Aug 2026· International Test Conference in Asia· pp. 1-5· 0 citations· 12 references
Abstract
Wafer map defect pattern recognition is useful for production-line yield diagnosis, but parameter-heavy classifiers are inconvenient for resource-constrained test-floor computers. This paper evaluates a compact nine-class, single-label classifier that combines MobileNetV3-Small with a Convolutional Block Attention Module (CBAM). Focal loss and class-weighted sampling are evaluated on the long-tailed WM-811K distribution. To avoid optimistic bias caused by wafers from the same production lot appearing in different subsets, complete lots—rather than individual wafer maps—are assigned to disjoint 70%/15%/15% training, validation, and test splits. The proposed model obtains 96.75% accuracy and macro-F1 = 0.8562 with 0.974 M parameters, 24.2 times fewer than ResNet-50. ShuffleNetV2 achieves higher absolute macro-F1 (0.9139) with 1.263 M parameters, whereas the proposed model has the smallest parameter count and highest macro-F1-per-parameter ratio among the evaluated neural baselines. Ablations reveal that removing CBAM, replacing focal loss with cross-entropy, or removing the weighted sampler raises macro-F1 to 0.9148, 0.9180, and 0.9058, respectively; hence the current single-run evidence does not support an accuracy benefit from these components. On the reference workstation, model-only batch-1 latency is 2.22 ms on an RTX 5090 GPU and 10.29 ms on the host CPU. PyTorch dynamic quantization reduces CPU latency to 10.19 ms but affects only the final linear layer, so its size reduction is negligible. These results position the model as a compact candidate for asynchronous classification on an ATE-side industrial computer, while also exposing class-specific and quantization limitations relevant to deployment.
A dual-task collaborative learning framework that decomposes conventional nine-class classification into binary defect detection and eight-class defect recognition and a shared convolutional backbone is combined with Multi-Scale Context-Aware Attention to enhance multi-scale defect representation and a task-gated Bottl...
Zhi-Zhu Ge, Zhao Ding, Ze-Tian Tang et al.· Measurement science and tech...· 0 citations
DiffWaferNet, a lightweight differential attention-based network, is proposed to improve the performance of minority defect classes while preserving the accuracy of majority classes, and is well-suited for real-time industrial applications.
Md. Ryhan Uddin, Md. Saiful Islam, Fahmid Al Farid et al.· PeerJ Computer Science· 0 citations
Wafer-level testing is essential for detecting defects caused by fabrication process deviations. However, test failures may introduce misjudgment and even physical damage to defect free wafers, resulting in wafer test-induced defects (WTIDs) with significant impact on yield management. Although convolutional neural net...
Zhen-Yu Wang, Zhao-Hui Guo, Guang-Sheng Chen et al.· International Test Conferenc...· 0 citations
The results show that targeted attention can improve moderate-overlap detection without the cost of an additional high-resolution head, but they also expose unresolved localization, source-domain, and false-accept limitations.
Yao-Yin Chen· Journal of Computing and Ele...· 0 citations
A data-efficient wafer defect classification framework that integrates generative data augmentation with lightweight deep learning models that enables reliable wafer defect classification in data-constrained manufacturing environments, making it well suited for deployment in resource-limited semiconductor fabrication f...