https://www.cas.cn/cm/202512/t20251212_5092190.shtml
Making ceramics is like cooking. From raw materials to finished products, it involves dozens of processes such as blending, ball milling, pressing, and sintering. The proportions of various raw materials, the precision of ball milling, and the sintering temperature all affect the final product’s properties. In the past, even with the involvement of the “national team” in developing key ceramic materials, each new generation of innovation required approximately 15 years.
At the CAS Shanghai Institute of Ceramics, an AI-based intelligent system found the optimal raw material ratio and process that would have required 10,000 trials in just 40 automated experiments. The newly created material exhibited extremely high stability, with its performance remaining unchanged even after more than 1,000 hours of testing.
“The window of opportunity for ‘AI + Manufacturing’ is within these three to five years,” researchers at the Shanghai Institute of Ceramics comment. The institute’s advantage lies in its long-term accumulation of over 200,000 data points in materials science, 10 million literature reviews, and 1.5 million patent data points. Leveraging the “flywheel effect” of large-scale models, it has gained a head start in the AI era. Currently, there are already mature robotic workstations on the market that can perform a series of tasks such as powder weighing, precise dispensing, and liquid stirring. Why does the Shanghai Institute of Ceramics need to build its own AI laboratory? The answer lies in two words—solid. Ceramics are made from solid powder, and solid-state experiments are more than ten times more difficult than liquid experiments. Take uniform mixing, for example: liquids only require vibration and stirring, while ball milling of solids requires adding grinding beads, making separation and material adhesion extremely difficult. Those not truly engaged in this research simply cannot build the equipment suitable for solid-state experiments.
Behind the AI agent stands the Institute of Ceramics’ large-scale materials model, MatMind, which is trained based on over 200,000 data points of materials science, 10 million literature documents, and 1.5 million patent data points accumulated over many years by the institute. Previously, developing new materials was like finding a needle in a haystack; even thousands of experiments might not yield the global optimum. But now, after each step, the intelligent agent provides suggestions, which, combined with human experience, greatly improves development efficiency.