Related Experiment Video
Updated: Mar 29, 2026
08:27
Author Spotlight: Efficient Image Recognition Using Directional Gradient Histogram Technique and Support Vector Machines
Published on: January 5, 2024
1.7K
Noise-Robust Wafer Map Defect Classification via CNN-ESN Hybrid Architecture
Hayeon Choi1, Dasom Im1, Sangeun Oh1
1Department of System Semiconductor Engineering, Sangmyung University, Cheonan 31066, Republic of Korea.
Micromachines
|March 28, 2026
Summary
A new hybrid CNN-ESN model improves wafer map defect classification robustness against noise. This approach enhances semiconductor manufacturing yield monitoring by maintaining accuracy under perturbations without needing noise-aware training.
Area of Science:
- Semiconductor Manufacturing
- Artificial Intelligence
- Machine Learning
Background:
- Wafer map defect classification is crucial for semiconductor yield monitoring and root-cause analysis.
- Current Convolutional Neural Network (CNN) models struggle with robustness against input perturbations and variability at test time.
- Existing methods are often evaluated on clean datasets, limiting their real-world applicability.
Purpose of the Study:
- To propose a hybrid CNN-echo state network (ESN) architecture for enhanced robustness in wafer map defect classification.
- To improve model performance under various perturbation scenarios, particularly the clean train/noisy test (CT-NT) setting.
- To investigate the mechanisms behind robustness improvements in the proposed hybrid model.
Main Methods:
- Developed a hybrid CNN-ESN model integrating spatial feature extraction (CNN) with sequential aggregation (ESN).
- Utilized a multidirectional scanline strategy to convert 2D CNN feature maps into sequences for ESN processing.
- Implemented a class-specific adaptive fusion mechanism to combine CNN and ESN features, evaluated using binary-consistent die-flip and Gaussian perturbations on the WM-811K dataset.
Main Results:
- The hybrid CNN-ESN model achieved a 0.61 pp test accuracy improvement over a ResNet34-based CNN on clean data, with greater gains for rare and structurally patterned defects.
- In the CT-NT scenario, the hybrid model maintained 87.30% accuracy, a significant 9.71 pp improvement over the CNN baseline (77.59% at σ = 0.10).
- Robustness gains were class-dependent, particularly benefiting defects like Loc and Edge-Ring, and attributed to enhanced representation stability and bounded reservoir dynamics.
Conclusions:
- The proposed CNN-ESN hybrid architecture significantly enhances robustness against input perturbations in wafer map defect classification.
- The model demonstrates superior performance in noisy test conditions without requiring noise-aware training or prior knowledge of perturbations.
- This approach offers a promising solution for reliable yield monitoring and root-cause analysis in semiconductor manufacturing under real-world conditions.