GAN Synthetic Defect Data Generation — PatSnap Eureka
GAN Synthetic Defect Data Generation 2026
Synthetic defect data generation using GANs addresses the chronic scarcity of labeled defect samples in industrial AI. This landscape covers core architectures, application domains, and the 2019–2026 patent record.
How GANs Solve the Defect Data Scarcity Problem
GAN-based synthetic defect data generation sits at the intersection of deep generative modeling, computer vision, and industrial quality assurance. Defective products are rare by design in high-quality manufacturing lines, creating severe class imbalance that degrades supervised deep learning classifiers. GANs address this by learning the underlying distribution of defect appearances and generating novel, photorealistic defect samples for downstream model training.
Four core technical clusters define the field: image-to-image translation for defect injection, conditional GANs enabling attribute-controlled synthesis, physics-coupled DCGAN pipelines combining simulation with GAN style transfer, and semiconductor-specific architectures trained on design data. A fifth emerging direction—transformer-based generative pre-trained models—signals a transition beyond classical GAN architectures in the most recent 2025–2026 filings.
The earliest records in this dataset date to 2019, with Applied Materials Israel’s foundational WO filing on training-set generation for semiconductor specimen examination and IBM’s coordinate-based GAN for layout pattern synthesis. The 2020–2022 window represents the field’s foundational industrialization phase, with at least 15 distinct patents or literature references identified across US, EP, CN, and WO jurisdictions.
Among retrieved records, KLA Corporation, Applied Materials Israel, and IBM collectively hold the largest share of active semiconductor and layout-domain patents in this dataset. Industrial surface inspection is more fragmented, with Chinese universities and equipment makers accounting for approximately 25% of records in retrieved records.
Filing Trends and Technology Cluster Distribution
Retrieved records span 2019–2026 across US, CN, WO, EP, IN, and SG jurisdictions. The dataset shows accelerating specialization from 2023 onward and a clear 2025–2026 pivot toward transformer and diffusion architectures.
Patent Filings by Technology Cluster — GAN Defect Synthesis (Dataset Snapshot)
Image-to-image translation and semiconductor-specific GAN clusters account for the largest share of retrieved records in this dataset, reflecting the dominance of wafer inspection and industrial surface defect use cases.
↗ Click bars to exploreFiling Activity by Period — GAN Defect Synthesis Dataset (2019–2026)
The 2025–2026 period shows the highest concentration of frontier filings in this dataset, driven by transformer, diffusion model, and 3D-engine-grounded synthesis approaches from NVIDIA, KLA, IMEC, and Applied Materials Israel.
↗ Click bars to exploreKey Deployment Domains for GAN Defect Synthesis
Retrieved records span four major application domains: industrial surface inspection, semiconductor wafer inspection, non-destructive testing, and medical imaging. Each domain presents distinct data scarcity profiles and generative architecture requirements.
Semiconductor Wafer Inspection
The most patent-dense specialised sub-domain in this dataset, with KLA Corporation, ASML Netherlands, and Applied Materials Israel holding the majority of semiconductor-specific filings in retrieved records. ASML’s US grant (2024) covers generators conditioned on defect attribute combinations including type, size, location, and strength applied to defect-free inspection images. Applied Materials Israel’s 2026 US filings describe systems generating millions of synthetic fault images at runtime during active wafer fabrication.
Semiconductor InspectionIndustrial Surface Defect Inspection
The most represented application domain across the dataset, addressing scratches, cracks, dents, and texture anomalies on metal, fabric, and composite surfaces. Shanxi Anshu’s DCGAN pipeline achieves synthesis at approximately 0.15 seconds per image on CPU. Shanghai Zhijing’s mask-guided pix2pix GAN targets grey cloth (raw fabric) defect texture synthesis, while Hefei University of Technology’s 2026 CN filing combines synthetic defect data with adversarial domain adaptation for automotive body surface inspection.
Manufacturing Quality ControlNon-Destructive Testing (NDT)
Indian Institute of Technology Madras filed across WO (2022), IN, and US (2024) jurisdictions for a system where a numerical simulation model generates NDT datasets with flaw geometrical features, which then train a DCGAN to produce diverse synthetic NDT inspection data covering ultrasonic, radiographic, and eddy-current modalities. This pipeline directly addresses the scarcity of labeled flaw samples in critical infrastructure inspection workflows.
Non-Destructive TestingMedical Imaging and Surgical AI
Siemens Healthineers AG (US, 2021) filed for GAN-generated synthetic image data training deep learning algorithms in tumor lesion characterization. Digital Surgery Limited (WO, 2024) describes GAN-based synthetic surgical data generation combining with real surgical video to improve training datasets for surgical AI. Nitte University (IN, 2024) filed a mobile-device-based GAN system using pixel-wise and latent space comparison for automated defect identification analogous to lesion detection.
Medical ImagingKey Patent Assignees in GAN Defect Synthesis (Retrieved Records)
In this dataset, KLA Corporation and Applied Materials Israel hold the deepest prosecution portfolios in the semiconductor-focused GAN defect synthesis space, with 5+ and 4+ retrieved records respectively. IBM accounts for 5 records spanning physical design layout GAN and feature-combination synthesis in retrieved records, while ASML Netherlands and IIT Madras each contribute 3–4 records in wafer and NDT domains.
Top Assignees by Filing Count — GAN Defect Synthesis (Dataset Snapshot)
↗ Click bars to exploreKLA Corporation
KLA holds 5+ retrieved records spanning US and WO jurisdictions, with active grants dating from 2021 and pending applications through 2025. Key patents include conditional GANs trained on design data-to-image pairs for wafer specimen simulation (US, 2021; US, 2024) and the DefectGPT transformer-based defect generative pre-trained model filed in both US and WO in April 2025, marking a pivot beyond classical GAN architectures. Patents include both active grants and pending applications targeting the semiconductor wafer inspection domain.
United StatesApplied Materials Israel Ltd.
Applied Materials Israel holds 4+ retrieved records across WO (2019), US (2020, 2026) jurisdictions, with the foundational WO filing establishing the framework for generating training sets for semiconductor specimen examination. The 2026 US filings cover automated generation and planting of synthetic defects-of-interest (DOIs) enabling millions of synthetic fault images at runtime during wafer fabrication, and a height-map-based 3D defect injection method followed by ML model-based image synthesis. Status includes active grants and pending US applications.
Israel — IL / United StatesFour Frontier Vectors in GAN Defect Synthesis (2025–2026)
Approximately 15 records published in 2025–2026 in this dataset signal four major shifts: transformer and diffusion displacement of classical GANs, text-conditioned self-supervised generation, 3D-model-grounded photorealistic synthesis, and runtime on-the-fly defect synthesis at scale.
Transformer and Diffusion Models Displacing Classical GANs
KLA Corporation’s DefectGPT (US/WO, 2025) marks the first clear pivot to transformer-based architectures in the semiconductor inspection domain, aligning synthetic defect images with process condition and wafer-level defect distribution. Simultaneously, IMEC VZW’s US filing (2026) applies denoising diffusion probabilistic models (DDPMs) to semiconductor defect image generation. Shenzhen LingYun Vishion Technology Co., Ltd.’s CN filing (2025) uses fine-tuned diffusion models per defect class alongside annotation synthesizers, confirming broad diffusion adoption across both Western and Chinese assignees.
Text-Conditioned and Self-Supervised Defect Generation
SAP SE’s active US and EP grants (2025) for a self-supervised defect generator (SDDG) integrate text embeddings as conditioning inputs, enabling generation of minority-class defect images with minimal or no defect examples. IBM’s 2021 US patent for text-encoded feature combination conditioning enables generation of defect images for previously unseen feature combinations. These filings together signal convergence of large language model embeddings with defect synthesis pipelines, creating a capability gap for organizations reliant on purely visual conditioning.
Classical GAN vs. Transformer/Diffusion Architectures for Defect Synthesis
Click any row to explore further.
| Dimension | Classical GAN Architectures | Transformer / Diffusion Architectures |
|---|---|---|
| Representative Filings | KLA GAN specimen simulation (US, 2021); ASML conditional wafer defect GAN (US, 2024); Delta Electronics cycle GAN (SG, 2022) | KLA DefectGPT transformer (US/WO, 2025); IMEC VZW DDPM for advanced node (US, 2026); Shenzhen LingYun diffusion per-class (CN, 2025) |
| Core Mechanism | Generator-discriminator adversarial training; image-to-image translation or conditional attribute injection | Transformer autoregressive token generation (DefectGPT) or iterative denoising via DDPM for image synthesis |
| Conditioning Inputs | Design data, defect attribute vectors (type, size, location, strength), spatial masks, Poisson blending | Process condition metadata, wafer-level defect distribution, text embeddings (SAP SE SDDG), per-class fine-tuning |
| Training Data Requirement | Requires paired or unpaired defect/non-defect image sets; class imbalance addressed via augmentation loops | SAP SE SDDG targets minority class generation with minimal defect examples; diffusion models leverage pre-training |
| Synthesis Speed | Shanxi Anshu DCGAN achieves ~0.15 seconds per image on CPU; Applied Materials Israel targets millions of images at runtime | Diffusion models typically slower at inference; transformer models’ runtime speed not specified in retrieved records |
| Jurisdiction Activity | Dominant in US, CN, WO, EP, IN, SG filing records 2019–2024 | Emerging primarily in US and WO records 2025–2026; CN diffusion filings appearing from Shenzhen assignees |
| 3D Integration | Changzhou Weiyi (2023): Blender + pix2pixHD; Raytheon (2024): 3D engine for ATR dataset | NVIDIA (2026): 3D model-based system with configurable lighting, environment, and severity levels |
| Patent Status (dataset) | Mix of active grants (KLA US 2021, ASML US 2024) and pending applications across jurisdictions | Predominantly pending applications (KLA DefectGPT US/WO 2025, IMEC US 2026, Applied Materials Israel US 2026) |
Frequently Asked Questions: GAN Synthetic Defect Data Generation
Defective products are rare by design in high-quality manufacturing lines, creating severe class imbalance that degrades the performance of supervised deep learning classifiers. GANs address this by learning the underlying distribution of defect appearances and generating novel, photorealistic defect samples for downstream model training, according to retrieved records.
In this dataset, KLA Corporation and International Business Machines Corporation each have approximately 5 retrieved records, Applied Materials Israel Ltd. has 4+ records, Indian Institute of Technology Madras has 4 records, and ASML Netherlands B.V. has 3 records. These counts are based on retrieved records only and do not represent total industry filing volumes.
The four core clusters are: (1) image-to-image translation approaches injecting defects into defect-free backgrounds; (2) conditional GAN architectures enabling attribute-controlled defect synthesis; (3) DCGAN variants trained on simulation or NDT data, sometimes coupled with physics-based rendering; and (4) semiconductor-specific GAN architectures leveraging design data and process condition metadata. A fifth emerging direction covers transformer-based and diffusion model approaches.
KLA Corporation’s DefectGPT, filed in US and WO jurisdictions in April 2025, is described in retrieved records as a transformer-based defect generative pre-trained model that aligns synthetic defect images with process condition and wafer-level defect distribution. It marks the first clear pivot to transformer-based architectures in the semiconductor inspection domain within this dataset, moving beyond classical GAN architectures.
Applied Materials Israel’s 2026 US filing describes systems capable of generating millions of synthetic fault images at runtime during active semiconductor fabrication, enabling continuous adaptive training of inspection models without stopping production. Earlier filings in this dataset focused on batch-mode training data generation rather than production-integrated real-time synthesis.
According to the retrieved records, KLA Corporation and Applied Materials Israel have built overlapping, deep prosecution portfolios targeting conditional GAN plus design-data input plus wafer specimen simulation in the US jurisdiction. New entrants are advised to conduct careful freedom-to-operate (FTO) analysis before commercializing wafer-oriented GAN synthesis systems, as noted in the strategic implications section of this dataset.
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.