AI accelerates the development of advanced heat-penetrating polymers

Machine learning methods developed by researchers at the Tokyo Institute of Science, Statistical Mathematics and other institutions accurately predict polymer liquid crystals with 96% accuracy. They screened over 115,000 polyimides and selected six candidates who were likely to exhibit liquid crystals. Successful synthesis and experimental analysis have led to these liquid crystal polyimides demonstrating thermal conductivity up to 1.26 WM⁻¹k⁻¹, facilitating the discovery of efficient thermal materials in next-generation electronics.

Finding new polymeric materials that can efficiently dissipate heat while maintaining high reliability is one of the biggest challenges of modern electronics. One promising solution is liquid crystal polyimides, a special class of polymers in which molecules naturally align with highly ordered structures.

These ordered chains create heat flow paths, making liquid crystal polyimides extremely attractive for thermal management of semiconductors, flexible displays, and next-generation devices. However, the design of these polymers has long been reliant on trial and error, as they lacked clear design rules to predict whether the polymer would form a liquid crystal phase.

In the breakthrough, researchers at the Tokyo Institute of Science (Science Tokyo), the Institute of Statistical Mathematics (ISM), and other institutions have developed machine learning models that can successfully predict the polyimide structure that forms the global liquid crystal phase of polyimma material research. The team then confirmed that they successfully synthesized these polymers, forming a smectic liquid crystal phase, demonstrating significantly higher in-plane thermal conductivity than traditional polyimides.

This study shows significant advances in data-driven polymer design, published online in the journal NPJ Calculation Materials July 2, 2025. Dr. Morikawa Nishikai (Tokyo Science) was led as the main investigator, and Yakawakawa (Tokyo Science) and Yoshida Ryoda (ISM) were led as the main collaborators. This project was mainly promoted by Associate Professor Stephen Wu (ISM) and graduate students Kuda and Nakagawa Fumi (Science Tokyo).

“This study shows that we take an important step in using machine learning to develop polymers and that we can successfully identify liquid crystal polymers with high thermal conductivity,” explains Hua.

The model developed by Yoshida's group serves as a binary classifier. Given the chemical structure of the polymer, it predicts whether the polymer forms an ordered liquid crystal state, achieving an impressive 96% classification accuracy.

To train the model, researchers used PolyInfo, the National Institute of Materials Science's comprehensive polymer properties database, containing 951 polymers confirmed to form 3,597 unlabeled polymers with liquid crystal phases. From these data, the model learned to recognize the chemical, physical, and structural properties that allow polymers to self-assemble into ordered phases.

The team then used the model to screen the desired liquid crystal properties of polyimide, a family of high-performance plastics known to withstand high temperatures and be used as insulating materials in electronic and aerospace applications.

To generate realistic candidates, researchers created a virtual library by splitting the polyimide template designed by Hayashi's group into five basic building blocks. The combination of various pairs of acid dianhydrides and diamines (molecular fragments that make up the polyimide chain) produced over 115,000 possible structures.

The model screened this virtual library and predicted more than 10,800 polyimide candidates that are likely to form a liquid crystal phase. From these, the researchers synthesized six diverse examples to form a smectic liquid crystal phase.

The thermal conductivity of these polyimides measured in the Morikawa group ranged from 0.72 to 1.26 Wm⁻¹k⁻¹. Researchers found that polyimides with stiffer molecular structures and better in-plane arrangements exhibit higher thermal conductivity and provide insight into the preferred molecular structure.

“This achievement is the first in polymer materials research where liquid crystal polymers have been discovered using machine learning. Our method opens up ways to investigate not only liquid crystal polyimides but other classes of liquid crystal polymers,” says Yoshida.

This method is a new trend to use machine learning for material design and convert it into rapid, data-driven exploration. It points to a future where materials with customized properties on a computer are designed and validated in just a few cases.