Domain Adaptation for Image Classification of Defects in Semiconductor Manufacturing
Adrian Poniatowski, Natalie Gentner, Manuel Barusco, Davide Dalle Pezze, Samuele Salti, Gian Antonio Susto · 2025
Read Domain Adaptation for Image Classification of Defects in Semiconductor Manufacturing by Adrian Poniatowski, Natalie Gentner, Manuel Barusco, Davide Dalle Pezze, Samuele Salti, Gian Antonio Susto (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-15260.txt.
Introduction
Adrian Poniatowski and co-authors help defect-spotting models survive the move between chip factories. Their 2025 paper applies domain adaptation to classifying flaws in electron microscope images of semiconductors. A model trained in one plant's conditions degrades in another's, and relabeling thousands of images by hand defeats the point of automation.
The team benchmarks adaptation techniques in unsupervised and semi-supervised settings, where the target plant provides few or no labels. Their own method, DBACS, builds on an image-translation architecture with extra loss terms to sharpen the transfer. Real production images validate the comparison, and their approach advances the state of the art for this setting. Time to market in chips is measured in weeks of yield learning; a classifier that arrives pre-adapted is worth money. The authors frame adaptation not as a research luxury but as the deployment requirement the industry actually faces.
Catalog record
| CREATOR | Adrian Poniatowski, Natalie Gentner, Manuel Barusco, Davide Dalle Pezze, Samuele Salti, Gian Antonio Susto |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 54 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-15260
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Domain Adaptation for Image Classification of Defects in Semiconductor Manufacturing Adrian Poniatowski, Natalie Gentner, Manuel Barusco, Davide Dalle Pezze, Samuele Salti, Gian Antonio Susto Abstract In the semiconductor sector, due to high demand but also strong and increasing competition, time to market and quality are key factors in securing significant market share in various application areas. Thanks to the success of deep learning methods in recent years in the computer vision domain, Industry 4.0 and 5.0 applications, such as defect classification, have achieved remarkable success. In particular, Domain Adaptation (DA) has proven highly effective since it focuses on using the knowledge learned on a (source) domain to adapt and perform effectively on a different but related (target) domain. By improving robustness and scalability, DA minimizes the need for extensive manual re-labeling or re-training of models. This not only reduces computational and resource costs but also allows human experts to focus on high-value tasks. Therefore, we tested the efficacy of DA techniques in semi-supervised and unsupervised settings within the context of the semiconductor field. Moreover, we propose the DBACS approach, a CycleGAN-inspired model enhanced with additional loss terms to improve performance. All the approaches are studied and validated on real-world Electron Microscope images considering the unsupervised and semi-supervised settings, proving the usefulness of our method in advancing DA techniques for the semiconductor field. K eywords Computer Vision \cdot Convolutional Neural Network \cdot Deep Learning \cdot Defect Classification \cdot Domain Adaptation \cdot Industry 4.0 \cdot Pseudo Labeling \cdot Scanning Electron Microscope Images 1 Introduction Semiconductors are constructed by linking circuit structures on various layers built upon a thin slice of semiconductor material called a wafer. The creation of those layers, including their complex individual structures and patterns, is done in a sequential manner through a combination of the steps. The repetition of process steps and re-entry into the production flow is part of the advanced and highly specialized fabrication process. Metrology is a set of quality assurance steps carried out in production to control and calibrate machines, process settings, process outcomes, and wafer properties. Within this framework, Defect Detection and Classification is a fundamental part of metrology to ensure high product quality and production performance. In this context, the use of Scanning Electron Microscope (SEM) images of the wafer can be used (see Figure 1 and 3 ). The collected data is examined and divided into different defect types before corrective actions or the declaration of ’defective’ products are triggered (the entire inspection procedure is depicted in Figure 1 ). Figure 1: Defect classification procedure for the semiconductor manufacturing process at hand: (i) an optical inspection tool locates defects (marked in red) by comparing multiple neighboring areas on a wafer and looking for anomalies; (ii) high-resolution SEM images are taken at each defect location. (iii) defects are classified: differences in the image background structure are visible, typically because images are taken in different positions on the wafer, or more generally, different wafers belong to different technologies, hence showing variations in their surface structure. Industry 4.0, which is enabled by technologies such as Deep Learning, enabled automatic defect classification which has become cost-effective and widely adopted [ 1 ] . However, a series of challenges hinder the scalability of these models, such as the differences in data distributions combined with the abundance of data not labeled [ 2 ] . Domain Adaptation (DA) is extremely fitted for Industry 5.0 applications which focus on human-centric approaches, sustainability, and resilience. Indeed, DA focuses on using the knowledge learned on a (source) domain to adapt and perform effectively on a different, but related (target) domain. DA minimizes the need for extensive manual re-labeling or re-training of models, allowing human experts to focus on high-value tasks and improving robustness and scalability. Specifically, in the context of the semiconductor field, to enable fab-wide usage of highly automated classification models, deployed methods need to be transferable and applicable when confronted with new optical properties and different image structures generated from technological diversity as well as reoccurring production steps in different stages of the manufacturing process. Therefore, in this work, we propose to study the effectiveness of DA techniques in the context of the semiconductor domain. In particular, we focus on state-of-the-art models able to perform in both the semi-supervised domain adaptation (SSDA) and the unsupervised domain adaptation (UDA) settings. We focus on the AdaMatch [ 3 ] approach for the pseudo-labeling group and DBACS [ 4 ] as a generative and adversarial model. While both have been proven to be of extreme applicability, they have never been applied so far in the context of image classification problems in semiconductor manufacturing. In particular, in this work, we consider an evolution of the DANN approach [ 5 ] called DANN-based Alignment with Cyclic Supervision (DBACS) [ 4 ] . This method was only considered for time-series data and it is studied for the first time in the image domain. While this is a CycleGAN-inspired model, we enhance and adapt the original version for the computer vision by adding the following modifications: cyclic adaptation, unpaired sample mapping and feature matching loss, identity loss and multi-scale structure similarity loss. These modifications improved the final performance. The main contributions of this work are summarized as follows: • We test and study for the first time the use of DA techniques to solve the defect classification problem in the semiconductor manufacturing field for the SEM images. • We introduce the use of the DBACS approach in the Computer Vision domain by adapting its architecture and by considering additional loss terms. • The effectiveness of our approach is validated on real-world Electron Microscope images considering the unsupervised and semi-supervised settings, proving also the usability and usefulness of DA techniques for the SEM images. Furthermore, we have made all the code used in our study, including the DBCAS implementation, publicly accessible 1 1 1 https://bitbucket.org/papers_vad_group/dbacs/src/main/ . This allows researchers and practitioners to utilize our implementation for further advancements. The rest of the paper is organized as follows: In Section 2 , we review the literature concerning the semiconductor sector and domain adaptation field. Section 4 defines the problem formally and presents our approach. Then Section 5 describes the methods used as comparison. Section 6 provides detailed information about the data used, the experimental setup, including the considered models, and the specific hyperparameters employed for each DA technique. Section 7 shows and discusses the obtained results. Finally, Section 8 summarizes our findings and future research directions are presented, 2 Related Work We delineate the relevant literature as follows. In Section 2.1 , an overview of previous works to solve the image classification problem in semiconductor manufacturing is provided. Then we discuss Domain Adaptation with a focus on the two main families, Pseudo-labeling in Section 2.2.1 and Generative Adversarial Models in Section 2.2.2 Task Labeled Unlabeled Distributions SSL source target source = target UDA source target source \neq target SSDA source + target target source \neq target Table 1: Comparison of different learning paradigm: Self-Supervised Learning (SSL), Unsupervised Domain Adaptation (UDA), Semi-Supervised Domain Adaptation (SSDA). 2.1 Image Classification in Semiconductor Manufacturing Image classification in semiconductor manufacturing is mainly applied for two different tasks: image defect classification and wafer-level/wafer map