Project Summary
- Customer and industry: TSMC / semiconductor foundry (Arizona fab and Southern Taiwan Science Park advanced packaging fab).
- Equipment: EFEM ATM wafer robot, used in advanced-node processing and CoWoS advanced packaging.
- Solution: Level 4 online life-prediction solution using the Aurora® 313MM2-R1 smart multivariate sensor and the RULe.TI002 edge AI inference gateway.
- Communication: Sensor to gateway over wired Modbus RS485; gateway to the FDC system over Modbus TCP industrial Ethernet.
Ultra-thin sensors are mounted on key mechanisms of the robot, and the edge AI gateway computes the health index (HI) and remaining useful life (RUL) in real time, helping the customer pursue zero unplanned downtime and zero-defect processing.
Pain Points
The ATM robot picks and places 12-inch wafers between the FOUP at the load port and the EFEM chamber, and is the hub of continuous production. When long-term high-frequency operation causes bearing wear, excessive backlash or a loose belt:
- Wafer scrap: One FOUP holds 25 12-inch wafers. At 3nm / 5nm or CoWoS packaging nodes, a position error that scratches, drops or hits a wafer costs an estimated USD 250,000 – 1,000,000 in direct scrap per incident.
- Downtime and qualification: After wafer breakage, cleanroom clean-up, robot re-teaching and run-card qualification (QUAL) take 4 – 12 hours, and a bottleneck tool loses tens of thousands of dollars per hour.
The customer used to rely on manual listening rounds or scheduled preventive maintenance (PM). Listening by ear cannot catch early, small vibration changes. Scheduled PM may replace parts that are still good, which raises cost, while blindly stretching the PM interval risks unplanned downtime.
Solution
- Ultra-thin and small: The Aurora® 313MM2-R1 measures only 21 × 21 × 11.5 mm and mounts non-invasively on the flange or joints of the ATM robot without affecting its dynamics or trajectory accuracy.
- Layered standard communication: Sensor to gateway uses industrial Modbus RS485, which is robust and stable. The RULe.TI002 then sends features and diagnostics over standard Modbus TCP straight into the customer's FDC / SCADA network, with no complex software conversion.
- Edge AI: The sensor has a built-in Arm® Cortex®-M7 core. Together with the RULe.TI002 it outputs a Physical AI health index (HI) and remaining useful life (RUL), accurately identifying bearing, gear and joint wear modes and avoiding false and missed alarms.
- Much less data: Spectral features are extracted at the edge, reducing the data stream from 307,200 Bytes/s to 200 Bytes/s (extraction ratio 651 ppm) and saving 99.9% of network bandwidth and storage.
Original-vendor systems or wired accelerometers can only return large volumes of raw data or alarm on a simple RMS threshold. This solution is accurate, lightweight and can be rolled out in parallel at scale.
Alarm and Maintenance Workflow
- Edge feature computing: The sensor captures tens of thousands of high-frequency vibration samples per second, filters them and reduces the data.
- Physical AI life inference: Features travel over Modbus RS485 to the RULe.TI002, which computes the robot's HI and RUL in real time.
- Central FDC alarm and interlock: The RULe.TI002 sends clean features, HI and RUL to the FDC over Modbus TCP. When HI drops or RUL falls below the safety threshold (for example RUL < 10 days), a central alarm and an automatic interlock stop are triggered so the robot does not keep running in a degraded state.
- Planned maintenance: The FDC reports a clear failure cause (such as a worn Z-axis screw or joint bearing), and the maintenance team replaces the part in a planned window.
Benefits
- Avoid costly scrap and downtime (high ROI): Avoids USD 250,000 – 1,000,000 of direct scrap per incident and 4 – 12 hours of cleanroom downtime. One site is expected to protect USD 500,000 – 700,000 of operating assets per year.
- Higher OEE: Physical AI life prediction turns sudden failures into planned repairs, raising availability of the EFEM and bottleneck tools.
- Very low TCO: The data stream drops from 307.2 KB/s to 200 B/s, saving 99.9% of storage and bandwidth cost. The direct FDC connection adds no IT development effort.
(Note: Financial figures are estimated from public industry reports and typical advanced-node production lines.)
Customer Track Record
- Customer: TSMC, Arizona fab and Southern Taiwan Science Park advanced packaging fab
- Equipment: EFEM ATM wafer robot
- Solution level: Level 4 life prediction
- Products: Aurora® 313MM2-R1 + RULe.TI002
- Result: Real-time analysis of the robot's motion caught warning signs of excessive backlash and joint bearing wear. Direct Modbus TCP connection to the FDC allows central management and prevents sudden gripper crashes and wafer scrap.
Solution
- Aurora Equipment Health Monitoring Platform
- Consulting Service
If you are interested in this solution, please contact your local distributor for more information and a quotation. The full case document (in Chinese) is available below.