We, along with the rest of the lithography world, have been charting the progress of extreme ultraviolet (EUV) technology. There are a number of puzzle pieces required to enable that technology fully, but most eyes have been on the light source for patterning wafers: increasing source power has been a major requirement for achieving the kinds of throughput needed for high-volume manufacturing.
In fact, this concept of turning out large volumes of material is what separates science-project technology from production technology, so much so that “high-volume manufacturing” has earned its own TLA: HVM.
If you are thusly focused, then it may come as a surprise that there’s a completely different EUV source project underway, one that requires less power – but more brightness. One that, for different reasons, is also critical to a smooth manufacturing process.
Here’s the problem: masks and wafers need to be inspected. Masks have to be shown to be defect-free during various stages of their manufacture as well as while in use. Wafers need to be inspected as well to make sure that the wafer itself doesn’t have or hasn’t acquired defects while being processed.
Here’s the challenge: these defects are miniscule and hard to see. They’re tough for humans, but, more importantly, they’re tough for machines; these inspections aren’t done manually. To make matters more complicated, a defect’s appearance can depend on the light you’re shining on it. It’s kind of like a nano-version of a crime scene detective who’s looking for traces of blood. Under normal room light, it might look like a miscreant did a thorough clean-up job. But focus UV light on the scene, and suddenly new evidence that was invisible under normal light becomes visible.
The mechanisms that cause the crime-scene example may be different from chip manufacturing, but the point is that the wavelength through which you view these defects matters. Under one type of light, one set of defects will be visible; under different light, a different set of defects may become evident.
So, given that, how the heck are you supposed to do a thorough examination without getting bogged down in inspecting over a wide range of light wavelengths? The reality here is similar to the proverbial tree falling that no one is around to hear. If there is a defect that can’t be seen, does it matter?
The whole reason for looking for mask defects is that they screw up the high-fidelity transfer of a mask pattern onto silicon. If there’s a defect on the mask, it will create a defect on the wafer. If there’s a particle or some other defect on the wafer, then it will interfere with the mask pattern being exposed.
But those problems happen only if a defect is visible. Because if it’s not visible, it’s as if it’s not there. This exposure happens under EUV light, so the defects matter only if they are visible at that wavelength. Any other defects are simply irrelevant, and identifying them and attempting to repair them would be a waste of time.
This results in a simple inspection requirement: you need to inspect under the same light as is being used for lithography – in this case, EUV light. In fact, defect visibility is said to change pretty dramatically with wavelength, so the inspection light source must match the lithography wavelength very closely.
This notion of inspecting using the same wavelength as the exposure wavelength has a name: actinic inspection. It can apply both to mask and wafer inspection, but the remainder of our discussion will focus on masks.
Now, since we need the same wavelength as exposure, you might think that the guys making the exposure sources are the same as those making the inspection sources. Not so. The technology is similar, but, although the erstwhile Cymer group dominates the exposure source discussion, I talked with a different company, Adlyte, about their inspection source efforts and, in particular, their recent performance milestone.
Inspection vs. Lithography
Let’s start with the fundamental technology. Both companies use the laser-produced plasma (LPP) approach. This way of generating EUV light relies on a ridiculous-sounding concept: establish a steady drip of ultra-tiny tin droplets and fire a laser at them in mid-air as they fall. When the laser blasts the droplet, it generates a tiny pulse of EUV light. Do this really often and you generate enough light to do something useful.
That “enough light” thing means around 100 W or more of delivered in-band power for exposure light. You need that much in order to expose the photoresists quickly enough to keep a short exposure time for high throughput. But with inspection, the light will reflect off of the target and form an image on an EUV sensor instead of a wafer. This requires less power – it could be lower than 1 W, ranging up to 20 W, depending on the application.
But what matters more is brightness: this is the power over a spherical arc area; the unit is W/mm2-steradian. This geometrical concept is referred to as the “etendue” (from the French étendue géométrique, or geometrical extent). Essentially, if you dial down the etendue, you’re focusing the given power of light onto a smaller area, increasing its brightness. Of course, when you do that, you reduce the size of the area inspected, which slows down inspection of the overall mask. Presumably one of many tradeoffs.
The brightness requirements go up as the silicon node gets more advanced. This feels to me all too familiar: those of you in my age grade may also be finding small print harder and harder to read. But often, a passage can be made more readable if you simply shine brighter light on it. This feels analogous to the need for brighter light to illuminate smaller features and keep the signal-to-noise ratio reasonable. Specifically, the brightness range can be from 20 to 150 W/mm2-steradian, but 30-100 is a typical range.
The LPP process we’ve seen before for lithography uses a “pre-pulse” approach. The idea is that the droplet being targeted is small as compared to the width of the laser beam. It’s the laser that carries the energy to be turned into UV light, so any of that laser beam that doesn’t hit a droplet is wasted. So the pre-pulse (like a red-eye reduction flash in a camera) hits the tin droplet with a lower-energy dose, which “puffs” it up, creating a bigger cross-section that intercepts more of the laser light on the following high-energy laser blast, taking advantage of more of the available laser energy.
The laser used for lithography has a beam focus somewhere in the range of a few hundred microns or so. By contrast, Adlyte’s laser has a 70-µm focus; aiming at a 40-50-µm tin droplet means that there is no real need for the pre-pulse.
Other differences include the frequency (the number of droplets per second): litho typically uses 30-50 kHz; Adlyte uses 6-10 kHz. Cymer uses a CO2 laser; Adlyte uses a YAG (yttrium aluminum garnet) laser with roughly 1/10th the power.
The inspection image is created by integration over many droplets (usually in the hundreds); the exposure time is 10-100 ms.
Keeping it clean
But there’s more to making this work than simply firing an insanely small laser at insanely small tin droplets insanely quickly without missing any. You also need to do this “cleanly.” Exactly what that means requires another definition: the “intermediate focus.”
These machines, whether for exposure or inspection, aren’t constructed solely by one company. In fact, companies like Zeiss and ASML build the final machines, integrating EUV sources made by an upstream supplier. So, in essence, there’s a point where light exits the source and enters the rest of the machine. In cases where the source has focusing collector optics (which is most cases, but, strictly speaking, not required), this boundary is referred to as the intermediate focus.
We also need to picture that this is an inherently violent process, with tin droplets being smashed to smithereens over and over again. It’s not like the tin vaporizes and ceases to exist (we would need anti-tin and a penchant for enormous explosions for that). The laser zap can result in debris, which can go… well, wherever it can go, which becomes a design issue.
Managing this debris in the source is a matter for maintenance; obviously, the easier the machine is to clean and the less often cleaning is needed, then the better the machine uptime will be. Just as we have collector hazing with exposure EUV sources, so debris in an inspection machine could impinge upon process control sensors and the optics. Because of this, Adlyte says that they’ve made debris mitigation an intense focus.
But the problem doesn’t stop in the source. One of the challenges of EUV is that pretty much no materials are transparent to it. So you don’t have “normal” lenses and other “tranmissive” optics; you have mirrors and reflective optics. This means that the pathway between EUV source and whatever the light is hitting has to be open. Put a lens or any other physical barrier there and you’d block the EUV light.
That means that there’s an open path between the scene of the tin droplet apocalypse and the target mask. And the concern is that debris could make it out of the source and end up, in the worst case, on a mask. This has obviously been a design consideration, so for a given droplet, it’s exceedingly unlikely that some of the resulting shrapnel will make it out to the mask – a significant distance to travel. But statistically, given billions of droplets, that “extremely unlikely” starts to sound possible.
Adlyte says that this cleanliness is more important for mask inspection than it is for wafer exposure. When running wafer production, if a mask were contaminated by some tin debris, it wouldn’t be good, but the messed-up mask could be swapped out with a clean mask so that production could resume while the sullied mask was cleaned or repaired.
That concept doesn’t apply when inspecting a mask itself. If the inspection process damages the mask, well, that’s just not good, and there’s no “swapping” that will help. In theory, you could end up in an infinite inspect-spatter-clean-re-inspect loop.
This goes to the relevance of Adlyte’s recent announcement: they simulated production conditions and ran the machine for hundreds of hours of tests with no contamination beyond the intermediate focus.
We’ve talked generically about “inspection,” but there are, by Adlyte’s count, three different types of actinic inspection: blank, pattern, and AIMS.
Blank unit inspection is more or less obvious: verifying whether the starting material is clean enough to begin work. With masks, there are actually two versions of this: pre- and post-“mirror.” Because EUV uses only reflective optics, that means that an EUV mask won’t look like the transparent masks we’re used to. Instead of shining light through a mask onto a wafer, light is reflected off of the mask and onto the wafer.
The reflection is managed by multiple layers that make up something akin to a Bragg grating. So a pre-mirror inspection checks the virgin mask blank to make sure it’s ok. Then the mirror layers are added (but, at this point, with no patterning). An inspection after the mirror is installed is the post-mirror check; the mask is still blank, but because it has been processed, you need another inspection to ensure that no defects were created.
Pattern inspection checks the mask after it has been patterned. But the details of this – and, frankly, the other methods – are beyond the scope of the source manufacturers. They’re implemented by the inspection machine manufacturers.
That said, the AIMS approach is worth a bit more of a mention. I poked about the internet to learn more – with limited success. What I did learn is that “AIMS” stands for “Aerial Image Metrology System” and has been put forward by the Carl Zeiss company (they appear to have trademarked it). I’ve also come to learn that the paucity of information is partly because many of the details are hidden behind NDAs.
This inspection approach hinges on the concept of an “aerial image,” which I saw described as an image that appears to “float” in mid-air, typically visible only if viewed from a particular angle. It strikes me as qualitatively similar to a hologram image, which also appears to float, and whose details change as you change viewing angle. In the context of AIMS, the intent is to mimic the lithography tool to identify which defects would or would not print onto the wafer. And… that’s about all I know at this point.
Much of Adlyte’s take-away messaging is, more or less, that they’re very close to being ready for HVM. The cleanliness data helped. Up-time is another big aspect of production-worthiness. Their goal is for the machine to be up 24/7, interrupted only for changes of tin cartridge and source cleaning. The cartridges provide the tin in the droplets. It takes just a few minutes to swap them out, and the goal is for a cartridge to last 3-5 days.
These days, Adlyte is running in development mode – bringing the machines up to run tests and such for 3-6 hours at a time. But this matches the needs of the work they’re doing; it’s not based on any uptime limitation of the machine. They say that they are capable of running continuously.
So, while there may be a few more wrinkles to iron out, Adlyte sounds optimistic about their readiness.