Enhancing semiconductor device characterization with deep learning-based keypoint detection

• Main Author: Elysia
• Other Authors: Kedar Hippalgaonkar
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Materials
• Online Access: https://hdl.handle.net/10356/168299
• Institution: Nanyang Technological University

The urgency to accelerate the development of electrical components due to the high demand for electronic devices in different applications has driven the implementation of automation in the research phase. Machine Learning (ML) and Artificial Intelligence (AI) have been utilized to boost and support the usage of automation applications in research, including the implementation of computer vision to automate the tedious electrical characterization process. However, algorithms currently used depend on morphological transforms that yield less robust keypoint predictions. In this research, deep learning using keypoint R-CNN was implemented to detect the electrical component’s contact coordinates for the probe to measure the device. Its capability to keep learning from the detection of false predictions opens the opportunity for more robust predictions. Transistors were used for the keypoint R-CNN training to detect the transistors’ contacts. Training the model on the transistor dataset that has been manipulate with OpenCV and Albumentations libraries resulted in accurate contact detection. The final weight of the trained model was then being implemented in the live imaging on microscope with a real time detection of the transistor’s contacts. The final contact coordinates were encoded in .csv format to be passed to the automated characterization’s probe. To expand the scope of the research in characterizing various types of electronic devices, the previous study on transistors is further extended to other electronic devices that share similar features as the transistors. This is to determine whether it could improve the detection accuracy for new devices. This can be done by either utilizing the transistor trained weights for other devices or integrating the transistor dataset into the training performance of other electronic devices’ contacts. By leveraging the deep learning techniques in the research phase, we aim to accelerate the process of materials discovery through a robust automated characterization tool.