Semiconductor Yield Virtual Metrology

Machine Learning-based Quality Index Prediction Model for Semiconductor Manufacturing Process

SK-Planet Data Analyst Training Program Project (2023.01.03 ~ 2023.03.09)

---

Overview

This project develops a Virtual Metrology (VM) model that predicts the quality index of semiconductor manufacturing processes using sensor data. The model leverages 665 sensor variables across 7 process steps, applying feature engineering, multicollinearity treatment, and hyperparameter optimization techniques.

Key Features

• Feature Engineering: 665 sensor variables → 800+ derived features
• Multicollinearity Treatment: PCA dimensionality reduction for VIF > 10 variables
• Feature Selection: SelectKBest (Mutual Information) based top 250 features
• Bayesian Optimization: Automated hyperparameter tuning for Ridge, Random Forest
• AutoML: Multi-model comparison and auto-tuning with PyCaret

Quick Stats

Metric	Value
Sample Size	611 observations (train)
Sensor Variables	665 features
Process Steps	7 steps (04, 06, 12, 13, 17, 18, 20)
Sensor Types	95 sensor categories
Target Variable	Quality Index (mean ~1263)
Generated Features	200+ engineered features

---

Sample Results

Target Distribution Analysis

Mean: 1263.41
Variance: 67.16
Distribution: Approximately Normal
Outliers: y < 1240 (clipped during preprocessing)

Process Duration Analysis

Total Process Time: 30.6 ~ 31.9 minutes
Critical Step Gap: Step 06 → Step 12 (longest duration)
Speed Classification: 1870 seconds threshold
  - Early (E): EQ7, EQ8 modules
  - Late (L): Other modules

Model Performance (AutoML)

Best Model: Extra Trees Regressor
Evaluation Metric: RMSE
Cross-validation: 5-fold

---

Quick Start

Prerequisites

# Python 3.8 or higher
python --version

# Install dependencies
pip install -r requirements.txt

Installation

# Clone the repository
git clone git@github.com:jinsoo96/Semiconductor_Yield_Virtual_Metrology.git
cd Semiconductor_Yield_Virtual_Metrology

# Install Python packages
pip install -r requirements.txt

Run Analysis

Full Pipeline:

# Step 1: EDA and Data Preprocessing
jupyter notebook 01_Code/01_eda_and_preprocessing.ipynb

# Step 2: Modeling (AutoML + Bayesian Optimization)
jupyter notebook 01_Code/02_modeling.ipynb

Note: Common functions are centralized in 01_Code/utils.py to avoid code duplication.

Execution Time: ~30-45 minutes for full pipeline

---

Repository Structure

Semiconductor_Yield_Virtual_Metrology/
│
├── 01_Code/
│   ├── utils.py                          # Common utility functions (shared module)
│   ├── 01_eda_and_preprocessing.ipynb    # EDA & Data Preprocessing
│   └── 02_modeling.ipynb                 # All modeling (AutoML + Bayesian Opt)
│
├── 02_Data/
│   ├── DATA_INFO.txt                     # Data documentation
│   └── raw/
│       ├── train_sensor.csv              # Training sensor data (~24.5MB)
│       ├── train_quality.csv             # Training quality labels (~25KB)
│       └── predict_sensor.csv            # Test sensor data (~10.5MB)
│
├── 03_Results/
│   ├── figures/                          # Generated plots
│   ├── tables/                           # Result tables
│   ├── preprocessed_data.pkl             # Preprocessed data (generated)
│   └── best_model.pkl                    # Trained model (generated)
│
├── 04_Documentation/
│   ├── final_presentation.pdf            # Final presentation slides
│   ├── Code_Structure.md                 # Detailed code guide
│   └── Analysis_Workflow.md              # Step-by-step workflow
│
├── archive/                              # Original development files
│
├── README.md                             # This file
├── requirements.txt                      # Python dependencies
├── LICENSE                               # MIT License
└── .gitignore                            # Git ignore rules

---

Methodology

1. Data Collection & Preprocessing

• Source: SK HYNIX semiconductor manufacturing data
• Period: October 2021
• Sample: 611 LOT observations with 665 sensor features

Preprocessing Steps:

1. Pivot transformation (Long → Wide format)
2. Feature generation from step_id + param_alias
3. Time feature extraction (process duration)
4. Missing value handling
5. Outlier treatment (IQR-based clipping)
6. Standardization (StandardScaler)

2. Feature Engineering

Feature Type	Description	Count
Original Sensors	Raw sensor measurements per step	665
Duration Features	Total and inter-step process time	21
Statistical Features	Sensor std/mean across steps	190
Categorical Features	Equipment category encoding	8+
Binned Features	Continuous variable discretization	30+

3. Feature Selection Pipeline

665 Original Features
    ↓
+200 Generated Features (865 total)
    ↓
Variance Threshold (remove zero-variance)
    ↓
VIF Analysis (identify multicollinearity)
    ↓
PCA (reduce VIF>10 features)
    ↓
SelectKBest (top 250 by Mutual Information)
    ↓
Final Feature Set (~300 features)

4. Model Training

Models Evaluated:

1. Ridge Regression (Bayesian Optimized)
2. Random Forest Regressor (Bayesian Optimized)
3. Extra Trees Regressor
4. CatBoost Regressor
5. LightGBM Regressor
6. Gradient Boosting Regressor

Training Strategy:

• Train/Valid/Test split: 64% / 16% / 20%
• RandomOverSampler for class imbalance
• Log transformation on target (optional)
• 5-fold Cross-validation

---

Key Findings & Insights

1. Target Variable Analysis

Finding: Target variable y follows approximately normal distribution (mean: 1263.41, variance: 67.16)

• Insight: Normal distribution assumption is valid, favorable for regression models
• Action: Clip outliers below y < 1240 for model stability

2. Equipment Effect (module_name_eq)

Finding: Significant quality index variation exists across equipment (EQ1~EQ8)

• Insight: EQ7, EQ8 show relatively lower quality indices compared to others
• Action: One-hot encode module_name_eq as categorical feature

3. Process Duration Impact

Finding: Total process duration clearly separates into two groups at 1870 seconds threshold

• Early Group (E): ~30.6 min, mainly EQ7, EQ8 modules
• Late Group (L): ~31.9 min, most modules
• Insight: Correlation exists between process speed and quality
• Action: Create tmdiff_speed categorical variable as model feature

4. Critical Process Steps

Finding: Step 06 → Step 12 shows longest duration and highest variability

• Insight: This interval is estimated to have the greatest impact on quality
• Action: Generate inter-step duration features (gen_tmdiff_0612, etc.)

5. Sensor Aggregation Value

Finding: Aggregated sensor statistics (std, mean) across steps show higher predictive power than individual step values

• Insight: Step-wise variability is a key indicator for quality prediction
• Action: Generate 95 sensor std (gen_{sensor}std) and mean (gen{sensor}_mean) features

6. Multicollinearity Issue

Finding: Multiple variables with VIF > 10 exist (multicollinearity problem)

• Insight: High correlation among sensor variables poses model instability risk
• Action: Apply PCA to VIF > 10 variables, reduce to 2 principal components

7. Feature Importance

Finding: Model performance maintained with top 250 features selected by Mutual Information

• Key Features:
  • Process duration features (gen_tmdiff_*)
  • Aggregated sensor statistics (gen_std, gen_mean)
  • Equipment categorical features
• Insight: Derived features show higher predictive power than original sensor variables

8. Class Imbalance

Finding: Target variable shows imbalance when binned into 3 categories (A: y<1242, B: 1242≤y≤1283, C: y>1283)

• Distribution: Most samples concentrated in B category
• Action: Apply RandomOverSampler for minority class oversampling

---

Code Example

import pandas as pd
import numpy as np
from sklearn.preprocessing import StandardScaler
from sklearn.feature_selection import SelectKBest, mutual_info_regression
from sklearn.model_selection import train_test_split

# Import utility functions
from utils import (
    load_data, make_dataset,
    gen_cate_feats, gen_duration_feats, gen_stats_feats,
    LST_STEPS, LST_STEPSGAP
)

# Load data
path = "./02_Data/raw/"
train_sensor, train_quality, predict_sensor = load_data(path)

# Create dataset
train = make_dataset(train_sensor, train_quality)

# Feature engineering
train = gen_cate_feats(train)           # Equipment category
train = gen_duration_feats(train, LST_STEPSGAP)  # Process duration
train = gen_stats_feats(train, sensors_nm, LST_STEPS)  # Sensor statistics

# Feature selection
skb = SelectKBest(score_func=mutual_info_regression, k=250)
X_selected = skb.fit_transform(X, y)

# Train model
from pycaret.regression import setup, compare_models, tune_model
reg = setup(data=train_df, target='y', normalize=True)
best = compare_models(sort='RMSE')
best_tuned = tune_model(best)

---

Requirements

Python Packages

pandas>=1.3.0
numpy>=1.21.0
scipy>=1.7.0
scikit-learn>=1.0.0
catboost>=1.0.0
lightgbm>=3.3.0
xgboost>=1.5.0
pycaret>=2.3.0
bayesian-optimization>=1.2.0
imbalanced-learn>=0.9.0
statsmodels>=0.13.0
matplotlib>=3.4.0
seaborn>=0.11.0

System Requirements

• Python: 3.8 or higher
• RAM: 8GB minimum (16GB recommended)
• Storage: 500MB for code and data
• OS: Windows, macOS, or Linux

---

Team Members

Name	Role
Yukyung Lim	Data Analysis & Modeling
Jin Soo Kim	Feature Engineering & Optimization
Hojin Lee	EDA & Visualization
Seungah Ahn	Preprocessing & Documentation

---

Documentation

Document	Description
README.md	Project overview (this file)
Code_Structure.md	Detailed code documentation
Analysis_Workflow.md	Step-by-step workflow
final_presentation.pdf	Final presentation slides

---

References

• PyCaret Documentation
• Bayesian Optimization
• Scikit-learn Feature Selection
• Virtual Metrology in Semiconductor Manufacturing

---

License

This project is licensed under the MIT License - see the LICENSE file for details.

MIT License
Copyright (c) 2023 Jin Soo Kim

---

Contact

• Author: Jin Soo Kim
• GitHub: @jinsoo96