Building a machine learning model while preserving data confidentiality
Six companies and two national R&D institutes collaborate to predict the long-term durability of diverse heat-resistant materials
image: Distributed learning conducted by each organization enabled the integration of model parameters without compromising data confidentiality, leading to improved accuracy in the lifetime prediction of heat-resistant materials.
Credit: Masahiko Demura, National Institute for Materials Science
NIMS and its collaborators (IHI Corporation, Kawasaki Heavy Industries, Ltd., Kansai Electric Power Co., Inc., Kobe Steel, Ltd., Electric Power Development Co., Ltd., the Japan Atomic Energy Agency, Mitsubishi Heavy Industries, Ltd. and Elix, Inc.) have developed a model designed to predict the long-term durability of a range of heat-resistant steel materials by performing machine learning while preserving the confidentiality of each organization’s data. This research was published online in Tetsu-to-Hagané on February 6, 2025.
Background
Private companies’ proprietary materials data is highly confidential, making cross-organizational sharing of it for collaborative R&D challenging. However, generating such data is extremely time-consuming and costly, so cross-organizational data collaboration is desirable. In particular, it can take more than a decade to acquire lifetime data for heat-resistant materials used in power generation facilities, highlighting the need for industry–public sector collaboration.
Key Findings
NIMS developed a system enabling multiple organizations (six private companies and two national R&D institutes) to independently perform machine learning using their own local data while preserving its confidentiality (i.e., through federated learning). As a result, they jointly constructed a “global model” capable of predicting the long-term durability of heat-resistant steel materials (see figure). The global model demonstrated significantly higher predictive accuracy than a local model built solely using NIMS’ data. This represents the first example of industry–public sector data collaboration through federated learning.
Future Outlook
These achievements are expected to promote industry–public sector data collaboration across a broad range of materials research fields. The federated learning system developed by NIMS is publicly available and open source. Going forward, NIMS plans to act as a coordinator, fostering collaboration to meet growing demands for industry–public sector partnerships.
Other Information
- The federated learning system used in this study was developed and released as open source (https://github.com/nims-federated-learning/NIMS-FL) by NIMS and Elix, with funding from the second-term SIP (Cross-ministerial Strategic Innovation Promotion Program) project entitled “Materials integration for revolutionary design system of structural materials.” This project was carried out under the NIMS Structural Materials DX-MOP framework with support from the MEXT DxMT project (grant number: JPMXP1122684766).
- This research was published in the online version of Tetsu-to-Hagané on February 6, 2025.