The National Institute for Materials Science (NIMS) and its partners have created a model that forecasts the long-term strength of various heat-resistant steel types. They accomplished this by using machine learning techniques while ensuring that each organization’s data remained confidential. This research is published in Tetsu-to-Hagané.

Data from private companies is often sensitive, which makes it difficult to share information between organizations for joint research and development. Gathering such data can also be lengthy and pricey—sometimes it takes over ten years to obtain lifetime information on heat-resistant materials used in power generation settings. This underscores the importance of partnerships between the industry and public sectors.

NIMS has developed a system that allows multiple organizations—including six private companies and two national research institutes—to independently conduct machine learning with their own local data while maintaining confidentiality (known as federated learning).

This collaborative effort led to the establishment of a “global model” that can accurately predict the long-term durability of heat-resistant steel materials. The accuracy of this global model was significantly higher than that of any local model built solely with data from NIMS. This marks the first successful collaboration between industry and the public sector using federated learning.

The success of this project is anticipated to encourage further collaboration across various materials fields. The federated learning system produced by NIMS is now available for public use. Moving forward, NIMS aims to serve as a facilitator, promoting partnerships that meet the increasing demand for collaboration between industry and the public sector.

The federated learning system utilized in this research has been developed and made available as open source by NIMS and Elix.