An article published in the journal “npj Computational Materials” reports the results of the use of a machine learning algorithm in the search for new materials. A team of researchers used an algorithm specialized in molecular design trained to recognize candidates with certain properties. In this way they designed polymers with high thermal conductivity, very useful to use for example in mobile technologies connected to 5G.
The use of machine learning algorithms is relatively new and in some fields its usefulness is yet to be verified. Today’s supercommuters can analyze huge amounts of data in a short time but specific training is needed for them to learn all they need to perform the type of analysis required. In the case of the design of new materials, a system must be trained so that it recognizes certain characteristics of the materials it analyzes.
To address this type of analysis, the researchers used a machine learning technique called transfer learning, in which the knowledge accumulated during training in a type of problem is adapted and applied to another different but related problem. Initial data on polymer properties were taken from PoLyInfo, the largest polymer database in the world hosted at the National Institute for Materials Science (NIMS), in Japan. Data on thermal transfer properties are limited so a preliminary training phase was needed using existing data then adapted for the training to the new targets.
The resulting model was combined with a machine learning algorithm for the molecular project called iQSPR, available as free / open source software under the MIT license to evaluate a series of candidates and select the ones considered to be best with respect to thermal conductivity. The code to produce the key results written in R language is available on GitHub.
Three polymers were selected by the candidates based on their ease in synthesis and processing. Tests on these products confirmed the expected high thermal conductivity, up to 0.41 Watts per meter-Kelvin (W/mK), a value that is 80% higher than typical polyimides, a type of polymers that have various uses and have been produced since 1955.
The image (Courtesy npj Computational Materials, Ryo Yoshida, Junko Morikawa et al. Creative Commons Attribution 4.0 International License) shows the chemical structure of the three synthesized polymers (a) and the routes to the targets (b-e).
Ryo Yoshida, one of the authors of the research, stated that there are still aspects to be explored, for example regarding the training of computer systems to work with limited data. Polymers have different properties from metals and ceramics and are not yet fully predicted by existing theories. The project that he’s pursuing together with his colleagues aims not only to develop materials informatics but also to contribute to a fundamental progress in materials science.
The use of machine learning for this type of research showed promising results and the applications are still to be discovered. Different applications require different characteristics so we can still expect many developments.
