Generating Input Data for Microstructure Modelling: A Deep Learning Approach Using Generative Adversarial Networks

Area of Science:

• Materials Science
• Computational Materials Science
• Metallurgy

Background:

• Representative volume elements (RVEs) are crucial for simulating material behavior.
• Traditional statistical methods often use simplified distributions (e.g., log-normal, gamma) for microstructural parameters.
• These methods fail to capture complex interdependencies between parameters like grain size, shape, and orientation.

Purpose of the Study:

• To develop a more accurate statistical description of metallic microstructures.
• To account for the interdependencies between microstructural parameters.
• To generate realistic synthetic microstructural data for RVEs.

Main Methods:

• Implementation of a Wasserstein generative adversarial network (WGAN).
• Statistical analysis of microstructural parameters (e.g., area, aspect ratio, grain axis slope).
• Validation of generated data against input microstructure data.

Main Results:

• The WGAN successfully captured the distribution of microstructural parameters.
• Crucially, the WGAN accurately modeled the interdependencies between these parameters.
• Validation confirmed a strong match between the input and synthetically generated microstructure data.

Conclusions:

• Machine learning, specifically WGANs, offers a powerful approach for statistically describing complex metallic microstructures.
• This method overcomes limitations of traditional distribution functions by accounting for parameter interdependencies.
• The generated microstructural data is suitable for creating accurate RVEs in materials simulations.