Federated Learning of Creep Rupture Time and High Temperature Tensile Strength Prediction Models

Abstract

Creep testing is time-consuming and costly, leading institutions to limit the number of tests conducted to the minimum necessary for their specific objectives. By pooling data from each institution, it is anticipated that predictive models can be developed for a wide range of materials, including welded joints and degraded materials exposed to service conditions. However, the data obtained by each institution is often highly confidential, making it challenging to share with others. Federated learning, a type of privacy-preserving computation technology, allows for learning while keeping data confidential. Utilizing this approach, it is possible to develop creep life prediction models by leveraging data from various institutions. In this paper, we constructed global deep neural network models for predicting the creep rupture life of heat-resistant ferritic steels in collaboration with eight institutions using the federated learning system we developed for this purpose. Each institution built a local model using only its own data for comparison. While these local models demonstrated good predictive accuracy for their respective datasets, their predictive performance declined when applied to data from other institutions. In contrast, the global model constructed using federated learning showed reasonably good predictive performance across all institutions. The distance between each institution's data was defined in the space of explanatory variables, with the NIMS data, which had the largest dataset, serving as the reference point. The global model maintained high predictive accuracy regardless of the distance from the NIMS data, whereas the predictive accuracy of the NIMS local model significantly decreased as the distance increased.