Source-linked AI summary
SePArate: Segmenting Patterns from Defects in Wafer Manufacturing Using Weak Supervision
Dain Kwon, Changmin Shin, Sunjong Park, Kanghyun Choi, Hyeyoon Lee, Jaewon Jang, Minseok Choi, Jinho Lee
TL;DR
Manual semiconductor defect inspection cannot scale, while existing automated methods often lack the geometric information needed for root-cause analysis. SePArate uses weak supervision and three-phase training to segment mixed defect patterns from image-level labels, outperforming baselines across two datasets.
Problem
Existing classification methods discard defect geometry, while segmentation is difficult to train without large pixel-level annotated datasets.
Method
SePArate trains a segmentation framework through encoder classification pretraining, spatial-cue transfer, and synthetic mixed-defect fine-tuning using only image-level labels.
Results
SePArate consistently outperforms baselines on defect datasets, achieving mean mIoU of 67.42% on MATDefects and 69.04% on MixedWM38.
Takeaways & Limitations
SePArate advances wafer inspection from image-level classification to scalable pattern-level segmentation without costly pixel-level annotations.
Abstract
from arXiv · showhide
In semiconductor manufacturing, defect analysis is essential, but manual inspection cannot scale. However, existing automated inspection methods remain insufficient for root-cause analysis and process optimization. To this end, we present SePArate, a weakly supervised wafer defect segmentation method. SePArate enables pixel-level separation of patterns by leveraging only image-level annotations. It consists of a three-phase training: encoder pretraining, knowledge transfer to learn spatial cues, and training on synthetic mixed-defect data for accurate segmentation. Experiments demonstrate that SePArate outperforms the baselines.
1 Introduction
Semiconductor defect analysis supports tracing defect causes and refining manufacturing processes, but inspection data volumes make manual review infeasible. Existing automated methods often lose geometric information needed for root-cause analysis, motivating SePArate’s weakly supervised pixel-level segmentation.
- Defect analysis helps engineers trace underlying causes and refine semiconductor manufacturing processes.
- Terabytes of daily inspection data and diverse pattern variations make manual inspection unable to scale.
- Classification methods detect predefined patterns but discard geometric attributes such as shape, area, and orientation.
- SePArate uses only classification labels to generate pixel-level masks that isolate individual defect patterns.
- Experiments on production-line DRAM mat-level and public wafer-level datasets show SePArate significantly outperforms baselines.
2 Background
Wafer defect datasets span mat-level DRAM and wafer-level images, while segmentation must separate mixed patterns using limited supervision. Because binary images lack visual semantics, SePArate leverages structural and distributional cues instead.
- Defect Data in Wafer Manufacturing: MATDefects represents DRAM cells as pixels and includes row-wise, column-wise, group, and mixed-type defect patterns.
- Defect Data in Wafer Manufacturing: MixedWM38 contains full-wafer images with die-level pixels and 38 single and mixed-type patterns.
- Segmentation without Ground-Truth Data: Available approaches include zero-shot foundation models and weakly supervised segmentation methods trained from image-level labels.
- Segmentation without Ground-Truth Data: Zero-shot models and conventional WSSS methods rely on visual cues such as color, texture, or shading that are largely absent from binary defect data.
- Segmentation without Ground-Truth Data: SePArate explicitly leverages structural and distributional cues without depending on visual semantics.
- Related Work: Prior segmentation works assume one defect instance per image, and SSB-Rec primarily targets classification rather than isolating mixed-type patterns.
3 Challenges in Defect Pattern Segmentation
Defect pattern segmentation is constrained by expensive annotations, binary-valued pixels, and sparse or imbalanced patterns. SePArate addresses these domain-specific challenges explicitly.
- Annotation and Data Constraints: Pixel-level annotations require substantial expert effort and are difficult to obtain through crowdsourcing because of semiconductor confidentiality requirements.
- Binary-Valued Pixels: Binary-valued images provide limited learnable features, forcing models to rely on structural patterns and distributional statistics.
- Sparsity of Defect Patterns: Defect patterns can be pixel-level sparse, with small defective regions surrounded by large backgrounds.
- Sparsity of Defect Patterns: Rare manufacturing causes create image-level sparsity and potential data imbalance among defect patterns.
- Proposed Response: SePArate explicitly addresses these domain-specific factors to achieve effective defect pattern segmentation quality.
4 Methodology
SePArate trains a U-Net in three phases, using image-level labels, spatial cues, and synthetic mixed-type defects to achieve weakly supervised segmentation. Soft supervision combines defect-density and classifier-derived pattern relevance, while weighted losses address scarce patterns and false positives.
- Overall Training: SePArate trains the model in three phases: encoder classification pretraining, spatial-cue adaptation, and final training with synthetic mixed-type defects.The process operates without expensive pixel-level annotations.
- Soft Supervision Maps: Phase 2 combines local density maps and CAMs to create pattern-aware soft supervision maps for U-Net training.Density maps highlight concentrated defect regions, while CAMs identify regions influencing classification decisions for each defect type.
- Pattern-Weighted Dice Loss: The pattern-weighted Dice Loss assigns image-pattern weights so optimization focuses on present patterns rather than absent-mask predictions.The loss is intended to stabilize learning across patterns with different prevalence and improve segmentation of scarce defects.
- Synthetic Mixed-Type Defects: Synthetic mixed-type masks are created by merging single-type defect masks, enabling Phase 3 fine-tuning with accurate segmentation masks.For binary-valued defect images, 1-valued pixels directly provide masks for single-type defects; mixed-type masks extend this process by combining them.
- Absence Penalty Loss & Scheduling: Phase 3 combines Pattern-Weighted Dice and Absence Penalty losses to penalize irrelevant or absent-pattern pixels.The Absence Penalty uses BCE-based supervision for predictions above a confidence threshold, whose value is gradually reduced from 0.5 to 0.0.
5 Experimental Results
SePArate is evaluated against weakly supervised segmentation baselines on MATDefects and MixedWM38, with ablations examining its phases and Phase 3 loss terms. It achieves strong segmentation performance, while the ablations show that spatial-cue learning and cumulative loss refinements are important.
- Performance Comparison: SePArate achieves average mIoU scores of 67.42% on MATDefects and 69.04% on MixedWM38, outperforming the evaluated baselines.The baselines show limited practical performance or fail on some binary defect patterns.
- Visualizations: SePArate successfully disentangles overlapping defect patterns in MATDefects and MixedWM38 while preserving morphological distinctions.Pixels belonging to the same pattern share colors in the visualizations, and the segmented regions retain pattern and shape information.
- Ablation Study of Phases: The full three-phase model reaches 67.42% mean mIoU, compared with 0.20% for Phase 1 alone and about 39.5% for partial phase combinations.Phase 2 or Phase 3 improves performance individually, but neither partial setting matches the complete model.
- Ablation Study of Phases: Without Phase 2, recall for the group pattern falls to 0.1%, whereas the full model better segments overlapping group and line defects.The ablation attributes this difference to Phase 2 supplying early spatial-context knowledge about defect regions.
- Loss Term Analysis: Adding the Absence Penalty and τ-Scheduling after Pattern-Weighted Dice Loss produces cumulative IoU gains by reducing false-positive pixels.The Absence Penalty supplies negative supervision for absent patterns, while decreasing τ makes training more conservative.
6 Conclusion
SePArate advances semiconductor defect inspection from image-level classification to pattern-level segmentation using only image-level labels. Experiments show that it consistently outperforms baselines while offering scalable, cost-effective pattern-level feedback.
- SePArate uses image-level labels to separate defect instances into pattern-level segmentations without costly pixel-level annotations.The framework is presented as a weakly supervised approach for automatic semiconductor defect inspection.
- Experiments on defect datasets show that SePArate consistently outperforms baselines.
- Pattern-level feedback from SePArate can support optical inspection, long-term monitoring, and continuous manufacturing improvement.