Advanced virtual metrology using distribution alignment loss for practical quality control

Virtual metrology (VM) model is essential for managing the manufacturing status in semiconductor processes without physical metrology. The model can be developed using machine learning or deep learning models from various sensor data collected from equipment. However, previous studies on VM have focussed on accurately predicting individual wafer metrology values. In the field, however, the distribution of metrology values is considered much more critical than individual wafer values for detecting the status of the equipment or production. This is because an uneven distribution or specific trend in wafer metrology values implies an abnormal state of the equipment or product. In this study, we aim to develop a novel VM that can effectively align the distribution of wafer metrology values. To achieve this, we utilise a Transformer-based deep learning architecture and distribution alignment loss (DAL). DAL consists of mean squared error loss, which follows the trends of metrology values, and mean variance error, which matches the actual variance of the metrology values. Experimental results using real semiconductor data showed that our proposed approach detects equipment abnormalities more effectively than other comparative methods. The proposed VM model will be practical for engineers in managing equipment and quality in real-world environments.