Abstract: In the semiconductor manufacturing industry, the yield prediction of a product is vitally essential. Its accuracy directly affects product cost and customer satisfaction. One commonly known yield prediction and management technique is statistical machine learning models due to their high generalization capacity. With the support of rapidly increasing data volume, driven by advanced wafer-level processes, an increasing number of Deep Learning architectures have recently been adopted for yield prediction in semiconductor manufacturing. With rapid advancements in wafer-level processes, it becomes feasible to collect defect density data through advanced sensors from testing equipment. However, constructing Deep Learning models from scratch requires a lot of expertise in both semiconductor manufacturing and machine learning, which is not easy to obtain in modern semiconductor industries. To automatically assist semiconductor manufacturing engineers in building accurate Deep Learning models in yield prediction, this study first designs and builds a fully-automated Deep Learning yield prediction framework. This framework can assist engineers in developing Deep Learning models in a timely manner without time-consuming data preprocessing, feature engineering, or architecture searching. It consists of yield simulation, data preparation, candidate models building, and ensembling formation. Besides yield prediction, another significant concern is the explanation of Deep Learning model outputs. On the one hand, many model-agnostic explanation algorithms have been successfully adopted in many fields, providing valuable information for improving model quality and transparency. The explanation logic is intuitive; e.g., it calculates the contribution of a feature to the output of the prediction, which helps in finding out pixel areas or categorical reasons to focus. On the other hand, due to the complexity of semiconductor manufacturing processes, existing interpretable models lack effective inductive bias for proper yield prediction in semiconductor manufacturing, and turned out to be unable to simulate disentangled understanding rules. Therefore, current model-agnostic explanation methods for explaining Deep Learning model outputs either fail to reason rules about semiconductors, due to insufficient representation ability, or contain prohibitive time complexity for searching the activation of nodes or features. As a result, existing approaches struggle to simultaneously achieve both high prediction accuracy and high-quality explanations.
Keywords : Data analytics, yield prediction, semiconductor manufacturing, machine learning, predictive modeling, process optimization, defect analysis, big data, statistical process control, anomaly detection, real-time monitoring, artificial intelligence, wafer-level data, equipment data, root cause analysis, pattern recognition, data mining, manufacturing intelligence, sensor data, quality control, production efficiency, regression analysis, classification models, predictive maintenance, deep learning, feature extraction, high-dimensional data, yield enhancement, data-driven decision-making, advanced analytics.
Copyright Form