Resolved Darkfield Imaging Extends Wafer Inspection
Alexander E. Braun, Senior Editor -- Semiconductor International, 7/1/2005
Darkfield technology has been around for some time, and has provided valuable contributions as a wafer inspection resource. As the industry has continued its hamster-on-a-wheel run after Moore's Law, OEMs have extended darkfield imaging technology's capabilities for wafer inspection applications, adapting to new developments — design rule by design rule — by shrinking the illumination spot. This has kept it ahead of the resolution race. However, traditional darkfield inspection is running out of tweaks and, at 65 nm, is expected to run out of steam. It faces a landmine field of trade-offs — how to reduce the illumination spot to cope with ever-smaller shrinks while maintaining sensitivity to avoid compromising throughput at the 65 nm node and beyond.
At the heart of most darkfield inspection systems is the acoustic-optic deflector (AOD). The AOD is a single crystal that exhibits certain properties when a high-frequency signal is applied to it, enabling it to refract a laser beam. If the beam is refracted very quickly, the result is similar to that of drawing a laser scan line on the wafer. It is this line that defines the height of each inspection pass across the wafer, which is directly tied to throughput.
Traditional darkfield architecture is simple: It is based on an illumination spot and photomultiplier tube (PMT) sensors.
Most traditional darkfield inspection systems today achieve ~100 Mpixel/sec. Theoretically, this is extendable to ~300 Mpixel/sec. However, wafer-structure scattering in darkfield can range from a few photons in one pixel to millions of photons in the next pixel; when sampling time is reduced to >3 nsec, the electronics that support a high dynamic range (>12 bit), single-tap digitization system become increasingly difficult to develop. Another barrier to extending laser-spot scanning is that, as the spot size is reduced to magnify inspection sensitivity, power density increases, raising the risk of material damage on advanced wafers. Additionally, the spot-scanning architecture cannot scale resolution and simultaneously preserve optical Fourier filtering. This is significant because Fourier filtering can increase the sensitivity of laser-spot-scanning systems to advanced SRAM and DRAM arrays by a factor of 10.
KLA-Tencor (San Jose) developed a solution to get around these fundamental limitations by moving the problem from the front end — where the illumination takes place — to the collection end. This simplifies the requirement of tracing a line along the wafer by using optical lenses. Their new Puma 9000 platform uses a proprietary architecture that enables it to image a long line on a wafer and process it, in real time, across multiple collection channels that create a double darkfield plane. A new linear sensor and architecture were designed — the company calls this combination "Streak" technology — which enable traditional darkfield limitations to be resolved differently, providing >500 Mpixel/sec, as well as ≤65 nm line monitor capability. There is sufficient computing power for sophisticated algorithms to analyze data from multiple channels, in real time, without extending scan time and sacrificing throughput.
Before, a laser light spot and a PMT were used; however, a PMT is a light detector, not a means to higher resolution. Resolution originates from the illumination spot's size. The platform controls the illumination spot and, because it has a pixilated sensor, there is also resolution in the collection path. The combination of the imaging technique with the illumination line allows the system to Fourier filter any advanced array structure, achieving a 10× sensitivity increase, while increasing resolution in the illumination line and imaging collector. This combination of illumination and collection optics with double darkfield geometry provides the best defect sensitivity at traditional darkfield wafer throughputs.
A typical application is in void detection on a layer like STI CMP after a silicon nitride strip. Voids can range anywhere from 200 to 20 nm. Puma has detected >50 nm voids at >10 wph throughput — impossible for traditional darkfield systems. Another example is the detection of single missing contacts or even partially etched contacts at 90 and 70 nm design rules, which, until now, has typically been a challenge for inspection systems because of the noise associated with the contact layer.
Since the platform is not limited by AOD requirements, it is expected to be extendable over several generations.
For additional information on inspection, measurement and test, go to www.semiconductor.net/imt.
