Deep learning-based automated metrology tool for semiconductor characterization

Semiconductor manufacturing involves the fabrication of billions of components with nano-scale specifications. While it is impossible to verify the dimensions of every part, process engineers can design a fabrication recipe that is reproducible, consistent, and leads to the desired uniformity of components.

However, developing a foolproof recipe requires extensive testing of different process parameters, followed by characterization of fabricated test elements. The latter step is performed mostly manually, and can be tedious and highly time-consuming. As a possible solution, Baderot et al. applied deep learning techniques to perform automatic and accurate metrology on electron microscopy images, which may ease the burden of manual characterization for the semiconductor industry.

“We estimate a reduction of the time spent on metrology between 25 to 75%, which brings different advantages — optimization of R&D costs and improvement of the quality of metrology and object characterization,” said co-author Julien Baderot. “Perhaps the most important advantage is the faster development of a fabrication recipe, which leads to reduced time to market.”

To create their tool, the team combined two approaches: deep learning-based object detection to account for microscopic images with large object variability and instance segmentation techniques to detect and segment the microscopic objects. Despite using pre-trained weights and very limited custom data to fine-tune the tool, it was able to outperform detection models built using traditional methods.

“Deep learning is known to require a large amount of data, which is a drawback,” said Baderot. “In this study, we show that the amount of data required can be reduced so that it can be easily managed by a human while maintaining high performances.”

Source: “Application of deep-learning based techniques for automatic metrology on scanning and transmission electron microscopy images,” by J. Baderot, M. Grould, D. Misra, N. Clément, A. Hallal, S. Martinez, and J. Foucher, JVST: B (2022). The article can be accessed at http://doi.org/10.1116/6.0001988 .