Silicon manufacturing technology is now able to shrink critical dimensions of structures down to scales well below 100 nanometers (roughly the width of the influenza virus). These nanostructures are too small to see, even with the most sophisticated imaging equipment. This presents a challenge for Intel engineers who manually inspect microprocessors to identify which nanostructures are defective, so that repairs can be made.
Silicon nanostructures are difficult to see for two reasons: First, the imaging tools used to visualize the nanostructures are not able to resolve them in great detail, so the images are very low-resolution. Secondly, most tools, once they zoom in far, provide very weak signals and add a lot of "noise" to the image-electric noise from amplification, and imaging noise created by the nature of the process being used to generate the image. Viewing these images is analogous to looking through a textured bathroom window pane at a frosted-glass object; the image is blurry and obscured by reflections.
The Computational Nanovision project team, led by Horst Haussecker, is addressing these challenges associated with nano-imaging. They are using computer vision techniques based on sophisticated mathematical models to enhance image data, and to measure and reconstruct the shape of silicon nanostructures in real time.
In carrying out their research, the team is collaborating with researchers at Brown University and MIT. As the project progresses, the team will continue to share knowledge and exchange ideas through presentations, publications, and participation in key conferences.
A Model-Based Approach
The Computational Nanovision team is researching model-based approaches to nanovision. Because of Intel's deep experience in semiconductor manufacturing processes, researchers have detailed models of the nanostructures of microprocessors. They know where features should appear, even if the human eye cannot discern them in the unprocessed images. By combining this information with knowledge of the physics underlying image formation, they are developing new model-based techniques for analyzing nanostructures in images; such as image analysis techniques for real-time image reconstruction, feature detection, and classification.
Right: Noisy image from a direct-write nano-machining tool, showing sub-micrometer silicon structures. Left: De-noised real-time reconstruction of the image.
Image and Surface Reconstruction. The smaller the structure, the noisier the image. As nanostructures continue to shrink in size, we are reaching a point where it will be virtually impossible for an imaging tool to directly visualize features such as parts of transistors in a microprocessor. To reconstruct noisy images of silicon nanostructures, researchers first create statistical models of the noise distribution of specific imaging tools and incorporate them into a Bayesian de-noising framework. This allows them to separate the real image structure from the noise, and provides the user with a significantly enhanced image.
In some cases, users wish to visualize the real three-dimensional structure of an object rather than a scanning beam image. To this end, the research team has recently developed a novel technique for rapidly generating three-dimensional structures from two-dimensional images of scanning electron microscopes (SEMs)-a process that was computationally intractable in the past. Using this new technique, 3-D reconstructions of nanostructures visible in an SEM image can be obtained within minutes.
Left: A two-dimensional image taken by the scanning electron microscope. Right: A three-dimensional reconstruction of the image.
Nanofeature detection and classification. In the process of debugging a microprocessor, an operator sits in front of a screen on which the image of the silicon is projected and uses a nano-machining tool to expose the layer where the defect is likely to be found. The operator must quickly interpret the blurry image on the screen and decide exactly when to stop the nano-machining tool.
As silicon structures get smaller, images become more blurry, and it is harder to identify what is a structure and what is noise. The operator must take more time to interpret the images, or risk milling away too much of the silicon and destroying the microprocessor. Researchers are automating this operation, so that when the system recognizes defective structures, the machine will stop automatically.
Research Progress
Since the Computational Nanovision project was launched in March 2002, researchers have developed a toolbox of solutions for specific application areas. They have also identified the prerequisites for any successful solution: a fundamental understanding of the imaging process (which varies by imaging tool) and the ability to mathematically model the process, which requires prior knowledge of what a given structure should look like.
The computational nanovision techniques that the team has pioneered are now routinely applied within Intel to find defects in lithography masks. They are also being used for failure analysis and silicon debug, enabling experimental rewiring of microprocessors and generating significant cost saving in manufacturing and development.
Numerous other potential applications have yet to be investigated. The probabilistic techniques developed by the Computational Nanovision team could be applied to nano-imaging a variety of objects, such as biological structures and organic crystals. Intel is actively seeking collaborations with researchers in these domains to explore the limits of this exciting new technology.
