Semi-supervised deep transfer learning for the microstructure recognition in the high-throughput characterization of nickel-based superalloys

Abstract

Nickel-based superalloys, owing to their superior resistance against mechanical and chemical degradation, have been widely applied in the aerospace, turbine engine, nuclear reactor, and chemical industries. The microstructure recognition plays a key role in the characterization and design of new superalloys. Although deep learning techniques have achieved satisfactory performance in the microstructure recognition, these methods usually suffer from generalization when the alloy composition or process changed, especially in the high-throughput experiments. In this paper, we propose a semi-supervised deep transfer learning framework for the microstructure recognition of nickel-based superalloys with different compositions and heat treatment procedures. To be specific, we achieve the knowledge transfer of recognition models from one condition (source domain) to another (target domain) by feature distribution alignment (FDA). To avoid the over-fitting, we design a dynamic alignment strategy to achieve the feature alignment based on the label guidance. Additionally, we effectively utilize the unlabeled samples in the target domain and achieve the distribution alignment between the two domains by adversarial training. The experimental results show that our method is superior to the commonly used deep transfer learning methods. In spite of few labeled samples, it can also approach the satisfactory accuracy. Codes are available at: https://github.com/yangchuanwu/FDA.