Semiconductor Test Positions for Success
Accurate feedback, nonlinear control algorithms, and sophisticated materials focused imagers and position probes for semiconductor test applications
By: Kristin Lewotsky, Contributing Editor
When it comes to semiconductor fabrication, time may be money, but yield trumps all. It doesn't matter how fast the fabrication steps take place or how many chips come out of a given batch if they don't work and have to be scrapped. As a result, while process steps must obviously clock with economically-viable throughputs, testing is every bit as important. With critical dimensions currently in the tens of nanometers across wafers as large as 450 mm, the ability to effectively and repeatably position wafers, optics and measurement equipment is crucial to capturing accurate data. That makes motion control an essential supporting technology for the semiconductor industry.
We can loosely divide semiconductor manufacturing into front-end and back-end processing. The separation between the two is less a line than a gray area, but for purposes of this article we will define front-end processing as everything involved in patterning the wafer, and back-end processing involving everything afterward, beginning with dicing and moving through packaging. Front-end test processes include wafer inspection and process-control metrology that takes place between process steps like etching and implantation. Back-end test processes include testing individual die prior to dicing the wafer and checking packaged devices.
Wafer inspection techniques range from whole-wafer optical imaging to localized optical microscopy, scanned-probe, and electron-imaging. The challenge here involves maintaining the area under inspection in focus, which becomes increasingly difficult as wafer sizes increase. Meanwhile, with shrinking critical dimensions, the size of defects that can impact chip performance has gone beyond the resolution of optical technology, forcing manufacturers to turn to alternatives like atomic-force microscopy.
For successful inspection, wafers need to be positioned in the focal plane of the system, and translated so that all of the wafer can be mapped at the highest possible speed. That requires a combination of millimeter-scale motion in the x-y plane, combined with sub-micrometer- or even nanometer-scale motion along the z axis, as well as along the x-axis and y-axis.
Once the wafer has been patterned, the dice need to be tested. As profit margins have tightened, manufacturers have expanded testing from a few samples for process control to, often, hundred percent inspections. To maintain throughput, a probe testing dice on a 450-mm wafer must move in 10- to 100-µm increments between test points in milliseconds.
In many ways, the motion problem for both front-end and back-end testing is the same: moving in rapid, small increments. The problem, of course, is that the faster the motion, the greater the likelihood of overshoot or ringing, which increases settling time. “If you look at the physics of settling, it's an exponential process," says Scott Jordan, Director of Nanoautomation Technologies at Physik Intrumente (PI). “The finer the tolerances, the longer it takes for things to fall into position. If you're in a field like semiconductor manufacturing, resolution is really important and so is time. Those are difficult challenges to meet simultaneously."
To address specifications, manufacturers use techniques like observer control, in which data and modeling yield sophisticated motion profiles that allow the system to decelerate as rapidly as possible while minimizing ringing and overshoot.
Two common positioning solutions for semiconductor test are linear motors—basically a rotary servo motor stretched flat—and piezoelectric positioners, which are formed of stacked ceramic material that expands under an applied voltage. Linear motors can achieve sub-micron resolution, with length of travel in theory limited only by the length of the magnet track. Piezoelectric stack positioners offer sub-nanometer resolutions with travel as long as 2 mm, depending upon the design. Often, the two technologies are combined to form stages with coarse- and fine-positioning capabilities (see figure 1). With both devices slaved to a high resolution encoder, such a system can provide nanoscale repeatability over hundreds of millimeters, allowing the system to scan a wafer and then accurately return to a known defect.
The numbers given above are ballpark figures--ultimately, accuracy, repeatability, and resolution depend upon the quality of the feedback. One increasingly common approach is to use multiple encoders, one on the motor and one on the load. System designers are also turning to newer technology, like synchronous serial interface (SSI) protocol. "In the last five to seven years, we've seen a lot of SSI encoders," says Kaushal Shah, Director of Technical Support Group at Galil Motion Control. "In the last three years or so, we're been seeing a lot of BISS encoders used for absolute positioning. They provide both the resolution and the high update rate that allows us to achieve high bandwidth in the system.”
Techniques like position-edge generation (PEG), in which the system generates a pulse or edge when it reaches a requested position, allow the controller to confirm that the load has reached the desired point. “We use high-speed position latches and output compare that we're comparing very quickly in hardware so when the signal comes in, we're capturing the position almost instantaneously," says Lisa Wade, VP of Sales and Marketing at Galil Motion Control.
Piezoelectric positioners more commonly use capacitive sensors that consist of opposing diamond-machined plates. As the piezoelectric material expands or contracts, it moves one of the plates. Changing the separation alters the capacitance of the device. The signal can be processed to yield nanoscale absolute displacement data from a high-bandwidth, environmentally stable source.
In the case of wafer inspection systems for which maintaining focus is key parameter, optical techniques provide another type of feedback. "Systems need really fast, repeatable, accurate focusing - not so much on a sensor of position as a sensor of focus," says Jordan. One approach is to use a dedicated optical sensor, ideally sensing the focal plane through the imaging optics to avoid offset issues. This method provides speed advantages while also allowing auto-focus of unpatterned areas, important in coating metrology. Real-time integration of such sensors into piezoelectric focusing systems is a recent area of significant advancement.
Fourier analysis of the image of the wafer provides another way to determine focus. An in-focus image of a patterned wafer will have lots of sharp edges and, hence, a large amount of high-spatial-frequency content. An out-of-focus image will contain a large amount of low-spatial-frequency content. The analysis can be performed rapidly and the positioner can drive the wafer to the ideal position.
That, of course, brings us to the question of control.
Keeping to control
One of the challenges of semiconductor test is that the application not only requires high resolution, accuracy, and repeatability, it requires that performance from multiple highly-synchronized axes. The more axes a system includes, the more cabling it has, which introduces points of failure, not to mention cost, and - important in the contamination-sensitive semiconductor environment - a source of particulation. One option to minimize cabling is to use distributed control, but it's not as popular in this application as you might think.
In the case of a food packaging line that stretches over hundreds of feet, a distributed architecture consisting of daisy-chained motor drives provides a good alternative to long cabling runs. Semiconductor test equipment, in contrast, features runs on the order of a meter or less. Although the axis counts are probably lower than for that high-speed packaging line, the performance requirements are far more stringent. As a result, systems typically use centralized control with a high-speed processor such as a reduced instruction set computer (RISC) chip. "The moves are so short and the axes are so tightly coupled that you need a dedicated processor," says Wade. "It's really important that you're not trying to use a bunch of independent, single-axis controllers and smart drives to coordinate through software through the host.”
Systems typically leverage motors on the order of a few hundred watts, at best. Improvements in electrical components and heat dissipation have reduced the size of motor drives to a fraction of their former bulk. Ten years ago, a controller/drive package for eight 500 W motors might have taken a space the size of a large shoebox. Today, a controller/drive package for the same eight axes is about the size of a hardbound novel (see figure 2).
Not only have they gotten more compact, they have become more economical, as well. "Drive components have come down so much in price and gotten so much more thermally efficient that you can get, in quantity, a four-axis 500-W drive for $100 per axis," says Wade. "That's why we’re still seeing a lot of centralized control because if you can figure out how to get the motion controllers and the drives all in one package in a small space, then you can solve that wiring problem. Especially if you're at eight axes or even 16 axes, it's going to be more cost effective than it will be with separate single-axis controller and drive packages."
That is not to say that designs don't on occasion leverage one of the many flavors of Ethernet to connect feedback to the controller or even tie unrelated axes to a set of highly synchronized ones. Consider an eight-axis controller, for example. Positioning a probe to test the individual dice on the wafer prior to dicing might require six axes of tightly coupled motion. Those six axes operate via a centralized architecture out of a central cabinet. Placing the controller in an enclosure with the drives, all as near to the probe as possible, minimizes cabling runs, as well as the cable flexing that can generate particles. At the same time, the other two axes handled by the controller may be on the other side of the machine, wired in via an Ethernet connection.
Controls present other challenges. In the case of piezoelectric positioners used with capacitive feedback, the controller needs to be able to handle the specialized analog feedback required for sub-nanoscale sensing. Likewise, the drive requirements are different. Piezoelectric devices can require as much as 100 or 120 V DC. Perhaps more important is the way they respond to drive voltage. "The position of a piezoelectric device is more or less proportional to the applied voltage, whereas with a motor, typically the speed or force is proportional to the voltage being applied to it, so it's a completely different physical domain,” says Jordan. Getting the desired performance, devices requires choosing the correct controller.
Ultimately, no matter how good the components are, the system needs to be properly tuned and designed. Most test platforms require motion in x, y, z, and ?x, ?y, and ?z. It is important to consider specifications for multi-axis system rather than stacking up the specifications for the six individual axes. It is also essential to remember that changing direction can introduce error, whether from material hysteresis, as in case electrics, or backlash from the motor/gearbox combination.
A system may look good on paper, but when the motion takes place in the actual physical structure, resonances may appear, vibrations may get amplified. "When you're talking about those very precision applications, everything has an effect," says Wade. "Even though you have a controller that can compensate for a lot of it, you still have to do system tuning to make sure that the nonlinearities in the system are addressed.” Using mechatronic principles in the design phase can eliminate many of these issues; adjustable tuning parameters in the controller along with tuning software can handle the rest. "Using tuning software in the time domain, I can put a disturbance into the system and see how much overshoot I get, find out my rise time, whether I have ringing,” says Wade. “In the frequency domain, I can find out whether I have resonances. If I have resonances at a certain frequency and put a notch filter into the controller, I can find out how it's going to look. I think a lot of it comes down to tools and making measurements at that level."
Another, increasingly pertinent, approach involves using parallel kinematics, in which the workpiece is actuated simultaneously by multiple actuators, rather than a stacked approach. “Attack the stack, and you eliminate mass, orthogonality issues, material deflection, cabling messes, and sheer bulk,” Jordan notes. Parallel kinematic designs appear in the highest-performance piezoelectric stages as well as motorized hexapod positioners.
There may be some question about whether Moore's Law will continue to hold over the long haul, but one thing is certain: The semiconductor industry will continue marching down the road to ever smaller feature sizes. The success of the effort will hinge upon the availability of accurate, repeatable, affordable positioning. Already, a flood of new motion technologies has entered the marketplace in the past few years, and the rate of innovation and advancement in the field will only continue.