Learning Wafer Map Defects with a Channel-Attentive DenseNet and Dual Pooling

DOI

10.2991/978-94-6239-616-6_78How to use a DOI?

Keywords

Wafer-map defect classification; convolutional neural networks; DenseNet-121; WaferNet; WM-811K/LSWMD; class-imbalance handling; minority-class augmentation (rotations, flips); lot-disjoint stratified splits; two-channel tensors; binary wafer mask; mask-aware dual pooling (GAP/GMP); confusion matrices; macro-F1; channel attention; Apple Silicon MPS; CUDA; PyTorch

Abstract

Automated classification of defects in wafer maps continues to face challenges. Wafer shapes vary, there is a large class imbalance with “none” defects being the largest class, and lot-level data leakage can lead to inflated performance estimates. To address these challenges, we have developed a robust and reproducible system that consists of a pragmatic data pipeline and an efficient model called WaferNet. We started by cleaning the large LSWMD dataset, which contains about 800,000 entries. We standardized the inconsistent class names and introduced a clear “unlabeled” category. To avoid data leakage and ensure accurate evaluation, we established lot-disjoint stratified splits. We kept wafers from the same production lot entirely within either the training or testing set while maintaining class ratios. We also standardized most inputs to a 96x96 dimension. In dealing with the class imbalance, we reduced the quantity of “none” class and raised the counts of the minority classes through simple augmentations that kept labels, such as rotations and flips. We are taking a two-channel tensor which holds the wafer-map and a binary mask of valid die areas as input. WaferNet, a light version of DenseNet, uses mask-aware pooling to ignore areas that contain no valid die. We also did class-prior logit adjustment during training to improve recognition of minority classes. Our primary contribution is this practical, end-to-end framework that highlights label hygiene, lot-disjoint evaluation, dimension-constrained masked inputs, and a balanced training strategy. We fully document the process to ensure reproducibility and provide a solid baseline for future yield-analysis workflows.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - B. K. Vishvajith
AU  - C. Gokularaman
AU  - Arkat Charishma
AU  - S. Margret Anouncia
PY  - 2026
DA  - 2026/03/31
TI  - Learning Wafer Map Defects with a Channel-Attentive DenseNet and Dual Pooling
BT  - Proceedings of the International Conference on Artificial Intelligence and Secure Data Analytics (ICAISDA 2025)
PB  - Atlantis Press
SP  - 1072
EP  - 1090
SN  - 1951-6851
UR  - https://doi.org/10.2991/978-94-6239-616-6_78
DO  - 10.2991/978-94-6239-616-6_78
ID  - Vishvajith2026
ER  -