Skip to main content
Information contained in this news release is current as of the date of the press announcement, but may be subject to change without prior notice.
June 28, 2021
Realizing DX in new materials development; Established AI technology
to reduce the number of experiments for organic material development by about 1/4
Tokyo, June 28, 2021 --- Mitsui Chemicals, Inc. (TSE: 4183, Mitsui Chemicals) and Hitachi, Ltd. (TSE: 6501, Hitachi) today announced that they have begun a practical demonstration test of materials informatics (MI) technology developed by Hitachi that utilizes artificial intelligence (AI) for actual development of new materials. Prior to the demonstration test, Hitachi's technology has been verified by using past organic material development records provided by Mitsui Chemicals, and it was confirmed that the number of experimental trials required to develop new high-performance compounds can be reduced to about 1/4 compared with that for conventional MI, shortening the development period.
This technology will be presented at the conference "DICOMO2021" (1) to be held on June 30, 2021.
In the future, Hitachi aims to establish this MI technology as a practical method for the "Materials Development Solution," (2) which is one of Hitachi's "Lumada" solutions with advanced technologies, to reduce time and costs for customers' and partners' development processes, and to strengthen their competitiveness. Mitsui Chemicals will continuously create innovative products, services, and business models in an agile manner to solve social issues through DX. Both companies will continue to promote collaborative creation (3) on materials development and contribute to realization of a sustainable society.
Development of new products is one of the most crucial activities in Mitsui Chemicals' business. However, it reluctantly spends significant time and incurs significant costs on processes such as problem setting, basic research, and scale-up experiments. At this time, combining the vast knowledge of Mitsui Chemicals obtained through its past new materials development with digital technologies provided by Hitachi will lead to a drastic reduction in the time and cost needed for new materials development.
Hitachi developed a new deep-learning technology to suggest chemical formulas expected to have characteristics better than existing chemical compounds as an advanced MI, which is a technology to develop new materials with AI and simulation technologies. The new technology, even in an organic material development that requires a huge experimental dataset, enables chemical formula suggestion without such a dataset.
Features of the technology are as follows.
Deep-learning technology that can automatically generate chemical formulas
for high-performance materials even with small amounts of experimental data
For more information, please visit Mitsui Chemicals' website (https://jp.mitsuichemicals.com/en/index.htm).
For more information, please visit Hitachi's website (https://www.hitachi.com/).
In the field of materials science, MI with advanced technologies such as AI and simulation is expected to improve the efficiency of R&D to meet the diversifying demand for materials such as plastic materials with low environmental impact. For example, AI is applied to analyzing data such as material blend ratios and manufacturing conditions (temperature, pressure, etc.) to estimate the optimal blend amounts and manufacturing conditions for creating high-performance materials.
In a general case of MI applied to the development of organic materials, AI is required to handle chemical formulas in textual format (e.g., ethanol (CH3CH2OH)) , but this is not an easy task. Therefore, as a known approach, the structure and properties of compounds are expressed numerically as descriptors, (8) and AI can predict performance from such descriptors. However, these descriptors cannot be easily inversely transformed into chemical formulas. In the conventional method, experts manually identify the chemical formulas of new materials by selecting from a large number of chemical formulas based on the predicted performance values of AI, and then evaluate their actual performance through real experiments (see Figure (1)).
In recent years, with advances in deep learning, AIs that can directly generate and suggest chemical formulas have been developed. However, such AI requires a large amount of experimental data which contains chemical formulas, experimental conditions, and their performance-index values (9) to learn the relationship between chemical structures and performance. Due to this requirement, the number of experiments needed to obtain training data increases.
Please contact the Research & Development Group, Hitachi, Ltd.