Director of Ryo Yoshida, Institute of Statistical Mathematics, Japan Information and Systems Research Organization, Professor Sunko Junko and material genetics expert at Tokyo Institute of Technology, and Professor Xu Yibin of Japan's National Materials Research Institute collaborate to allow computers to learn existing data through artificial intelligence to find high thermal conductivity New molecular materials. The related research results were titled "Machine-learning-assisted discovery of polymers with high thermal conductivity using a molecular design algorithm" and were published in "Nature" top journal "npj Computational Materials". The team's machine learning approach involves "transfer learning", which is characterized by the ability to find materials with the required characteristics from extremely small databases.
Stones from other hills, can learn
PoLyInfo of Japan's National Materials Research Institute is the world's largest polymer database. The data of this research is derived from PoLyInfo's polymer performance database. Although the database is very large, the amount of data on the heat transfer properties of polymers is limited. In order to predict the heat transfer characteristics of polymers based on the given limited data, the machine learning model is pre-trained using properties related to thermal conductivity properties, and there is enough data in the database for related tasks. These pre-trained models capture features related to target attributes (here the heat transfer characteristics are studied). Reusing these machines on target tasks to obtain feature attributes, even with very small data, can produce excellent performance predictions. For example, this is like the ideal reasoning of human beings. Machine learning has also achieved this breakthrough in the field of molecular design of polymer thermally conductive materials.
iQSPR algorithm + machine learning algorithm
Yoshida and colleagues previously developed the iQSPR algorithm, which the research team combined with a specially crafted machine learning algorithm for molecular computing design. Using this technique, thousands of promising "virtual" polymers can be identified. Among a large number of candidate materials, three polymers were selected based on their ease of synthesis and processing. Test results confirm that the thermal conductivity of the new polyimide is as high as 0.41 W / mK. This result is 80% higher than conventional polyimide.
Multilateral cooperation
By verifying the computational heat transfer performance of the design polymer, the research achieved a key breakthrough in the use of machine learning methods to design materials, and a fast and economical machine learning method to design thermally conductive polymer materials. This research requires the team's collaboration in data science, organic synthesis, and advanced measurement technologies.
Yoshida commented that many aspects remain to be explored, such as how to "train" computing systems to handle limited data by adding more appropriate descriptors. "Using machine learning to design polymers or soft materials is a challenging but promising area because these materials have different properties from metals and ceramics, and existing theories have not yet fully predicted the performance,"
This research is the starting point for discovering other innovative materials, Morikawa added: "We want to create a machine learning driven high-throughput computing system to design the next generation of soft materials for applications in the 5G, 6G, 7G … nG era. Through us Project, our goal is not only to pursue the development of materials informatics, but also to help advance the basic science of materials, especially in the field of phonon engineering. "