FEM-GAN: A Physics-Supervised Deep Learning Generative Model for Elastic Porous Materials.

Area of Science:

• Materials Science
• Computational Imaging
• Artificial Intelligence

Background:

• X-ray micro-computed tomography (μCT) is crucial for material characterization but is costly and time-consuming.
• Existing generative models often lack physical grounding, limiting their application in scientific imaging.
• There is a need for physically informed generative models to enhance X-ray μCT image accessibility.

Purpose of the Study:

• To develop and evaluate a physics-supervised generative adversarial network (GAN) for generating X-ray μCT images.
• To integrate finite element method (FEM) simulations for physical guidance in the GAN training process.
• To assess the quality and utility of generated X-ray μCT images for various applications.

Main Methods:

• A physics-supervised generative adversarial network (GAN) was developed, incorporating elastic coefficients from FEM simulations.
• Hostun sand X-ray μCT images were used as target data.
• Nonparametric statistics were employed for posterior metric evaluation during training.
• A parametric study investigated various loss functions and FEM evaluation frequencies.

Main Results:

• The FEM-GANs model generated X-ray μCT images that outperformed reference images for most elasticity coefficients.
• While not perfectly reproducing all coefficients, the model demonstrated significant improvement over standard GANs.
• The generated images showed potential for improving data augmentation and image analysis tasks.

Conclusions:

• Physics-supervised GANs offer a viable approach to generating realistic X-ray μCT images.
• This method can mitigate limitations associated with traditional X-ray μCT imaging, such as cost and time.
• The generated images have practical applications in data augmentation, tool calibration, and multiscale modeling.