Virtual metrology and Data Drift Recovery for a Leading Memory Products OEM
Background
A top manufacturer for memory products is producing over half a million memory wafers yearly. This use case highlights a single processing step and its following quality test, for simplicity.
- Wafer processing is composed of over 1,000 steps (of different modules: Etching, CMP, Litho, etc.)
- Time-series data from sensors installed in the equipment is being collected for each wafer during the process.
- The fabrication process includes multiple quality control steps, which are expensive and time- consuming. Each wafer is going through metrology inspection in 13 fixed areas.
- Data drifts occur naturally as processing equipment is aggregating particles residue and its condition changes over time.
Challenges
- Decrease overall cycle-time:
- Reduce quality sampling time during the physical metrology inspection stage
- Detect defective wafers early in the manufacturing process
- Deploy a model immune to the naturally recurring data drifts in this process.
Solution process
- Reduce the quality sampling time and detect defective wafers early:
- This is achieved through training a model to predict the metrology test results based on data from the etching process.
- Step 1: The user defined the desired acceptance criteria for Vanti’s model performance by MAPE and P95 error thresholds for the sampled areas on the wafer.
- Step 2: Historical time-series data from the etching step alongside their corresponding quality test results for hundreds of wafers was used to train 13 regression models1
- Each model successfully complied to the pre-set success criteria and corresponded to a specific set of coordinates (X,Y) for each of the 13 sampled areas during metrology inspection.
- The models provided real-time defect predictions of all sampled areas, thereby reducing the OEM’s dependency on physical metrology testing.
- Deploy a model immune to naturally recurring data drifts:
- Vanti’s service includes a proprietary drift-recovery mechanism. Once a drift is detected, production models can be automatically replaced within minutes with a new model.
- Vanti continuously monitors the performance of deployed models to automatically improve performance over time.
Results
Using Vanti’s service the OEM is now able to drastically minimize the previous
dependency on physical quality sampling during metrology inspections, thereby
reducing cycle time. Moreover, employing Vanti’s proprietary drift recovery mechanisms
and auto-optimization of deployed models ensure a high quality of service at all times.
