Products In Action

Machine Vision Software Improves Accuracy of Photomask Repair System

As ICs continue to become more sophisticated and minimum feature sizes keep shrinking, photomasks have also become increasingly complex. Instead of being based primarily on square and rectangular chip designs, many of today's advanced masks feature curves, arcs, and other odd-shaped, submicron-sized patterns. Optical proximity correction (OPC) masks, for example, use a variety of strategically placed, randomly shaped patterns in order to compensate for the optical diffraction of UV light as it is exposed through the ultrafine lines.

Producing highly complex masks, like the IC fabrication process itself, is an extremely intricate process involving many steps, and the slightest defect that occurs during the process can jeopardize the circuitry of an entire wafer. Masks have become increasingly expensive, ranging from $5000 to $50,000 for a single mask. "With masks having become such highly valued components of the semiconductor manufacturing process, the ability to accurately repair mask defects has become a critical need among both mask makers and semiconductor manufacturers," says Greg Athas, senior software engineer at Micrion (now part of FEI), a Peabody, MA—based supplier of photomask repair systems and metrology equipment.

Pattern-matching operation steps: a mask defect is selected (top left); the defect site is overlaid on top of a good pattern so that the software precisely aligns the two images and performs image subtraction (top right); the resulting image indicates what must be cut away during repair (bottom left); and the final result is displayed (bottom right).

To better ensure the accuracy of the mask repair process, Micrion has incorporated the PatMax machine vision software technology from Cognex (Natick, MA) into its latest advanced mask repair system, the Micrion 8000. The software is used to perform a process that the system supplier refers to as pattern copy. This is essentially a pattern-matching operation during which an image of a mask defect is selected using a mouse, and is then cut and pasted on top of a good image of the pattern. The vision software then precisely aligns the two overlaid images, typically within a few nanometers, and then "subtracts" the regions that match. The resulting image represents the defect areas or regions on the mask where unwanted material should be removed or missing material replaced.

"At that point," explains Athas, "the machine has the information it needs to go in and perform the repair, which involves using a focused ion beam to etch away excess chrome on the mask or, in situations where voids are present, deposit an opaque carbon film to make the repair. Once a site is repaired, the result should look like a 'copy' of the nondefective pattern."

While the process sounds relatively simple, precisely matching up good and bad patterns of advanced photomasks has traditionally been a great challenge for machine vision, according to Athas. "In some of our older machines we implemented a homegrown pattern-matching technology, however it didn't do well in situations where the shapes were highly complex or where there was a lot of contrast variation between good and bad patterns."

The software engineer explains that isolated chrome areas on the mask surface can charge up from the effect of the ion beam over time, which can affect the contrast of the mask surface and cause patterns to image poorly. "Unlike many vision applications, our machine's imaging system does not use a camera and frame grabber to capture images. Instead, it uses a multichannel plate detector which measures the number of electrons emitted from the mask surface and turns this data into a bitmap image. However, since it is difficult to completely control the charging of the mask surface and no way to manipulate the image, the image we end up capturing can have significant contrast variations from the good image it needs to be aligned with."

To handle these variations, the machine vision software uses a patented geometric image analysis technique as the basis for comparing images of good and bad mask patterns. It seeks out unique geometric features found in both images and uses those features for aligning the two images, regardless of contrast differences. "What is also significant is that PatMax can automatically find the best alignment features common to the two images, which can be a difficult task because there are usually a lot of features present in the defect pattern that are not present in the good pattern," says Athas. "This relieves the operator of having to spend time analyzing a scene where there is a lot of image clutter and trying to guess which features would work best for lining the two images up."

The vision software's alignment is accurate up to 1/40th of a pixel, which is critical for the repair of highly advanced mask patterns. "With a lot of the smaller, more unique patterns we're seeing today, many of our customers need to perform repairs as small as 0.5 µm," Athas points out. "Thus, being able to align with subpixel accuracy is essential."

While the pattern copy process can leave repaired sites visually indistinguishable from nondefective areas, knowing how well the process has worked is more than a matter of seeing the results. To verify alignment accuracy, PatMax reports a score that indicates how closely the repaired image matches the good image. For example, a score of 1000 would indicate a perfect match between the two images, while a score of 0 indicates that the two patterns don't match at all. The score is based on how well the individual geometric features of the two images conform to each other. "Score values appear right on the monitor as the alignment happens, giving you instant feedback on how good an alignment you have," says the Micrion engineer.

Athas feels the new vision technology will go a long way in helping to improve the accuracy of mask repair. "So far, PatMax has helped our customers repair the most advanced masks out there, no matter how small the linewidths are or how complex the design is. This helps to eliminate the very expensive task of having to rewrite masks and improves overall mask turnaround time."

MicroHome | Search | Current Issue | MicroArchives
Buyers Guide | Media Kit