Contrastive Learning for Defect Classification 2026
Contrastive Learning for Defect Classification
Contrastive learning is addressing the labeled-data bottleneck in industrial defect inspection. This dataset spans 19 patent records and 22 literature records across US, CN, WO, TW, KR, and AU jurisdictions from 1999 to 2026.
Self-Supervised Defect Detection: From Academia to Production IP
Contrastive learning for defect classification sits at the intersection of self-supervised representation learning and automated optical inspection (AOI). The field encompasses contrastive loss-based feature embedding, few-shot and low-data classification via momentum encoders, semi-supervised teacher-student frameworks with feature-contrast regularization, and knowledge distillation architectures that reward dissimilarity on defective samples.
The foundational challenge driving the field — confirmed across multiple records — is the scarcity of labeled defect data in real production environments. As noted in the literature on momentum contrastive learning for wafer defect classification, manual labeling is time-consuming and labeled data is insufficient for standard supervised approaches.
The innovation timeline spans from foundational ensemble and confidence-gated classification concepts (NEOPATH, INC., 1999) through KLA-Tencor’s deep learning maturation phase (2016–2020), into the explicit contrastive learning emergence period (2021–2023) marked by momentum encoders, aligned contrastive loss, and semi-supervised feature-contrast architectures.
In this dataset, the most recent frontier filings (2024–2026) originate from ASML Netherlands B.V. (WO 2025), KLA Corporation (US and WO 2024–2025), IMEC VZW (US and WO 2025), SAP SE (US 2026), and Juhaokan Technology Co., Ltd. (CN 2026), reflecting a broadening assignee base in retrieved records beyond the historically dominant KLA portfolio.
Filing Trends and Technology Cluster Distribution
Analysis of the 19 patent records and 22 literature records in this dataset reveals a concentration in semiconductor wafer inspection and an acceleration of frontier filings from 2024 to 2026 across multiple jurisdictions.
Patent Records by Technology Cluster (This Dataset)
Semiconductor wafer inspection accounts for at least 18 of 19 patent records in this dataset, reflecting the dominance of KLA Corporation and ASML in the retrieved records.
↗ Click bars to exploreFrontier Filings by Year: 2021–2026 (Retrieved Records)
Filings in retrieved records accelerated sharply from 2024 to 2026, with ASML, KLA, IMEC, SAP SE, and Juhaokan Technology all contributing records in this dataset within that window.
↗ Click bars to exploreWhere Contrastive Learning for Defect Classification Is Being Applied
The dataset spans four principal application domains: semiconductor wafer and IC inspection, display panel manufacturing, industrial component and surface defect detection, and fault detection in rotating machinery and industrial assets.
Semiconductor Wafer Inspection
This is the dominant application domain in this dataset, with at least 18 patent records targeting semiconductor specimens including wafer bin map (WBM) classification and SEM image analysis. The 2023 momentum contrastive learning literature record explicitly benchmarks WBM defect classification under low labeled-data conditions. ASML’s 2025 WO patent references IC inspection with electron beam microscopes, while IMEC VZW’s federated RL approach enables distributed training across multiple fab sites.
Semiconductor InspectionDisplay Panel Manufacturing
Samsung Display Co., Ltd. filed two records (US and TW, 2023) covering a system and method for defect detection in display manufacturing where labeled defect data is scarce, employing a cost function that rewards similarity on normal images and dissimilarity on defect-introduced images. NEOPATH, INC. filings from 1999 also explicitly named liquid crystal display (LCD) defect classification alongside semiconductor wafers as target applications.
Display ManufacturingIndustrial Components and Surfaces
South China Agricultural University (CN, 2025) filed a two-stage contrastive learning patent for industrial component defect detection combining contrastive learning with adversarial fine-tuning. Shanghai Jiao Tong University’s semi-supervised feature-contrast filings also target general industrial components. Literature records confirm application to steel surface defects (Severstal benchmark), fabric defect detection, and wear debris classification for industrial condition monitoring.
Industrial ComponentsFault Detection in Industrial Assets
A 2021 literature record on contrastive learning for fault detection and diagnostics extends the methodology to rotating machinery and complex safety-critical industrial assets under changing operating conditions and novel fault types. This represents a distinct use case beyond visual defect classification, applying contrastive feature learning to sensor-based condition monitoring scenarios not covered by optical inspection patents in this dataset.
Condition MonitoringKey Patent Assignees in Contrastive Learning Defect Classification (Retrieved Records)
In this dataset, KLA Corporation / KLA-Tencor Corporation holds at least 20 distinct patent records spanning US, WO, TW, CN, SG, and DE jurisdictions — the largest single filing concentration in retrieved records. ASML Netherlands B.V. represents the most significant recent entrant with its 2025 WO contrastive deep learning patent carrying EP priority dates from 2023 and 2024.
Top Assignees by Filing Count — Dataset Snapshot (Retrieved Records)
↗ Click bars to exploreKLA Corporation / KLA-Tencor
KLA Corporation holds at least 20 distinct patent records in this dataset spanning US, WO, TW, CN, SG, and DE jurisdictions, the largest filing concentration in retrieved records. Its portfolio covers ensemble learning classifiers (2016–2019), adaptive automatic defect classification (2017), active learning for classifier training (2019–2020), and ensemble deep learning with pseudo-loss calibration converging to approximately 0.5 (US and WO, 2024–2025). KLA operates as both the algorithmic innovator and the principal filing entity across all major jurisdictions in this dataset.
United StatesASML Netherlands B.V.
ASML Netherlands B.V. filed one WO record in this dataset — Contrastive deep learning for defect inspection (WO, 2025) — with EP priority dates tracing back to November 2023 and August 2024, representing the first dedicated contrastive deep learning patent from a major lithography OEM in retrieved records. The patent explicitly claims a contrastive deep learning methodology for IC inspection referencing electron beam microscopes. This filing signals ASML’s deliberate IP strategy in contrastive methods and a direct competitive signal against KLA in next-generation inspection workflows.
Netherlands — NL / WOFive Frontier Signals in Contrastive Defect Classification (2024–2026)
Among the most recent filings in this dataset (2024–2026), five directional signals emerge spanning OEM-level contrastive IP, federated multi-site learning, semantic auto-labeling, incremental class learning, and model compression with contrastive feature distance loss.
OEM-Level Contrastive IP: ASML’s WO 2025 Entry
ASML Netherlands B.V.’s Contrastive deep learning for defect inspection (WO, 2025) carries EP priority dates from November 2023 and August 2024, representing the first OEM-level contrastive learning patent for lithography-adjacent inspection in this dataset. This signals that contrastive methods are transitioning from academic prototypes to production-grade IP at major semiconductor equipment manufacturers. R&D teams at both ASML and KLA should map freedom-to-operate boundaries around contrastive loss functions applied to SEM and optical inspection imagery.
Federated Learning for Multi-Site Defect Classification
IMEC VZW’s Reinforcement Learning (RL) Based Federated Automated Defect Classification and Detection (US and WO, 2025) introduces a consensus/voting-based federated global model trained from local SEM-based ML models without sharing raw training data. With 3 records in this dataset, IMEC VZW is the only assignee filing explicitly on federated defect model training in retrieved records. This architecture directly addresses data privacy constraints across multi-fab semiconductor environments — a structural industry need that no major inspection OEM has yet claimed at the patent level in this dataset.
Contrastive Pre-Training vs. Semi-Supervised Feature Contrast: Approach Comparison
Click any row to explore further.
| Dimension | Contrastive Pre-Training (Momentum Encoder) | Semi-supervised Feature Contrast (Teacher-Student) |
|---|---|---|
| Primary Mechanism | Momentum contrastive pre-training on large-scale unlabeled data; prototypical network fine-tuning with minimal labeled samples | High-confidence pseudo-labels forwarded for supervised training; low-confidence pixels optimized via contrastive feature learning in the student network |
| Key Patent / Literature Anchor | Momentum Contrastive Learning Framework for Low-Data Wafer Defect Classification (Literature, 2023); ASML WO 2025 contrastive deep learning patent | Semi-supervised industrial defect detection based on feature contrast — Shanghai Jiao Tong University (CN, 2022 and 2025) |
| Application Domain | Semiconductor wafer bin map (WBM) classification; IC inspection with electron beam microscopes | General industrial components; agricultural processing lines; display manufacturing |
| Labeled Data Requirement | Minimal labeled data required for fine-tuning; pre-training uses large-scale unlabeled defect images | Both labeled and unlabeled data used simultaneously in a unified pipeline; pseudo-labels generated iteratively |
| Key Assignees in This Dataset | ASML Netherlands B.V. (NL/WO); KLA Corporation (US) — momentum and ensemble approaches | Shanghai Jiao Tong University (CN); Samsung Display Co., Ltd. (KR/TW/US) |
| Filing Jurisdiction Concentration | WO (ASML 2025); US and WO (KLA 2024–2025); Literature (2022–2023) | CN-primary (Shanghai Jiao Tong University 2022, 2025); US and TW (Samsung Display 2023) |
| Earliest Record in Dataset | 2022 (wear debris few-shot contrastive learning literature); 2023 (wafer WBM momentum contrastive) | 2022 (Shanghai Jiao Tong University CN filing) |
| Uncertainty Handling | Prototypical network fine-tuning provides classification confidence; ensemble pseudo-loss calibration (KLA) addresses runtime uncertainty | Unreliable pseudo-label pixels separated from high-confidence labels and regularized via contrastive feature optimization |
Frequently Asked Questions: Contrastive Learning for Defect Classification
The primary challenge is the scarcity of labeled defect data in real production environments. As noted in the literature on momentum contrastive learning for wafer defect classification in this dataset, manual labeling is time-consuming and labeled data is insufficient for standard supervised approaches.
KLA Corporation / KLA-Tencor Corporation holds at least 20 distinct patent records in this dataset, spanning US, WO, TW, CN, SG, and DE jurisdictions, covering ensemble learning classifiers (2016–2019), active learning, and ensemble deep learning with pseudo-loss calibration (2024–2025).
ASML’s Contrastive deep learning for defect inspection (WO, 2025) directly claims a contrastive deep learning methodology for IC inspection, referencing electron beam microscopes, with EP priority dates tracing back to November 2023 and August 2024. It is described as the first dedicated contrastive deep learning patent from a major lithography OEM in this dataset.
IMEC VZW’s Reinforcement Learning (RL) Based Federated Automated Defect Classification and Detection (US and WO, 2025) introduces a consensus/voting-based federated global model trained from local SEM-based ML models without sharing raw training data, addressing data privacy constraints across multiple fab sites. IMEC is the only assignee in this dataset filing explicitly on federated defect model training.
KLA Corporation’s ensemble deep learning patents (US and WO, 2024) train an ensemble by adjusting parameters until a pseudo-loss function is approximately equal to but not greater than 0.5, directly addressing calibrated uncertainty for runtime defect labeling in high-volume manufacturing workflows.
US leads in filing volume with approximately 25+ US-jurisdiction records in this dataset, followed by CN (8–9 records), WO (7–8 records), and TW (6–7 records), with minor presence in SG, AU, DE, and IN. WO filings from KLA, IMEC, and ASML indicate global protection intent for the most recent innovations.
Data and insights on this page are based on a limited patent and literature dataset and are for reference only. Figures may not represent the complete technology landscape.