eprintid: 14283 rev_number: 8 eprint_status: archive userid: 2 dir: disk0/00/01/42/83 datestamp: 2024-09-19 23:30:14 lastmod: 2024-09-19 23:30:15 status_changed: 2024-09-19 23:30:14 type: article metadata_visibility: show creators_name: Shahroz, Mobeen creators_name: Ali, Mudasir creators_name: Tahir, Alishba creators_name: Fabian Gongora, Henry creators_name: Uc Ríos, Carlos Eduardo creators_name: Abdus Samad, Md creators_name: Ashraf, Imran creators_id: creators_id: creators_id: creators_id: henry.gongora@uneatlantico.es creators_id: carlos.uc@unini.edu.mx creators_id: creators_id: title: Hierarchical Attention Module-Based Hotspot Detection in Wafer Fabrication Using Convolutional Neural Network Model ispublished: pub subjects: uneat_eng divisions: uneatlantico_produccion_cientifica divisions: uninimx_produccion_cientifica divisions: uninipr_produccion_cientifica divisions: uniromana_produccion_cientifica full_text_status: public keywords: Wafer hotspot detection, hierarchical attention module; autoencoder; data augmentation; hybrid attention module; deep learning; image classification; convolutional neural networks abstract: Wafer mappings (WM) help diagnose low-yield issues in semiconductor production by offering vital information about process anomalies. As integrated circuits continue to grow in complexity, doing efficient yield analyses is becoming more essential but also more difficult. Semiconductor manufacturers require constant attention to reliability and efficiency. Using the capabilities of convolutional neural network (CNN) models improved by hierarchical attention module (HAM), wafer hotspot detection is achieved throughout the fabrication process. In an effort to achieve accurate hotspot detection, this study examines a variety of model combinations, including CNN, CNN+long short-term memory (LSTM) LSTM, CNN+Autoencoder, CNN+artificial neural network (ANN), LSTM+HAM, Autoencoder+HAM, ANN+HAM, and CNN+HAM. Data augmentation strategies are utilized to enhance the model’s resilience by optimizing its performance on a variety of datasets. Experimental results indicate a superior performance of 94.58% accuracy using the CNN+HAM model. K-fold cross-validation results using 3, 5, 7, and 10 folds indicate mean accuracy of 94.66%, 94.67%, 94.66%, and 94.66%, for the proposed approach, respectively. The proposed model performs better than recent existing works on wafer hotspot detection. Performance comparison with existing models further validates its robustness and performance. date: 2024-07 publication: IEEE Access volume: 12 pagerange: 92840-92855 id_number: doi:10.1109/ACCESS.2024.3422616 refereed: TRUE issn: 2169-3536 official_url: http://doi.org/10.1109/ACCESS.2024.3422616 access: open language: en citation: Artículo Materias > Ingeniería Universidad Europea del Atlántico > Investigación > Producción Científica Universidad Internacional Iberoamericana México > Investigación > Producción Científica Universidad Internacional Iberoamericana Puerto Rico > Investigación > Artículos y libros Universidad de La Romana > Investigación > Producción Científica Abierto Inglés Wafer mappings (WM) help diagnose low-yield issues in semiconductor production by offering vital information about process anomalies. As integrated circuits continue to grow in complexity, doing efficient yield analyses is becoming more essential but also more difficult. Semiconductor manufacturers require constant attention to reliability and efficiency. Using the capabilities of convolutional neural network (CNN) models improved by hierarchical attention module (HAM), wafer hotspot detection is achieved throughout the fabrication process. In an effort to achieve accurate hotspot detection, this study examines a variety of model combinations, including CNN, CNN+long short-term memory (LSTM) LSTM, CNN+Autoencoder, CNN+artificial neural network (ANN), LSTM+HAM, Autoencoder+HAM, ANN+HAM, and CNN+HAM. Data augmentation strategies are utilized to enhance the model’s resilience by optimizing its performance on a variety of datasets. Experimental results indicate a superior performance of 94.58% accuracy using the CNN+HAM model. K-fold cross-validation results using 3, 5, 7, and 10 folds indicate mean accuracy of 94.66%, 94.67%, 94.66%, and 94.66%, for the proposed approach, respectively. The proposed model performs better than recent existing works on wafer hotspot detection. Performance comparison with existing models further validates its robustness and performance. metadata Shahroz, Mobeen; Ali, Mudasir; Tahir, Alishba; Fabian Gongora, Henry; Uc Ríos, Carlos Eduardo; Abdus Samad, Md y Ashraf, Imran mail SIN ESPECIFICAR, SIN ESPECIFICAR, SIN ESPECIFICAR, henry.gongora@uneatlantico.es, carlos.uc@unini.edu.mx, SIN ESPECIFICAR, SIN ESPECIFICAR (2024) Hierarchical Attention Module-Based Hotspot Detection in Wafer Fabrication Using Convolutional Neural Network Model. IEEE Access, 12. pp. 92840-92855. ISSN 2169-3536 document_url: http://repositorio.unib.org/id/eprint/14283/1/Hierarchical_Attention_Module-Based_Hotspot_Detection_in_Wafer_Fabrication_Using_Convolutional_Neural_Network_Model.pdf