Comparative study of the experimentally observed and GAN-generated 3D microstructures in dual-phase steels

Area of Science:

• Materials Science
• Computational Materials Science
• Artificial Intelligence

Background:

• Deep learning, specifically generative adversarial networks (GANs), shows promise for generating artificial microstructures.
• Reproducing three-dimensional (3D) microstructures from two-dimensional (2D) images is a key challenge for materials science applications.
• Assessing the quantitative accuracy and mechanical relevance of GAN-generated microstructures is crucial for practical use.

Purpose of the Study:

• To evaluate the reproducibility and accuracy of 3D microstructures generated by GANs from 2D images of dual-phase steels.
• To compare the mechanical behavior of GAN-generated microstructures with experimentally observed ones using finite element analysis.
• To identify the limitations of current GAN algorithms in capturing essential microstructural features.

Main Methods:

• Utilized an automated serial sectioning technique to obtain observed 3D microstructures.
• Generated artificial 3D microstructures using a GAN-based algorithm from three orthogonal 2D surface images.
• Constructed finite element models from 3D voxel data of both observed and generated microstructures.
• Performed finite element analysis (FEA) on representative volume elements (RVEs) to simulate mechanical behavior.

Main Results:

• GAN-generated microstructures captured macroscopic anisotropy but showed quantitative discrepancies in mechanical responses compared to observed microstructures.
• Inaccuracies in reproducing the volume fraction of ferrite/martensite phases were identified as a primary cause for the misalignment.
• The GAN algorithm struggled to replicate fine microscopic morphology, especially the connectivity of the martensite phase when its volume fraction was low.

Conclusions:

• Current GAN-based algorithms have limitations in accurately reproducing the quantitative features and morphology of complex microstructures like dual-phase steels.
• The inability to precisely replicate phase volume fractions and intricate morphologies limits the predictive capability of generated microstructures for mechanical behavior.
• Three-dimensional observation and validation remain essential for reliable materials modeling and design, underscoring the need for advancements in generative algorithms.