Material structure-property linkages using three-dimensional convolutional neural networks. (March 2018)

• Record Type: Journal Article
• Authors: Cecen, Ahmet; Dai, Hanjun; Yabansu, Yuksel C.; Kalidindi, Surya R.; Song, Le
• Is Part Of: Acta materialia. Volume 146(2018)
• Page Start: 76
• Page End: 84
• Publication Date: 2018-03
• Subjects: Convolutional neural networks -- Spatial correlations -- Structure-property linkages -- Principal component analysis; Materials -- Periodicals; Materials science -- Periodicals; Materials -- Mechanical properties -- Periodicals; Metallurgy -- Periodicals; Chemistry, Inorganic -- Periodicals; 620.112
• DOI: 10.1016/j.actamat.2017.11.053
• Languages: English
• ISSNs: 1359-6454

Abstract

Abstract: The core materials knowledge needed in the accelerated design, development, and deployment of new and improved materials is most accessible when cast in the form of computationally low cost (reduced-order) and reliable process-structure-property (PSP) linkages. Quantification of the material structure (also referred as microstructure) is the core challenge in this task. Conventionally, microstructure quantification has been addressed using highly simplified measures suggested by the governing physics, with the list of measures often suitably augmented by the intuition of the materials expert. In this paper, we develop an objective (data-driven) approach to efficiently and accurately link a three-dimensional (3-D) microstructure to its effective (homogenized) properties. Our method employs a 3-D convolutional neural network (CNN) to learn the salient features of the material microstructures that lead to good predictive performance for the effective property of interest. We then utilize 3-D CNN learned features as estimators of higher-order spatial correlations, and formulate an integrated framework combining 3-D CNN features with 2-point spatial correlations. In this work, we created an extremely large microstructure-property benchmark dataset of 5900 microstructures, and demonstrated that our CNN based approach not only learns interpretable microstructure features, but also leads to improved accuracy in property predictions for new microstructures, while achieving a dramatic reduction in the computation time. Graphical abstract: Image 1