The continued improvements in lithography have been the driving force that has upheld Moore's Law. Shorter wavelengths, better lenses, and adaptive optics have all contributed to this success story, which has allowed commercial chips to be fabricated with 45nm feature sizes, with 32nm in testing. This process has succeeded beyond everyone's wildest expectations, but that does not mean that future feature reductions that rely on the same basic technology are guaranteed. This week, researchers published an adapted maskless lithography technique that may provide an alternative path to sub-32nm features.
The maskless part is pretty important. Normal lithography creates an image of the features that will be etched into a chip by imaging a negative mask. This mask is big, expensive, and can take quite a while to make. When you are just trying to get a chip design right, each little change means a new mask. You can imagine that this might increase the development cost somewhat, while also slowing circuit development.
Maskless lithography gets around the problem by writing the chip directly. This can be done by milling the silicon with a beam of electrons or ions. Electrons and ions can be focused to very small spots, meaning that features just a few nanometers in size can be written, but this is a seriously slow way to write an entire chip. Running many beams in parallel would speed the process up. Unfortunately electrons and ions are charged, so the beams interact with each other—parallel beams don't focus, and don't go where they are directed.
Light, on the other hand, doesn't interact so strongly with itself, making it the perfect candidate for parallel beam maskless lithography. Unfortunately, it doesn't focus so well—a really good lens will focus light to, at best, a quarter of the wavelength. Fortunately, there may be another option: the field of nanophotonics and something called a plasmon, and plasmon lenses.
A plasmon is created by the interaction between light and the sea of electrons in a metal. The light causes the electrons to oscillate and, when conditions are right, they oscillate as a group, creating extremely large electric fields. These fields are exactly like the light field that created the plasmon, except that they extend just a hundred or so nanometers from the metal surface. A lens, based on plasmons, can be created by a set of concentric metal rings. The fields from the plasmons in each ring act in such a way as to create a tightly focused spot of light. In principle, these lenses could focus light tightly enough to create features about five to ten nanometers in size.
There is, of course, a catch—the distance between the lens and the focal point is about 20nm. This is a bit of headache, because it means that the lens has to move accurately over a silicon wafer while maintaining a distance of just 20nm between it and the wafer. This is precisely the step forward described in the new paper.
The paper actually describes a bit more than that. The problem isn't so much maintaining the separation, but rather maintaining the separation while moving the lens with any speed. Since we only have a limited array of lenses, they have to move over the wafer in order to write the whole circuit; to write with enough speed, they need to move fast. Feedback electronics cound maintain the lens at the correct height, but their response time is too slow.
To get around this problem, the researchers used airflow over and under the lens to maintain its height. Basically, if the lens lifts too high, the air pressure under the lens falls, forcing it back down. Likewise, if it sinks too low, the air pressure increases, forcing it back up. The response time is governed by the mass of the lens, which is very light, allowing it respond very quickly.
In the actual apparatus, the airflow is provided by spinning the wafer under an array of 16 lenses at 2000 revolutions per minute. The write position is controlled by translating the lenses from the center of rotation outwards to the edge of the wafer and flashing laser light into the lenses at appropriate times—getting the sequencing right for a complicated structure must be quite tricky, but that is what computers are for. Now, since the rotation rate is constant, the speed of the wafer under the lens depends on where the lens is, which means that the write speed and control over the height vary as the lenses move outwards. The researchers showed that they could maintain the lens within 20nm of the surface over the full speed range, and importantly, keep the lens nearly flat as well (within a few millidegrees).
It is important to emphasize that there are still some big hurdles to overcome. First of all, for this to compete with the speed of commercial lithography, the number of lenses will need to be of the order of 1000. Second, they used a small glass wafer to test this on. A 12 inch silicon wafer will be quite another story. Third, the feature resolution was 145nm, which is far from ideal.
I worry about the last part the most. Once feature sizes get below 32nm, the silicon wafer doesn't look very smooth anymore. In fact, it rather resembles downtown Manhattan. Obtaining features that are smaller, or even about the same size, as the depth variations in the silicon surface is going to be very challenging for any lithographic technique.
That said, this work has a lot going for it. No nasty vacuums, no horrible discharge lamps, no complicated optics train, and no expensive mask. Instead, only a laser, a spindle, some precision (OK, a lot of precision), and a complicated computer program are required.
