Label Efficient Segmentation of SiC–C–Si Phases in SiC-based Ceramics from Optical Micrographs Using a Compact U-Net

Silicon carbide is a material built for endurance. It is a ceramic so hard it can cut through steel, yet light enough to be used in aerospace, and it can withstand temperatures that would melt most metals. Because of these qualities, engineers use it to make parts for engines, brakes, and high-temperature furnaces. However, the material is not a single, uniform substance. Inside a block of silicon carbide ceramic, there is a complex mixture of three distinct components: the main silicon carbide crystals, leftover free silicon, and pockets of carbon. The performance of the final part depends entirely on how these three ingredients are arranged and how much of each is present. If the silicon content is too high, the part might become brittle; if the carbon is unevenly distributed, it might fail under stress. To make better materials, scientists need to know exactly what is inside, but seeing these tiny, mixed-up phases under a microscope is surprisingly difficult.

For decades, researchers have tried to count these invisible ingredients by looking at reflected-light photographs of the material's surface. The challenge is that the three components look very similar to the human eye and to standard computer programs. They are all shades of gray, and the lighting can vary from one photo to the next, making it hard to tell where one phase ends and another begins. Traditional methods often rely on simple rules, such as "anything darker than this shade is carbon," but these rules fail when the contrast is low or when the different materials are woven together in thin, intricate threads. When these old methods fail, scientists must manually trace every single grain by hand, a process that is slow, tedious, and prone to human error. This bottleneck has slowed down the development of better ceramics, leaving engineers guessing about the exact composition of their materials.

A team of researchers from the Osipyan Institute of Solid State Physics and Irkutsk State University has developed a new way to solve this problem. Instead of relying on rigid rules or spending hours drawing lines by hand, they created a smart computer system that learns to recognize the three phases on its own. The key to their success was a technique called "weak supervision." Normally, training a computer to see these details requires a human to painstakingly label thousands of images, marking every single pixel as silicon, carbon, or ceramic. The researchers realized they could skip this exhausting step. They first used a mathematical method to group the colors in the photos into three rough categories. These groups were not perfect, but they were good enough to serve as a starting point. They then used these rough drafts to teach a compact neural network, a type of artificial intelligence inspired by the human brain, how to distinguish the phases accurately.

The system they built is based on a design known as a U-Net, which is shaped like the letter U. It works by first looking at the whole image to understand the general context, then zooming in to see the fine details, and finally combining those views to draw a precise map of the material's interior. The researchers trained this system on hundreds of micrographs of silicon carbide ceramics made through a process called liquid silicon infiltration. In this process, molten silicon is poured into a porous carbon block, reacting to form silicon carbide while leaving some silicon and carbon behind. The team tested their new system against both the old rule-based methods and the slow manual tracing done by human experts. They found that the old rule-based methods often collapsed completely when the contrast was low, failing to separate the silicon from the ceramic and producing wildly inaccurate counts. The new AI system, however, maintained a high level of accuracy even in the most difficult, low-contrast images.

While the new system did not always produce a slightly more perfect pixel-by-pixel map than a highly skilled human using a traditional watershed method, it offered a massive advantage in speed and consistency. The traditional method, which requires a human to adjust settings and guide the computer, can take tens of minutes or even an hour to process a single image. In contrast, the new AI system can analyze a full, high-resolution micrograph in less than a second on a standard graphics card, or in just a few seconds on a regular computer processor. More importantly, the system is reproducible. If you run the same image through the system ten times, it gives you the same result every time, whereas a human might draw slightly different lines each time due to fatigue or distraction. The researchers found that the system could estimate the volume of silicon in the material with an error of less than two percent for most samples, a level of precision that is sufficient for routine industrial quality control.

The study also revealed why some images are harder to analyze than others. The researchers developed a way to measure the "difficulty" of a micrograph by looking at how thin the connections between the phases are and how much the colors overlap. They found that when the silicon forms very thin bridges connecting larger pools, or when the colors of the silicon and ceramic are nearly identical, simple thresholding methods fail. The new AI system, however, was able to navigate these tricky areas by learning from the context of the surrounding image, effectively seeing the whole picture rather than just isolated pixels. This ability to handle complex, messy microstructures means that scientists can now rapidly screen new ceramic recipes, testing how different manufacturing conditions affect the final material without being slowed down by manual analysis.

This work represents a significant step forward in making advanced ceramics more accessible and reliable. By automating the counting of microscopic ingredients, the researchers have removed a major barrier to optimizing these materials. The system is not a magic solution that works perfectly in every single case; the researchers noted that for the very coarsest structures with the largest features, the error rate was slightly higher than with traditional methods. However, the trade-off is overwhelmingly positive, offering a fast, automatic, and consistent way to understand the inner life of silicon carbide ceramics. As the team looks to the future, they plan to refine the system further and apply it to even more complex materials, potentially helping to design the next generation of heat-resistant components for industry and space. The ability to see clearly into the microscopic world of ceramics is no longer a slow, manual struggle, but a rapid, automated process that can keep pace with the demands of modern engineering.