The global energy crisis necessitates efficient thermal energy use. Polymers, while cost-effective and versatile, often exhibit low thermal conductivity (TC). Improving the intrinsic TC of polymers is crucial for applications like high-power electronics. Current approaches involve adding high-TC fillers, but enhancing the polymer matrix itself is vital. Amorphous polymer TC arises from intramolecular (bonded) and intermolecular (non-bonded) interactions, with bonded interactions dominating. Strategies to increase TC include polymer chain alignment (achieving high TC but limited to the alignment direction) and modifying non-bonded interactions through techniques like π–π stacking or blending polymers with strong intermolecular interactions. Polymer blending offers tunability but presents conflicting experimental results, with some blends exceeding the rule of mixtures and others following it. Molecular dynamics (MD) simulations provide an alternative, but the vast chemical space of polymer blends makes exhaustive simulation computationally expensive. Machine learning (ML) and polymer informatics (PI) offer potential solutions, but most PI research focuses on single-component polymers. This study combines high-throughput MD simulations with active learning (AL) to efficiently explore the chemical space of polymer blends and uncover those with superior TC.

Prior research has explored various methods to enhance the thermal conductivity of polymers, including the introduction of high-TC fillers and modification of the polymer matrix itself. Studies on polymer chain alignment through stretching have demonstrated significant increases in thermal conductivity, but this method is often incompatible with conventional processing techniques. Other studies have focused on engineering the non-bonded inter-chain interactions, such as π–π stacking, electrostatic interactions, and blending polymers with strong intermolecular interactions. The effect of hydrogen bonding on thermal conductivity in polymer blends has been a subject of debate, with some studies reporting significant enhancement while others find no significant effect. Existing experimental and computational studies on polymer blends have yielded mixed results, with some suggesting non-monotonic relationships between blend composition and thermal conductivity. The application of machine learning to polymer informatics has primarily focused on single-component polymers, largely neglecting the vast and unexplored chemical space of polymer blends. This work addresses this gap by applying machine learning and active learning to the exploration of polymer blend thermal conductivity.

This study employed a high-throughput molecular dynamics (MD) simulation combined with active learning (AL) to identify polymer blends with enhanced thermal conductivity (TC). Five datasets were used: Dataset 1 contained TC values for 608 amorphous single-component polymers; Dataset 2 was a stratified subset of Dataset 1; Dataset 3 contained TC values for 216 binary polymer blends with varying ratios (1:5, 1:1, 5:1); Dataset 4 combined Datasets 1 and 3; and Dataset 5 contained approximately 550,000 unlabeled polymer blends. Two polymer representation methods were used: Morgan fingerprints (MF) and polymer embeddings (PE). Two blend representation methods were investigated: weighted summation and weighted concatenation, leading to four representation combinations. A random forest (RF) regression model trained on Dataset 4 identified PE with weighted summation (PE-WS) as the optimal representation for polymer blends. An active learning framework, combining an RF classifier with certainty-based acquisition function, was used to iteratively select polymer blends from Dataset 5 for MD simulation and labeling. The process continued until a satisfactory proportion of high-performance blends (TC higher than both constituents) was reached (approximately 5%). The radius of gyration (Rg) and hydrogen bonding were analyzed to understand the molecular origins of TC enhancement. A log-linear model was used to analyze the relationship between TC improvement, Rg improvement, and hydrogen bond improvement. Odds ratios were calculated to quantify the association between these variables.

The active learning framework significantly accelerated the identification of high-performance polymer blends. The PE-WS representation method showed superior performance in capturing the structure-property relationships compared to other methods. The iterative AL process successfully increased the proportion of high-performance blends from 2.79% to 5.04%. Certainty-based sampling consistently outperformed random sampling in identifying high-performance blends. Analysis of the top 10 predicted blends revealed a strong positive correlation between increased Rg and enhanced TC in eight out of ten cases. A similar trend was observed for hydrogen bonding in seven out of ten cases. Log-linear model analysis of a broader dataset (387 blends) indicated a significant positive association between TC improvement and Rg improvement (conditional odds ratio = 2.87 when H-bond improvement was present). Hydrogen bond improvement did not show a direct association with TC but might contribute indirectly by affecting Rg.

The findings demonstrate the efficacy of combining high-throughput MD simulations with active learning for efficiently discovering high-performance polymer blends. The strong correlation between increased Rg and enhanced TC highlights the importance of extended polymer chain structures in improving thermal transport through stronger bonded interactions. The role of hydrogen bonding in enhancing TC appears to be indirect, potentially by influencing the Rg and chain conformation. The study's success in identifying high-performance blends highlights the potential of this approach for materials discovery and optimization. The use of polymer embeddings and weighted summation for blend representation demonstrates a significant advancement in polymer informatics.

This work successfully demonstrated the accelerated discovery of high-thermal conductivity polymer blends using active learning and high-throughput MD simulations. The strong correlation between increased Rg and improved TC was identified, and the indirect role of hydrogen bonding was suggested. Future research could explore other factors influencing thermal transport in polymer blends, such as miscibility, crosslinking, and interfacial thermal resistance. Extending this methodology to other material properties and blend compositions is a promising avenue for future work.

The study's findings are based on MD simulations and might not perfectly reflect real-world behavior. The specific force field used (GAFF2) and the choice of parameters in the MD simulations could influence the results. The analysis focuses primarily on the relative change in TC compared to the constituents and does not incorporate aspects like miscibility or cross-linking. The analysis of the relationship between Rg, hydrogen bonding, and TC relies heavily on the data obtained through the active learning process, and further studies could help confirm the findings using independent datasets.