A MetroSpection use case: enabling CNN ADC on CDSEM images for robust R&D developments and improved manufacturing quality control

<strong>Background</strong>: The microelectronics industry faces constant challenges in improving the quality control of silicon chip manufacturing. One of them is to quickly detect and diagnose process or equipment drifts in production, as well as enabling very early detection of process marginalities during R&amp;D phase. In previous publication, we have shown that Convolutional Neural Networks (CNN) are of high interest to perform Automatic Defect Classification (ADC) on many types of images gathered during various manufacturing process control steps [1]. More recently, an illustration of a real time CDSEM image computing framework, the ImageLab sandbox, was presented [2], but was mainly focused on contour metrology [3-4]. <strong>Aim</strong>: In this paper, we are evaluating the natural extension of this “in-house” infrastructure to perform CNN-based ADC on CDSEM images, that were originally produced for metrology and not inspection purposes. Indeed, anomalies can sometimes be seen on the CDSEM images, but those do not disturb the measurement. Since the measurement is within specifications, there is no reason for the manufacturing control system to stop the process, nor the equipment. Because this measurement process is automated, most images are not manually checked. Adding a CNN-based ADC systematic classification on those already existing images could be an efficient way to quickly identify non-quality events. We are also investigating use cases leveraging a dedicated CDSEM images capture step for ADC. Indeed, when defect mechanisms and locations are well known, metrology-SEM based ADC presents various advantages compared to Inspection-SEM that are more suitable for an Inspection-Review mode. <strong>Approach</strong>: The ImageLab has been presented at SPIE 2022 [5] and EMLC 2023 [2]. It provides real time CDSEM images processing capabilities. In this work, we will focus on AI based solutions, especially open-source CNN (Convolutional Neural Network) architectures. The main interest of such an environment is to enable Big Data approaches, by emulating data collection on production line in parallel of the real production, instead of limiting the use of metrology data to perform single lot analysis. <strong>Results</strong>: This CNN-based ADC approach on metrology images was successfully proven useful on several use cases, for example in the context of the introduction of a new derivative technology, on top of 4 already existing ones on that node. It allowed R&amp;D engineers to quickly highlight that the new technology’s process integration modifications and product layout influenced the overall process. Our MetroSpection approach was able to detect anomalies even though the CDSEM measurements were within the specification limits. Suited Graphical User Interface clearly illustrates this process deviation, while showing that the other 4 technologies in production were not affected. It also helped to better define where this marginality happened during the process flow, in this case, the last 2 metal levels. Besides the native ADC CNN capability, other challenges are addressed to allow adoption of such capability in a mass production environment. One of these is model training. The first model trained appeared to be not accurate enough and therefore not usable in production because it was too noisy. In fact, sensitivity to CDSEM machine type was predominant. To be able to retrain the model efficiently, appropriate visualization of the last embedded layer results of the first model is performed. This layer is the final and finest data treatment layer used by CNNs before classification and contains the extracted characteristics of the images. The visualization is therefore an illustration of how CNNs interpret our image dataset and can be used to identify weaknesses in classification performances, often exhibited by cluster merging. It also enables data selection using lasso techniques, and re-labeling of the images can be operated, leading to an improved dataset, that will be able to produce, in turn, a better tuned model. <strong>Conclusion</strong>: The ADC CNN on CDSEM images is now running in production for several use cases. Applications are either for manufacturing control or R&amp;D. Use cases for this approach are numerous and some of them will be presented in this paper. Promising user friendly tools are also in development to increase and speed up mass production environment adoption.