Yield Decision WorkflowExplainable RecommendationCross-Factory Analytics
Yield Optimization Workflow
A manufacturing optimization workflow built for adoption, not just prediction.
AI Engineer / Optimization Lead2021 - 2023AUO Corporation
Yield Optimization · Cross-Fab Workflow
Cross-Factory AnalyticsGenetic OptimizationLIME Explainability
- 01
Factory Data
Multi-Site Process & Sensor
- 02
Standardize
Unified Process Definitions
- 03
Feature Build
SQL / ETL Signal Prep
- 04
GA Optimize
Genetic Algorithm Search
- Trust Layer05
LIME Explain
Transparent Factor Impact
- 06
Action Recs
+1.43% Yield · −12.62% Def
Frames optimization as explainable equipment & parameter guidance rather than opaque prediction scores.
Before
Similar products made in different factories showed yield gaps, but teams lacked a practical way to turn that variability into usable improvement actions.
After
A cross-factory analytics and optimization workflow translated process differences into explainable recommendations for engineering teams.
Context & Background
- Different factories produced similar products with meaningful yield variation.
- The real business need was not another model score, but recommendations teams could operationalize.
Constraints
- Data semantics varied across sites and needed normalization first.
- Recommendations had to be explainable enough for engineers to review and trust.
Problem Solved
- Cross-factory yield differences were hard to analyze consistently.
- Model outputs without explanation were difficult for operations teams to use.
My Role
- The workflow was designed around adoption: how data should be standardized, how optimization should run, and how results should be communicated.
- Analysis output was shaped into something production teams could use in decision-making.
System Design
- Combined cross-site data normalization, optimization logic, and explainability into one workflow.
- Framed outputs as parameter and equipment recommendations rather than opaque prediction scores.
Implementation Highlights
- Standardized process and sensor meaning across factories before optimization.
- Built a recommendation workflow instead of a score-only analytics pipeline.
- Added explainability to help engineers review and act on output.
Outcome & Impact
- Helped teams identify actionable optimization directions across factories.
- Improved adoption by making model output easier to interpret and discuss.
Tech Stack
PythonScikit-learnLIMEGenetic AlgorithmSQL / ETL