Paper deep dive
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
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 9/15/2026, 8:35:37 AM
Summary
The paper introduces SePArate, a weakly supervised semantic segmentation framework designed to isolate individual defect patterns from mixed-type binary defect images in semiconductor wafer manufacturing. Unlike existing methods that rely on pixel-level annotations or visual cues absent in binary data, SePArate leverages only image-level labels. It employs a three-phase training strategy: encoder pretraining for multi-label classification, knowledge transfer using soft supervision maps (combining Class Activation Maps and local density maps) with pattern-weighted Dice loss, and fine-tuning on synthetic mixed-defect data with an absence penalty loss. Experiments on MATDefects and MixedWM38 datasets demonstrate that SePArate significantly outperforms baseline methods in mean Intersection over Union (mIoU).
Entities (12)
Relation Signals (10)
SePArate → evaluatedon → MATDefects
confidence 95% · Results on a production-line DRAM mat-level dataset... MATDefects... demonstrate that SePArate significantly outperforms baselines.
SePArate → evaluatedon → MixedWM38
confidence 95% · Results on... a public wafer-level dataset MixedWM38... demonstrate that SePArate significantly outperforms baselines.
SePArate → uses → U-Net
confidence 95% · We use the U-Net [20] architecture... SePArate trains the model through three phases... adapting it to learn spatial cues... U-Net Model
SePArate → employs → Pattern-Weighted Dice Loss
confidence 90% · The U-Net is trained to align its output with these maps, using the pattern-weighted Dice loss
SePArate → employs → Absence Penalty Loss
confidence 90% · When training the model in Phase 3, we train the model with the aforementioned Pattern-Weighted Dice Loss and the Absence Penalty (AP) Loss.
SePArate → generates → Soft Supervision Maps
confidence 90% · Soft supervision maps act as intermediate pseudo-labels... combine local density maps and class activation maps (CAM)
SePArate → outperforms → CLIPSeg
confidence 90% · Table 1: Comparison of Mean IoU... SePArate 67.42 (MATDefects Mean) vs CLIPSeg 4.21
SePArate → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2608.30410v1
- Canonical: https://arxiv.org/abs/2608.30410v1
Trouble viewing inline? Open PDF directly →
Full Text
37,507 characters extracted from source content.
Expand or collapse full text
SePArate: Segmenting Patterns from Defects in Wafer Manufacturing Using Weak Supervision Dain Kwon Seoul National University Seoul, South Korea dain.kwon@snu.ac.kr Changmin Shin Seoul National University Seoul, South Korea scm8432@snu.ac.kr Sunjong Park Seoul National University Seoul, South Korea ryan0507@snu.ac.kr Kanghyun Choi Seoul National University Seoul, South Korea kanghyun.choi@snu.ac.kr Hyeyoon Lee Seoul National University Seoul, South Korea hylee817@snu.ac.kr Jaewon Jang SK Hynix Seoul, South Korea jaewon2.jang@sk.com Minseok Choi SK Hynix Seoul, South Korea choineral@gmail.com Jinho Lee Seoul National University Seoul, South Korea leejinho@snu.ac.kr Abstract In semiconductor manufacturing, defect analysis is essential, but manual inspection cannot scale. However, existing automated in- spection 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 an- notations. 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 demon- strate that SePArate outperforms the baselines. The code is available at https://github.com/meowrowan/SePArate. Keywords Semiconductor Defects, Mixed-type Defect Patterns, Image Seg- mentation, Weakly Supervised Semantic Segmentation ACM Reference Format: Dain Kwon, Changmin Shin, Sunjong Park, Kanghyun Choi, Hyeyoon Lee, Jaewon Jang, Minseok Choi, and Jinho Lee. 2026. SePArate: Segmenting Patterns from Defects in Wafer Manufacturing Using Weak Supervision. In 63rd ACM/IEEE Design Automation Conference (DAC ’26), July 26–29, 2026, Long Beach, CA, USA. ACM, New York, NY, USA, 7 pages. https://doi.org/10. 1145/3770743.3804046 1 Introduction Defect analysis is an essential step in semiconductor manufacturing to minimize faults during production. By inspecting patterns of failures, engineers can trace underlying causes of defects, leading to process refinement based on the found issues [2,4,23]. Figure 1 illustrates an inspection scenario. During the production, defects This work is licensed under a Creative Commons Attribution 4.0 International License. DAC ’26, Long Beach, CA, USA © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2254-7/2026/07 https://doi.org/10.1145/3770743.3804046 Database for past defects Wafer Manufacturing Process mat-level wafer-level Extract defect patterns Inspection & Quality Control Flow Categorized pattern instances Defect Analysis Locate Causes Feedback for continuous process improvements Pattern Classification Defect Pattern Images Further Automation Identify Issues Assemble defect instances for comparison Automation with classification models Automation with SePArate Process requiring manual inspection Figure 1: Automation of the overall visual inspection of de- fects and continuous improvement by integrating SePArate. are detected, either at the wafer-level, or at the specific component- level (e.g., mat-level for a DRAM [8]). These defect maps become classified images organized into datasets. Then, experts analyze these to identify issues, trace their causes, and provide feedback. However, the sheer volume of inspection data makes manual inspection infeasible. Modern manufacturing processes generate terabytes of defect data each day [12,17], containing diverse and intricate pattern variations that could guide further manufacturing process optimization. In such conditions, human inspection can- not scale to the full dataset. Therefore, automated defect pattern analysis is a natural progression for wafer defect detection. Despite recent advances, current AI-based defect analysis meth- ods [5,16] remain limited in practical utility. Many of them are classification methods, i.e., they detect whether predefined patterns exist on wafers. This leads to a significant limitation: They discard geometric attributes such as shape, area, and orientation. These at- tributes provide essential information about the root cause of the defects and clues on how to fix them. Consequently, automated pixel-level defect pattern annotation is necessary, but AI-based seg- mentation methods do not support wafer defect analysis and are hard to train due to the absence of a large, annotated dataset. arXiv:2608.30410v1 [cs.CV] 31 Aug 2026 DAC ’26, July 26–29, 2026, Long Beach, CA, USADain Kwon et al. To address this, we propose SePArate: a weakly supervised de- fect pattern segmentation approach requiring only classification labels without costly pixel-level annotations. SePArate generates pixel-level masks that isolate individual patterns from defect im- ages, enabling finer categorization, new pattern discovery, and automated database construction that previously required manual curation. An end-to-end automation of the defect analysis pipeline can be initiated, as illustrated in Figure 1. Using a combination of self-generated supervision and a task-specific fine-tuning ap- proach, SePArate effectively isolates diverse defect patterns. Results on a production-line DRAM mat-level dataset and a public wafer- level dataset demonstrate that SePArate significantly outperforms baselines. The key contributions of our work are as follows. • We are the first to tackle segmentation for binary mixed-type defect pattern images with weak supervision. •We devise SePArate, a framework that trains a defect pattern segmentation network under image-level supervision, without re- quiring additional costly instance annotations. •SePArate greatly surpasses existing baselines, highlighting the value of its defect-aware design. 2 Background 2.1 Defect Data in Wafer Manufacturing Defects occur at various granularities in wafer manufacturing, as shown in the leftmost side of Figure 1. For generalization, we use two datasets of different granularity: a private, DRAM mat-level de- fect dataset MATDefects, obtained from an actual DRAM manufac- turing process, and a public wafer-level defect dataset MixedWM38 [25]. The pixels are binary: 0 for normal, and 1 for defects. MATDefects contains data from DRAM products, where each pixel represents a DRAM cell (i.e., a bit). A DRAM die [8] often exhibits a regular structure composed of multiple banks, each com- prising multiple mats. Each mat is a square cell array, typically 512-2,048 in size, depending on the product. MATDefects includes single and mixed-type defect patterns. The single defect patterns are classified as: row-wise (wordline or wordline driver issues), column-wise (bitline or sense amplifier issues), or group (individual cell issues). As in Figure 2, multiple single-type defects may appear in an image, forming a mixed-type defect pattern. MixedWM38 contains images generated from full wafers, where each pixel corresponds to a die. It contains 38 single and mixed-type patterns, with 8 single-type patterns, also shown in Figure 2. 2.2 Segmentation without Ground-Truth Data Defect pattern datasets usually provide only image-level ground- truths, but we need pixel-level annotation for each pattern to do segmentation. Two approaches are feasible: zero-shot foundation models for segmentation [11,19], or weakly supervised semantic segmentation (WSSS) methods [1, 9, 26]. Zero-shot models such as CLIPSeg [11] and GroundedSAM [19] use vision-language pretraining to align visual features with text semantics. As they enable text-guided segmentation without labeled examples, defect-pattern text prompts may be used to guide them. There are also WSSS methods [1,9,26], which train segmentation models with image-level labels. Most methods train a classifier, generate pixel-level pseudo-labels from intermediate activations MATDefects Example Images MixedWM38 Example Images Mixed-Type R + C + Group Column (C)Row (R)Group Single-Type Defects R + C + Group Mixed-Type Defects Single-Type Defects NormalRandom (R)Scratch (S) NearFull (NF)Loc (L) L + S EdgeRing (ER) EL + S EdgeLoc (EL) Donut (D)Center (C) D + L C + ER D + S EL + L C + L + ER + S D + L + EL+ S C + L + EL+ S Figure 2: Single and mixed-type defect pattern examples for MATDefects and MixedWM38. (e.g., Class Activation Maps (CAMs) [1, 9] or attention maps [26]), then train a segmentation model with these synthetic labels. The main challenge with these alternatives is that they rely on visual cues–such as color, texture, or shading–that are largely absent in our binary data. With no diversity in pixel values to infer about shape, size, or depth, we are only left with geometric structure to differentiate between patterns. Therefore, a new defect-aware approach is required: a weakly supervised framework that explicitly leverages structural and dis- tributional cues without depending on visual semantics. SePArate is explicitly designed for this setting, providing a practical solution for defect pattern segmentation. 2.3 Existing Works for Binary Defect Data For binary defect images, existing works have focused on multi- label classification, detecting the presence of each single-type defect pattern in mixed-type defects. The classification models evolve from CNNs [6,10,21,24,25,28] to transformers and state space models [5,16]. There exist a few segmentation works, [14,15] but they assume only a single defect instance per image and thus cannot separate multiple single-type defect patterns. One work, SSB-Rec [27], employs defect segmentation-like train- ing. However, its primary objective remains to enhance classifica- tion quality. We include it as a baseline to test its applicability to defect pattern segmentation. To our knowledge, no prior work has attempted to isolate single-defect patterns from mixed-type defects. 3 Challenges in Defect Pattern Segmentation The goal is to separate defect pixels into pattern-specific subsets within a defective region. However, defect pattern segmentation poses unique challenges that are not present in natural image seg- mentation. This is because, although both are image representations, they have basically different characteristics. We list them as below. (1)Difficulty of obtaining high-quality labels. Obtaining pixel- level annotations is particularly challenging and expensive. The pixel-level labels require substantial efforts from experts, com- pared to labels for image classification. Moreover, common SePArate : Segmenting Patterns from Defects in Wafer Manufacturing Using Weak SupervisionDAC ’26, July 26–29, 2026, Long Beach, CA, USA Unet Model Phase 2 Model Output � CLS � CLS 3x3 Kernel Conv. 1 1 1 1 1 1 1 1 1 Images [1,0, ... ,0] Labels C 1 C 2 C k ... ... Unet Model Classification Head Element-wise Multiplication � DICE Density MapClass-level CAMsSoft Supervision Map Images Unet Model Model Output � DICE Ground-Truth Masks for I syn Synthetic Multi-Defect Image I syn Single-Defect Real Images SumStack Phase 3Phase 1 GradCAM Unet Enc. Classif. Head Classification Head Images ... ... [1,0, ... ,0] Labels C 1 C 2 C k ... Images ... ... U-Net Model Phase 2 [1,0, ... ,0] Labels C 1 C 2 C k ... Model Output Classification Head Mean Filter 1 1 1 1 1 1 1 1 1 Classification head U-Net Encoder CAM Images [1,0, ... ,0] Labels C 1 C 2 C k ... ... U-Net Model Hadamard Product Density MapClass-level CAMsSoft Supervision Map Images U-Net Model Model Output Ground Truth Masks Synthetic Mixed-Type Defect Image Single-Defect Real Images Sum Stack Phase 3Phase 1 ℒ PWDL λ dice ℒ PWDL +λ ap ℒ AP ℒ CLS Figure 3: A full overview of our proposed method, SePArate. i) Phase 1 trains the encoder for multi-label classification, enabling the model to learn about defect patterns using image-level labels. i) Phase 2 transfers classification knowledge to the U-Net by training with soft supervision maps and pattern-weighted Dice loss. i) Phase 3 further tunes the model using synthetic mixed-defect data, improving precision with pattern-weighted Dice loss and absence-penalty scheduling. crowdsourcing for annotation is often infeasible in the semi- conductor domain due to strict confidentiality requirements. (2) Binary-valued pixels. Defect image pixels only have binary values: normal (0) or defective (1). This implies that there are limited features to learn from, as the image only comprises a white foreground region with a lack of visual cues. Without such cues, models must rely on structural patterns and distributional statistics, making the task uniquely challenging. (3)Sparsity of defect patterns. Defect patterns may exhibit pixel- level sparsity (small defective regions within large backgrounds) and image-level sparsity (rare occurrence of certain patterns). As shown in Figure 2, some patterns have extremely sparse appearances, where background (non-defect regions) dominate the image. This indicates pixel-level sparsity. Some patterns may be scarce due to specific causes in manufacturing. This leads to image-level sparsity where data imbalance may occur. Unlike existing baselines that neglect these factors, SePArate explic- itly addresses such domain-specific challenges to achieve effective defect pattern segmentation quality. 4 Methodology In Figure 3, we illustrate the overall training with SePArate. The entire process operates without expensive pixel-level annotations. SePArate uses the image-level labels to first learn basic pattern-level features, then leverages this knowledge for identifying key spa- tial features. Specifically, SePArate trains the model through three phases: 1) Pretraining the encoder for classification (section 4.1), 2) Adapting it to learn spatial cues of defect patterns (section 4.2), and 3) Final training on synthetic mixed-defect data for precise pattern segmentation (section 4.3). We use the U-Net [20] architecture, as it is widely used for segmentation on real-world images. 4.1 Phase 1: Encoder Training Phase 1 initializes the U-Net encoder with basic knowledge required to identify each single-type defect pattern. This ability is later ex- ploited to generate cues that guide the model to spatially localize each defect pattern. In Phase 1, we use the available image-level labels to train the encoder for classification task. For this, an MLP- based classification head is attached after the encoder to predict the presence of each single-type defect in the image. The model is trained with the mean of Binary Cross-Entropy Loss for each label, as defect pattern classification is a multi-label task. 4.2 Phase 2: Spatial Knowledge Transfer In Phase 2, we train the full U-Net architecture (encoder and de- coder). The goal at this stage is to learn spatial indicators of defect patterns. This is done by training the model with soft supervision maps with a pattern-weighted Dice loss. Given the absence of se- mantic information in the images, we generate soft supervision maps to supply pattern-aware hints, stabilizing the model’s early learn- ing process to differentiate defect patterns. The pattern-weighted Dice Loss helps this process by assigning loss weights, ensuring balanced optimization across defect patterns. 4.2.1 Soft Supervision Maps. Soft supervision maps act as inter- mediate pseudo-labels that guide the model toward defect regions. They incorporate global and local spatial cues, each extracted from the image-level classifier, and the defective regions of the image. To construct these maps, we combine local density maps and class activation maps (CAM, [3, 22]) from the pretrained classifier. Dataset for Phase 2 CAMs per defect pattern Filter Real Image Encoder CAM Soft Supervision Maps Real Image Pseudo-labels Stack Local Density Maps Hadamard Product Figure 4: Generation of soft supervision maps. For each im- age, local density map (bottom left) and CAM (top center) are computed, combined to generate soft supervision maps. DAC ’26, July 26–29, 2026, Long Beach, CA, USADain Kwon et al. On the one hand, we generate a local density map of the defect images. Local density maps provide global cues of the image by highlighting regions where defect pixels are concentrated. Defect patterns typically form clusters, so we compute these maps using a simple mean filter that amplifies dense defect regions and smooths out sparse areas. High-value regions in the density map represent probable defect zones, as illustrated in Figure 4. On the other hand, we also utilize CAM, which complements the density maps by identifying regions that influence the classifier’s decision for each defect type. While density alone offers spatial importance, it does not differentiate between patterns. Therefore, CAM introduces type-specific relevance, enabling the supervision to carry pattern-specific information. Finally, we generate the final soft supervision maps by multi- plying each type-specific CAM with the density map, followed by normalization on pixel values. These score-like pixel values rep- resent the joint contribution of defect concentration and pattern- specific signals from the classifier. The U-Net is trained to align its output with these maps, using the pattern-weighted Dice loss which balances training across defect patterns. 4.2.2 Pattern-Weighted Dice Loss. Due to their sparse and mono- tonic regions, defect patterns are difficult for the model to separate in the early stages of training. This often leads the model to default to predicting mostly absent masks for some patterns. To mitigate this issue and encourage the model to first detect coarse regions of patterns, thereby improving early-stage recall, we introduce the pattern-weighted Dice Loss. We use the loss to ensure the model to primarily capture the existing defects and further enable balanced training for all defect patterns, regardless of the available samples of each defect pattern. The original Dice loss [13] is widely used for segmentation tasks: L dice,푖,푝 = 1− 2( Í 푦 푖,푝 · ˆ 푦 푖,푝 )+휖 Í 푦 푖,푝 + Í ˆ 푦 푖,푝 +휖 (1) where ˆ 푦 푖,푝 denotes the predicted mask and푦 푖,푝 denotes the ground truth mask for image푖and single-type pattern푝. We observed that directly using the Dice loss led to a drastic degradation of less frequent and scarce defect patterns. This behavior stems from the loss, pushing the model toward absent-mask predictions for rarely occurring patterns. Therefore, we assign loss weights to every image-pattern pair to help the model focus more on the presence of each pattern, rather than being biased by the absence of patterns. For each image 푖 and pattern 푝, we define the indicator: Let 훿 푖,푝 = ( 1,if pattern 푝 exists in the image 푖, 0,otherwise. (2) Let the dataset-level occurrence count of pattern푝be퐾 푝 = Í 푁 푖=1 훿 푖,푝 where 푁 is the total number of images in the train set. The Pattern-Weighted Dice Loss is then defined as L PWDL = 푃 ∑︁ 푝=1 1 퐾 푝 푁 ∑︁ 푖=1 훿 푖,푝 1− 2( Í 푦 푖,푝 · ˆ 푦 푖,푝 )+휖 Í 푦 푖,푝 + Í ˆ 푦 푖,푝 +휖 ,(3) Equivalently, L PWDL = 푃 ∑︁ 푝=1 1 퐾 푝 +휖 ∑︁ 푖:훿 푖,푝 =1 L dice,푖,푝 (4) The Pattern-Weighted Dice Loss stabilizes learning across pat- terns of different prevalence and improves segmentation on scarce defect patterns. 4.3 Phase 3: Synthetic Defect Training Phase 3 further refines the performance of SePArate with fine- tuning using a synthetic dataset with accurate segmentation masks. 4.3.1 Synthetic Dataset Generation. As discussed in the previous section, because the datasets are binary-valued images, any 1- valued pixels correspond to one of the target patterns. Therefore, for single-type defects, all 1-valued pixels can directly serve as the seg- mentation masks for those patterns. The remaining problem is that this only applies to the single-type defects, and not to the mixed- type defects. To tackle this problem, we synthesize multi-type defect masks by merging single-type defect masks. For example, in the MATDefects dataset, a row-defect map and a column-defect map can simply be added together with slight shifts for augmentation. This creates a synthetic, but realistic mixed-type defect maps, which are shown in Figure 5. Thus, for the training, we use the synthetic dataset’s segmentation masks as the ground truth. 4.3.2Absence Penalty Loss & Scheduling. When training the model in Phase 3, we train the model with the aforementioned Pattern- Weighted Dice Loss and the Absence Penalty (AP) Loss. Here, the Absence Penalty Loss is introduced to reduce the false-positive pixels. Since the absent sample pairs are ignored in Phase 2 by design Equation (3), the trained model has a tendency to include irrelevant pixels, or sometimes show defect patterns that are not present. To address this, we add a loss term that penalizes such a prediction with Binary Cross-Entropy (BCE) Loss. Let ˆ 푦 ′ 푖,푝 (푥)= ( ˆ 푦 푖,푝 (푥),if ˆ 푦 푖,푝 (푥) ≥ 휏 0,otherwise. (5) where푥indexes the location of pixels in ˆ 푦 푖,푝 , and휏is the confi- dence threshold. If the confidence is greater than or equal to휏, it is included in the loss. We gradually reduce the value of휏from 0.5 to 0.0, depending on loss convergence. Define the pixel-wise BCE against an all-zero target as L BCE ˆ 푦 ′ 푖,푝 , 0 =− 1 |Ω| ∑︁ 푥∈Ω log 1− ˆ 푦 ′ 푖,푝 (푥)+휖 ,(6) (Real) Single-Type Defect Ground truth segmentation masks (Synthetic) Mixed-Type Defect Synthetic images D + ER + S Group + R + C Sum Sum ᄑ Col (C)Row (R)Group Scratch (S)EdgeRing (ER)Donut (D) Figure 5: Generation of synthetic mixed-type defect images. SePArate : Segmenting Patterns from Defects in Wafer Manufacturing Using Weak SupervisionDAC ’26, July 26–29, 2026, Long Beach, CA, USA Table 1: Comparison of Mean IoU (mIoU) for Defect Pattern Segmentation (%) Method MATDefectsMixedWM38 ColumnRowGroup MeanCenterDonutEdgeLocEdgeRingLocNearFullScratchRandom Mean CLIPSeg [11]4.204.324.124.215.165.064.944.824.874.875.175.285.02 MCT [26] 42.5350.6031.6641.605.6016.8915.3433.735.7421.434.3682.8623.24 ReCAM [1]37.230.0023.7120.3110.5237.519.6311.568.6963.928.3393.5530.46 S2C [9] 60.5740.1332.4944.401.430.2013.2010.7518.180.008.900.006.58 SSB-Rec [27]55.5950.8944.8150.435.4418.710.8426.5756.3232.9416.4367.5328.10 Ours, SePArate70.9166.6664.6867.4294.1060.5632.2069.2375.3583.3339.2198.3169.04 ∗ MCT: MCTFormer, SSB-Rec: Semantic Segmentation-Based Wafer Map Mixed-Type Defect Pattern Recognition whereΩdenotes the set of all pixels above the threshold in the image for ˆ 푦 ′ 푖,푝 . Then the Absence Penalty loss is L AP = 푃 ∑︁ 푝=1 1 퐾 ′ 푝 +휖 푁 ∑︁ 푖=1 푎 푖,푝 L BCE ˆ 푦 ′ 푖,푝 , 0 ,(7) where 푎 푖,푝 = 1−훿 푖,푝 and 퐾 ′ 푝 = Í 푁 푖=1 1−훿 푖,푝 . Consequently, the total loss is L total = 휆 dice L PWDL + 휆 ap L AP ,(8) where 휆 dice ,휆 ap ≥ 0 are weighting hyperparameters. 5 Experimental Results 5.1 Experimental Settings We use two datasets for evaluation: MATDefects and MixedWM38. Each dataset was randomly split into 80% training, 10% validation, and from the remaining 10%, we manually annotated 1,000 samples and used them as test sets. Model prediction was performed for defective regions in each image. In MixedWM38, all non-defective pixels are unified to 0, resulting in binary images. While class labels were provided, test-set masks were manually annotated at the pixel level. Our framework was implemented with PyTorch [18]. Experiments were run on NVIDIA RTX 4090 GPUs. We chose baselines compatible with our scenario: Segmenta- tion methods that do not require pixel-level annotations (See 2.2): CLIPSeg [11], MCT [26], ReCAM [1], S2C [9] and SSB-Rec [27]. CLIPSeg [11] is a widely used zero-shot segmentation model. S2C [9] generates pixel-level masks refined by SAM [7]. MCT [26] uses class-specific tokens in the transformer to obtain discriminative attention maps. ReCAM [1] introduces a loss function to improve object boundaries in CAM-based pseudo-labels. Lastly, SSB-Rec [27] also leverages the idea of generating a pseudo dataset with single-type defect images, but uses it to train classification models. We modify the method to do a segmentation task, which is trivially done as it already uses U-Net with a segmentation-based loss. 5.2Comparison of Defect Segmentation Results 5.2.1 Performance Comparison. Table 1 shows the performance comparison by evaluating mean Intersection-over-Union (mIoU) results. As shown in table 1, SePArate outperforms baselines in both datasets by a great margin. For the MATDefects dataset, the average mIoU is 67.42%, which shows a superior value to be used in prac- tice. For the MixedWM38 dataset, the average mIoU is 69.04%, also Table 2: Ablation Study of SePArate with MATDefects Phase 1Phase 2Phase 3 mIoU (%) ColumnRowGroup Mean ✓✗0.000.000.610.20 ✓✗42.1842.3233.9839.49 ✓✗✓ 62.3056.540.1439.66 ✓70.9166.6664.6867.42 showing a reliable performance, considering the larger number of classes. On the other hand, the baselines do not show practical per- formance, or even fail to work on some patterns. This degradation occurs because baselines rely on rich visual features that are not present in binary defect images. Therefore, its general knowledge does not help the segmentation of the defect patterns when using only weak visual cues from the binary pixels. Unlike the baselines, only SePArate successfully performs defect pattern segmentation, validating the design and its suitability for practical use. 5.2.2Visualizations. We visualize defect segmentation results for the test images of MATDefects and MixedWM38 in Figure 7. Each example shows the original defect image where multiple pattern types overlap. SePArate successfully disentangles these mixed pat- terns. Pixels of the same pattern share identical colors. The seg- mented regions exhibit clear morphological distinctions that accu- rately represent each defect pattern. This shows the practical value of leveraging SePArate for automated defect analysis, preserving critical pattern and shape information. ColumnRowGroup 50 60 70 80 90 100 Precision (%) Precision per defect pattern Phase 1+3 Phase 1+2+3 ColumnRowGroup 50 60 70 80 90 100 Recall (%) 0.1 Recall per defect pattern Phase 1+3 Phase 1+2+3 Figure 6: Precision and Recall comparison with and without Phase 2 in SePArate, with MATDefects. It shows improved metrics when trained using all 3 phases. DAC ’26, July 26–29, 2026, Long Beach, CA, USADain Kwon et al. Defect Pattern Segmentation in MixedWM38Defect Pattern Segmentation in MATDefects Figure 7: Examples of segmented defect pattern in test images, using our proposed method, SePArate. For proprietary reasons, the examples of MATDefects were slightly modified from the original data. Key patterns were preserved regardless of modification. 5.3 Further Analyses of SePArate 5.3.1Ablation Study of Phases. Table 2 shows the ablation study of SePArate where we evaluate models trained with selective phases. The first row shows that Phase 1 yields almost no segmentation capability, with a mIoU of 0.20%. When adding Phase 2, the mean score increases to 39.49%, indicating that soft supervision helps the model learn spatial cues but remains insufficient. Similarly, using only Phase 3 with Phase 1 increases the mean mIoU further to 39.66%, showing that refinement is effective but still incomplete, as the model fails on "group" patterns. The final setting with all phases corresponds to SePArate, achieving the highest performance with a mean mIoU of 67.42%, validating the overall design. A notable observation is the substantial mIoU gap in "group" pattern segmentation when comparing the third and final rows. Without soft supervision maps (third row), the model fails to iso- late group defects. This confirms the crucial role of Phase 2. As group defects typically co-occur with line-defects (Figure 7), Phase 2 proves essential for disentangling such overlapping patterns. We further analyze the synergistic effects of phases in SePArate, shown in Figure 6 with MATDefects. The results for the model trained with all phases of SePArate are shown in green. The results in red are obtained only with Phase 1 and Phase 3. Notably, the 0.1% recall for the "group" pattern when Phase 2 is excluded suggests that C D EL ER L NF S R 20% 40% 60% 80% 100% C D EL ER L NF S R 20% 40% 60% 80% 100% PWDLPWDL + AP PWDL + AP +-scheduling Figure 8: Loss term ablation for Phase 3 with MixedWM38. The left figure compares mIoU across single-type defect pat- terns, and the right depicts the difference in Precision. the model only captures a very small portion of true defect pixels. The model collapses without an early-stage understanding of spatial context about defect patterns, highlighting Phase 2’s primary role in supplying knowledge of defect regions. By comparing the two, we can see that the overall segmentation is better with Phase 2. 5.3.2Loss Term Analysis. Figure 8 illustrates the improvement in segmentation as additional loss terms and scheduling strategies are added to Phase 3. From left to right, each setting introduces a new component: Pattern-Weighted Dice Loss (L 푃푊퐷퐿 ) only,L 푃푊퐷퐿 + Absence Penalty (L 퐴푃 ), and finallyL 푃푊퐷퐿 +L 퐴푃 +휏-Scheduling. This results in a gradual increase in IoU across all defect patterns. Pattern-Weighted Dice Loss Only serves as the baseline. While it handles class imbalance, it shows limited performance on sparse patterns (e.g., Scratch, EdgeRing), which remain underrepresented. Adding the Absence Penalty term provides negative supervision for pattern-absent regions, reducing false positives. This yields noticeable IoU gains on previously over-predicted patterns (e.g., EdgeRing). Finally,휏-Scheduling further improves the model by suppressing all false-positive pixels during training. As training progresses, decreasing휏makes the model more conservative, re- ducing false positives and improving IoU for nearly all patterns. Overall, the figure demonstrates the cumulative benefit of the three phases, showing the best performance when used altogether. 6 Conclusion In this work, we present SePArate, a pioneering weakly supervised framework that advances automatic semiconductor defect inspec- tion from image-level classification to pattern-level segmentation. Unlike conventional approaches requiring costly pixel-level anno- tations, our method uses only image-level labels to separate defect instances, while remaining scalable and cost-effective. Experiments on defect datasets show that SePArate consistently outperforms baselines. By providing a single type of pattern-level feedback for defects, our framework can enhance optical inspection, long-term monitoring, and continuous manufacturing improvement. Acknowledgments This paper was the result of the research project supported by SK Hynix Inc (MDUS001T). Jinho Lee is the corresponding author. SePArate : Segmenting Patterns from Defects in Wafer Manufacturing Using Weak SupervisionDAC ’26, July 26–29, 2026, Long Beach, CA, USA References [1]Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. 2022. Class re-activation maps for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 969–978. [2]Mehret Getachew, Birhanu Beshah, and Ameha Mulugeta. 2024. Data analyt- ics in zero defect manufacturing: a systematic literature review and proposed framework. International Journal of Production Research (2024), 1–33. [3]Jacob Gildenblat and contributors. 2021. PyTorch library for CAM methods. https://github.com/jacobgil/pytorch-grad-cam. [4]Shaun S Gleason, Kenneth W Tobin Jr, Thomas P Karnowski, and Fred Lakhani. 1998. Rapid yield learning through optical defect and electrical test analysis. In Metrology, Inspection, and Process Control for Microlithography XII, Vol. 3332. SPIE, 232–242. [5]Hongquan He, Guowen Kuang, Qi Sun, and Hao Geng. 2024. PaLM: Point Cloud and Large Pre-trained Model Catch Mixed-type Wafer Defect Pattern Recognition. In 2024 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 1–6. [6]Tongwha Kim and Kamran Behdinan. 2023. Advances in machine learning and deep learning applications towards wafer map defect recognition and classifica- tion: a review. Journal of Intelligent Manufacturing 34, 8 (2023), 3215–3247. [7] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. 2023. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision. 4015–4026. [8]Kibong Koo, Sunghwa Ok, Yonggu Kang, Seungbong Kim, Choungki Song, Hyey- oung Lee, Hyungsoo Kim, Yongmi Kim, Jeonghun Lee, Seunghan Oak, Yosep Lee, Jungyu Lee, Joongho Lee, Hyungyu Lee, Jaemin Jang, Jongho Jung, Byeongchan Choi, Yongju Kim, Youngdo Hur, Yunsaing Kim, Byongtae Chung, and Yongtak Kim. 2012. A 1.2V 38nm 2.4Gb/s/pin 2Gb DDR4 SDRAM with bank group and×4 half-page architecture. In IEEE International Solid-State Circuits Conference. [9] Hyeokjun Kweon and Kuk-Jin Yoon. 2024. From sam to cams: Exploring segment anything model for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 19499– 19509. [10] Kiryong Kyeong and Heeyoung Kim. 2018. Classification of mixed-type defect patterns in wafer bin maps using convolutional neural networks. IEEE Transac- tions on Semiconductor Manufacturing 31, 3 (2018), 395–402. [11] Timo Lüddecke and Alexander Ecker. 2022. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 7086–7096. [12]Anne Meixner. [n. d.]. Too Much Fab And Test Data, Low Utilization. https: //semiengineering.com/too-much-fab-and-test-data-low-utilization/. [13]Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. 2016. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV). Ieee, 565–571. [14]Subhrajit Nag, Dhruv Makwana, Sparsh Mittal, C Krishna Mohan, et al.2022. WaferSegClassNet-A light-weight network for classification and segmentation of semiconductor wafer defects. Computers in Industry 142 (2022), 103720. [15] Takeshi Nakazawa and Deepak V Kulkarni. 2019. Anomaly detection and seg- mentation for wafer defect patterns using deep convolutional encoder–decoder neural network architectures in semiconductor manufacturing. IEEE Transactions on Semiconductor Manufacturing 32, 2 (2019), 250–256. [16] Mu Nie, Shidong Zhu, Aibin Yan, Cheng Zhuo, Xiaoqing Wen, and Tianming Ni. 2025. Efficient Modulated State Space Model for Mixed-Type Wafer Defect Pattern Recognition. In 2025 Design, Automation & Test in Europe Conference (DATE). IEEE, 1–6. [17] Upasana Pandya. [n. d.]. Modernizing Semiconductor Manufacturing Analytics Platform using AWS Data Analytics Services. https://designthesolution.org/wp- content/uploads/2024/10/process-track-upasana-modernizing-semiconductor- manufacturing-analytics-aws-data-analytics-services.pdf . [18]Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al.2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019). [19]Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al.2024. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 (2024). [20]Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention. Springer, 234–241. [21]Muhammad Saqlain, Qasim Abbas, and Jong Yun Lee. 2020. A deep convolu- tional neural network for wafer defect identification on an imbalanced dataset in semiconductor manufacturing processes. IEEE Transactions on Semiconductor Manufacturing 33, 3 (2020), 436–444. [22]Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedan- tam, Devi Parikh, and Dhruv Batra. 2017. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE interna- tional conference on computer vision. 618–626. [23]NG Shankar and ZW Zhong. 2005. Defect detection on semiconductor wafer surfaces. Microelectronic engineering 77, 3-4 (2005), 337–346. [24] Ho Sun Shon, Erdenebileg Batbaatar, Wan-Sup Cho, and Seong Gon Choi. 2021. Unsupervised pre-training of imbalanced data for identification of wafer map defect patterns. IEEE Access 9 (2021), 52352–52363. [25]Junliang Wang, Chuqiao Xu, Zhengliang Yang, Jie Zhang, and Xiaoou Li. 2020. Deformable convolutional networks for efficient mixed-type wafer defect pattern recognition. IEEE Transactions on Semiconductor Manufacturing 33, 4 (2020), 587–596. [26]Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. 2022. Multi-class token transformer for weakly supervised semantic segmenta- tion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 4310–4319. [27] Jinda Yan, Yi Sheng, and Minghao Piao. 2023. Semantic segmentation-based wafer map mixed-type defect pattern recognition. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 42, 11 (2023), 4065–4074. [28]Jianbo Yu. 2019. Enhanced stacked denoising autoencoder-based feature learn- ing for recognition of wafer map defects. IEEE Transactions on Semiconductor Manufacturing 32, 4 (2019), 613–624.