Abstract
In recent years, materials infomatics (MI), a technology that combines materials engineering and machine learning, has become popular and is used for discovering new materials. In this research, we aimed to verify whether MI can be applied to the problem of “development and maintenance of technology,” which is becoming more difficult due to the decrease in the number of engineers caused by the declining birthrate and aging population in Japan. We selected “discrimination of optical electron microscope images” as the verification target, and used Convolutional Neural Networks (CNNs) as the machine learning technology to discriminate between seven types of sintered metal objects under different sintering conditions, hoping for general applicability to the discrimination problem, and confirmed a discrimination accuracy of 98.5%. In addition, we verified the effectiveness of using pseudo-samples for the discrimination problem using Generative Adversarial Networks (GANs) in the hope of improving accuracy by increasing the number of samples, and confirmed the improvement of accuracy by adding pseudo-samples to the training data.
