correlation coefficients (Pearson, Spearman, Distance, and MIC), and (iii) final selection using a random forest (RF) model and recursive feature elimination (RFE). This resulted in a set of 20 optimized descriptors. Three different ML models (RF, XGBoost, and MLP) were trained on this reduced descriptor set to predict the logarithmic TC (log₂TC) of the polymers. The SHAP (SHapley Additive exPlanations) method was used to analyze the contribution of each descriptor to the prediction. To identify promising high-TC polymers, the trained models were applied to the PoLyInfo and PI1M databases. The predicted TC values from the three models were combined to select candidates for further validation through MD simulations. Finally, symbolic regression (SR) using genetic programming was employed to derive mathematical formulas for TC prediction based on the optimized descriptors obtained from the machine learning process. Phonon dispersion relations were computed using phonon spectral energy density (Phonon-SED) analysis to gain insights into the thermal transport mechanisms within the identified high-TC polymers. The thermal transport linkages between individual polymer chains and amorphous polymers were also investigated by calculating the thermal conductivity of amorphous polymers and analyzing the radius of gyration.

The hierarchical down-selection process successfully reduced the dimensionality of the descriptor space, improving the accuracy and interpretability of the ML models. The three ML models (RF, XGBoost, MLP) achieved R² values above 0.80, outperforming models trained with traditional graph descriptors. SHAP analysis identified key descriptors, including cross-sectional area and average dihedral force constant (Kd_average), which are strongly correlated with TC. The ML models successfully identified 107 promising polymer structures with TC > 20.00 W m⁻¹K⁻¹, many of which exhibit a relatively high synthetic accessibility (SA) score. Symbolic regression yielded mathematical formulas that capture the relationship between the optimized descriptors and TC, facilitating rapid screening of high-TC polymer candidates. Phonon dispersion analysis revealed that the high TC of the selected polymers is associated with strong chain stiffness and high phonon group velocities, particularly in π-conjugated systems. Analysis of amorphous polymers showed a strong positive correlation between the radius of gyration (Rg) and thermal conductivity, suggesting that strong intra-chain interactions and large Rg enhance thermal transport in the amorphous state. Energy flux decomposition analysis demonstrated the dominant role of intra-chain interactions (bonds, angles, and dihedrals) in the thermal conductivity of amorphous polymers, particularly for π-conjugated systems.

The findings demonstrate the effectiveness of combining interpretable machine learning and physical descriptors for high-throughput screening of high-TC polymers. The hierarchical down-selection process significantly enhanced model performance, while SHAP analysis provided valuable insights into the key structural features influencing TC. The identification of a large number of promising high-TC polymers, many with high SA scores, offers significant potential for experimental validation and further development of high-performance materials. The derived mathematical formulas provide a rapid screening tool, complementing the more computationally intensive MD simulations. The analysis of phonon dispersion relations and amorphous polymer behavior has provided valuable insights into the underlying mechanisms of thermal transport in these polymers, confirming the importance of chain stiffness, intra-chain interactions, and radius of gyration. This research bridges the gap between computational design and experimental realization of high-TC polymers.

This study presents a novel, interpretable ML framework for the discovery of high-TC polymers. The framework successfully reduced the descriptor space, improved model accuracy, and revealed key structural features driving high TC. The identification and validation of numerous promising candidates, along with the development of predictive mathematical formulas, significantly advance the field. Future research should focus on experimental validation of the predicted polymers and exploration of advanced synthesis techniques to realize the potential of these high-TC materials. Further investigation into the interplay between various structural parameters and thermal transport mechanisms at different length scales could further refine the prediction models and design guidelines.

The accuracy of the MD simulations depends on the accuracy of the force field employed. The GAFF2 force field, while widely used, may not perfectly capture all interatomic interactions, potentially affecting the accuracy of TC predictions. The focus on linear polymer chains in this study limits the applicability of the findings to other polymer architectures. The synthetic accessibility (SA) score, while useful for initial screening, does not fully account for the complexities of polymer synthesis, and some predicted polymers might prove challenging to synthesize in practice. The number of polymers in the training set, although larger than in many previous studies, could still be considered limited, potentially affecting the generalization ability of the ML models to unseen polymer structures.