# Noise-Robust Wafer Map Defect Classification via CNN-ESN Hybrid Architecture

## TL;DR

This paper introduces a hybrid CNN-ESN model for wafer map defect classification that improves robustness to input noise and perturbations.

## Contribution

The novel CNN-ESN hybrid architecture enhances robustness under test-time perturbations without requiring noise-aware training.

## Key findings

- The proposed model achieved 0.61 pp higher test accuracy than ResNet34 under clean conditions.
- In noisy test scenarios, the model maintained 87.30% accuracy compared to 77.59% for CNN baselines.
- Robustness improvements were most pronounced for defect types with repetitive structural patterns.

## Abstract

Wafer map defect classification plays a critical role in yield monitoring and root-cause analysis in semiconductor manufacturing. Although recent convolutional neural network (CNN)-based approaches have achieved high classification accuracy, most existing models are evaluated primarily on clean datasets and remain vulnerable to unseen perturbations and representation-level variability at test time. In this paper, we propose a hybrid CNN–echo state network (ESN) architecture that integrates spatial feature extraction with sequential aggregation to enhance robustness under input perturbations. The CNN backbone extracts two-dimensional feature maps, which are converted into ordered sequences using a multidirectional scanline strategy and processed by an ESN reservoir. The resulting sequential representations are combined with CNN features through a class-specific adaptive fusion mechanism. Using the defect-only eight-class version of the WM-811K dataset, we systematically evaluate robustness under multiple perturbation scenarios, with particular focus on the clean train/noisy test (CT-NT) setting. To ensure a controlled robustness evaluation aligned with the binary nature of wafer map data, we introduce binary-consistent die-flip perturbations and additionally employ additive Gaussian perturbations as a representation-level stress test. Under clean-data conditions, the proposed model showed a 0.61 pp improvement in test accuracy compared to the ResNet34-based CNN, with notably larger gains for rare classes and defect types exhibiting strong structural patterns. In the clean train/noisy test scenario, where the model was trained on clean wafer map data and evaluated under controlled test-time perturbations, the accuracy of the CNN baseline dropped to 77.59% at σ = 0.10, whereas the proposed hybrid model maintained an accuracy of 87.30%, resulting in an absolute improvement of 9.71 pp. Per-class analysis reveals that the robustness gain is class-dependent, with pronounced improvements for defect types exhibiting clear and repetitive structural patterns, such as Loc and Edge-Ring. Further mechanistic analysis demonstrates that the robustness improvement arises from enhanced representation stability and bounded reservoir dynamics, rather than from changes in CNN feature extraction or training regularization. These results demonstrate that the proposed CNN-ESN hybrid architecture provides meaningful advantages in terms of robustness under noisy evaluation conditions without requiring noise-aware training or prior knowledge of perturbation characteristics.