Generation of synthetic microstructures containing casting defects: a machine learning approach.

Area of Science:

• Materials Science
• Computational Materials Science
• Additive Manufacturing

Background:

• Casting defects like shrinkages and pores significantly impact material properties, particularly metal fatigue.
• Understanding defect morphology and spatial distribution is crucial for material development.
• Existing methods for defect simulation lack the realism needed for accurate mechanical property prediction.

Purpose of the Study:

• To develop a novel strategy for generating realistic synthetic samples with casting defects.
• To characterize and model the spatial distribution and morphology of shrinkages and pores in Inconel 100.
• To validate the generated synthetic samples against real material data.

Main Methods:

• Utilized X-ray tomography for characterizing four Inconel 100 reference samples with casting defects.
• Integrated Spatial Point Pattern (SPP) analysis to describe defect spatial distributions.
• Employed deep learning techniques, including Generative Adversarial Networks (GANs) and Convolutional Neural Networks (CNNs), for defect generation.

Main Results:

• SPP analysis identified two distinct void nucleation mechanisms during metal solidification.
• The deep learning model successfully generated synthetic casting defects (100 µm to 1.5 mm) with realistic shapes.
• Generated samples maintained global defect statistics consistent with reference samples.

Conclusions:

• The developed deep learning approach effectively generates synthetic casting defects that closely mimic real ones.
• This method provides a powerful tool for creating realistic microstructures for material simulation and testing.
• The findings contribute to a better understanding of void nucleation during metal solidification.