AI Platform
Achieving 90% Defect Detection Accuracy with AI Vision and MLOps for Autonomous Manufacturing | Shinhan Diamond Industrial
■ The Challenge of Defect-Free Semiconductor Components: Manual Inspection and Complex Process Variables
Chemical mechanical polishing (CMP) creates a mirror-smooth surface on semiconductor wafers. CMP diamond disks play an essential role by conditioning the polishing pad. Each disk contains 50,000 microscopic diamond particles—and if even one becomes dislodged or protrudes above the intended height, it can cause severe wafer scratches and directly reduce semiconductor yield.
Despite these demanding quality requirements, inspecting micron-scale particles relied on operators’ eyesight and sense of touch. Fatigue led to inconsistent inspection results, while growing labor shortages constrained productivity. The company also faced a structural limitation: complex, constantly changing electroplating variables could not be captured in real time, leaving quality control dependent on inspection after processing.
■ A Unified MonPlant Infrastructure Connecting 3D Vision and Physics-Based Algorithms
Shinhan Diamond Industrial adopted MonPlant, Mondrian AI’s enterprise AI platform, to build an autonomous manufacturing workflow that continuously converts process conditions into precise data and supports automated decision-making.
Using MonPlant, Mondrian AI designed a robust data pipeline to support model training on more than 100,000 high-quality data samples manually collected by shop floor operators.
The platform brought previously disconnected equipment into a unified intelligence framework—from ultra-precision AI vision sensors that detect micron-scale defects and 3D vision systems that guide diamond particle placement to electrochemical measurement equipment that monitors additive concentrations.
For flatness control, Mondrian AI also established computing resource allocation at the edge and visual monitoring of system status. This infrastructure supported the application of a physics-informed neural network algorithm that combines data-driven learning with physical laws.
■ Turning Shop Floor Expertise into Digital Assets with MonPlant
Following deployment, decades of tacit knowledge held by experienced operators became digital assets that enabled the system to control process variables autonomously. Sensors and algorithms interacted in real time at the edge, creating an intelligent feedback loop that corrected deviations and delivered measurable manufacturing improvements.
- AI training dataset: 100,000 samples
- Defect detection accuracy: Increased from 60% to 90%
- Product inspection time: Reduced from 22 minutes 36 seconds to 13 minutes 20 seconds
- Hourly production output: Increased from 138 to 143 units
- Defect rate: Reduced by 10%, from 6% to 5.4%
These results were supported by preventive, process-specific control that helped avoid production downtime. Shinhan Diamond Industrial’s experience demonstrates how MonPlant connects manufacturing data, analysis, and control in a continuous workflow—strengthening the competitiveness of Korea’s semiconductor tooling supply chain.
