More data does not automatically mean a clearer decision.
As the owner of a laser-based wafer inspection system at Transphorm, I worked across inspection recipes, maintenance, data analysis, and production support. The challenge was to distinguish consequential defects from benign features and turn those distinctions into reliable grading workflows.
An optical inspection system can describe a single feature through many correlated attributes. Physical understanding helps determine which signals are meaningful; statistics helps make those distinctions repeatable.
Connecting the analysis tools.
I built Python-based data-mining and ETL tools to extract inspection results, work with defect imagery, and connect measurements to SQL data systems. I used statistical analysis in JMP including clustering, discriminant analysis, and extensive scripting to develop and continuously refine classification rules.
The work combined inspection knowledge, image characteristics, and historical examples. The result was a workflow for automated defect classification and wafer grading that replaced substantial manual inspection effort.
- 01 / INGESTCollect and organize inspection results.
- 02 / CLASSIFYUse physical signatures and statistical analysis.
- 03 / DECIDEConnect classifications to wafer grading.
Validate against reviewed defects and wafer dispositions.
I compared the automated results with engineer-reviewed defect populations and known wafer dispositions. Nuisance detections and visually similar defect classes exposed incorrect classifications, so I tuned the workflow around the false-positive/false-negative tradeoff that mattered to manufacturing.
Greater than 95% wafer-grading accuracy.
The resulting system achieved greater than 95% grading accuracy against the reviewed wafer dispositions and saved more than 1,000 labor hours per year. Engineers could inspect automatically classified populations and focus on exceptions instead of manually grading every wafer.