Deep-learning approach to predict a severe plastic anisotropy of caliber-rolled Mg alloy

• Published in: Materials letters 2020-06, Vol.269, p.127652, Article 127652
• Main Authors: Lee, Taekyung, Kwak, Byung Je, Yu, Jinyeong, Lee, Jeong Hun, Noh, Yoojeong, Moon, Young Hoon
• Format: Article
• Language: English
• Subjects: Alloys; Anisotropy; Artificial neural networks; Crystal structure; Deep learning; Deformation and fracture; Magnesium; Magnesium base alloys; Materials science; Metals and alloys; Neural networks; Plastic anisotropy; Regression analysis; Texture
• ISSN: 0167-577X; 1873-4979
• DOI: 10.1016/j.matlet.2020.127652

[Display omitted] •Deep neural network (DNN) with 5 hidden layers was trained by 85,967 examples.•DNN gave a good prediction for the severe plastic anisotropy of caliber-rolled Mg alloys.•DNN exhibited a high generalization ability as well as high applicability. Mg alloys have a strong plastic anisotropy due to their intrinsic crystal structure. In particular, an anisotropy of caliber-rolled Mg alloy is difficult to anticipate using a traditional method due to the unique texture developed by this process. This study adopted a deep neural network (DNN) with optimized hyperparameters to predict the severe plastic anisotropy. The DNN model was trained with 85,967 examples, and then evaluated in comparison with other approaches, such as ‘shallow’ neural networks, multiple linear regression, and constitutive analytical equations. The optimized DNN model exhibited the best prediction among these approaches. Furthermore, it showed a high generalization ability, which is indispensable for interpreting a plastic anisotropy. It has been verified that the deep-learning approach has a vast potential for interpreting the anisotropy problem of Mg alloys.