Generative artificial intelligence (genAI) has revolutionized multiple disciplines by drastically changing content creation approaches. This study focuses on implementing genAI to facilitate efficient investigations of isothermally aged interfacial Cu-SAC305 microstructures. A C-DDPM is developed, conditioned by ageing time and ECD-Cu metallisation growth rates correlating with Cu-impurity contents. The model generates high-quality microstructural images that are qualitatively indistinguishable from real FESEM-BSE micrographs. Quantitative analysis shows that the virtually generated data generally lies within the standard deviations of the real data. The C-DDPM’s ability to generate physically accurate microstructural images is assessed by evaluating physical descriptors like IMC-layer thicknesses and Kirkendall pore areas. The results demonstrate that genAI has the potential to reduce experimental work for microstructural imaging significantly by training on a relatively small dataset. The C-DDPM is proven to be a promising tool for generating microstructural images of unknown conditions when only relatively few real micrographs are available. The findings have great potential to advance failure analysis and material development by enabling holistic microstructural imaging of critical conditions with high efficiency.
Introduction
GenAI has disrupted multiple disciplines, from daily life to education, medicine, scientific research, and art, by changing content creation approaches. In materials science, genAI can be used to generate microstructural images, reducing experimental work and time.
Methodology
A C-DDPM is developed and trained on FESEM-BSE micrographs to generate microstructural images of Cu-SAC305 interfaces. The model is conditioned by ageing time and ECD-Cu metallisation growth rates. The generated images are validated using physical descriptors like IMC-layer thicknesses and Kirkendall pore areas.
Results
The C-DDPM generates high-quality microstructural images that are qualitatively indistinguishable from real FESEM-BSE micrographs. Quantitative analysis shows that the virtually generated data lies within the standard deviations of the real data. The C-DDPM’s ability to generate physically accurate microstructural images is demonstrated.
Conclusion
This study highlights the application of genAI for microstructural image generation in materials science. The developed C-DDPM exhibits negligible hallucinations and can extract significant cause variables from a relatively scarce number of seen conditions. The findings have great potential to advance failure analysis and material development.
