Writing with light
Lithos is Greek for stone. Today the stone is a 300 mm disc of silicon, and the finest pen is light with a wavelength of 13.5 nm: a virus is almost seven of those waves across.
A projector running backwards.
A cinema projector throws a small film frame large onto a screen. A lithography scanner does the opposite. It takes a mask, the reticle, and throws its image 4× smaller onto the wafer, through a projection lens of many elements.
Look at the F. It lands upside down and mirrored, like any image through a lens. The chip designer never sees this. The mask is simply drawn the other way round.
The reticle is a quartz plate 6 inches square, about 152 mm, with the pattern in a thin film that blocks light. A pellicle, a thin film on a frame, sits a few millimetres above the pattern: dust that lands on it is too far out of focus to print. One field on the wafer is 26 by 33 mm, about the size of a €2 coin or a 500-yen coin.
Why shrink at all? A flaw on the mask shrinks by four as well. Berkeley Lab writes that the reduction greatly relaxes what the mask must meet, in the quality of its features and in defects.
The print is a stencil.
Light deposits nothing. It changes a thin light-sensitive film called photoresist. Where light lands, positive resist becomes soluble and washes away in the developer. What stays behind is a stencil for the real work: etching, implanting, depositing metal.
Then the resist is stripped and the next layer begins. Nikon writes that a chip goes round this loop several tens of times.
Bare wafer: silicon with a layer of oxide on top.
Coat: the wafer spins at high speed, so the resist spreads thin and even.
Expose: light passes the clear parts of the mask and changes the resist below them.
Develop: the developer dissolves the exposed resist. The stencil is open.
Etch: a plasma eats the oxide wherever the stencil is open, and nowhere else.
Strip: the resist has done its job and is removed. The pattern now lives in the oxide.
Fill: metal fills the trenches, and the top is polished flat. One layer done; dozens to go.
This is why lithography rooms are lit yellow: the light is filtered to hold back the ultraviolet that resist reacts to, and to pass the longer wavelengths, above about 500 nm, that people work by. It slows exposure rather than stopping it. In one university cleanroom in Korea, the old yellow tubes still fully exposed a resist in 60 minutes.
Shorter waves draw finer lines.
You cannot draw a line thinner than your pen. For light, the pen is its wavelength. The industry has walked down the spectrum: mercury-lamp lines, then excimer lasers, then EUV.
The waves are drawn to scale against each other. The bars below them are the lines each could print, worked out at k₁ = 0.3.
A mercury lamp line, visible blue. Early tools had a numerical aperture of 0.16, in Chris Mack's table. Nikon's first wafer stepper, in 1980, printed with g-line.
The next mercury line, just into the ultraviolet. Nikon shipped its first i-line stepper in 1984. Its i-line tool today has an NA of 0.62, Canon's 0.57 and ASML's 0.65, and Nikon sells it for the less critical layers.
Krypton fluoride excimer laser: the step from lamps to lasers. Nikon shipped its first KrF stepper in 1988 and its first KrF scanner in 1995. Nikon's current KrF scanner has an NA of 0.82, Canon's up to 0.86 and ASML's up to 0.80.
Argon fluoride excimer laser, in air. Nikon announced its first ArF scanner in 2001. Its dry ArF scanner today has an NA of 0.92, and ASML's up to 0.93: in air the practical limit is about 0.93.
The same 193 nm laser with water under the lens. Nikon writes that it shipped the first production immersion scanner in January 2006, at NA 1.07. ASML dates its first immersion machine to 2003 and its first for volume production to 2006. Both makers' current ones reach 1.35.
Tin plasma, mirrors, vacuum: a wavelength fourteen times shorter, bought with a much smaller NA of 0.33. One company builds EUV scanners, ASML, and these are its NXE machines. The first electronics made with EUV went on sale in 2019.
The same 13.5 nm light with an NA of 0.55, in ASML's EXE machines. ASML gives a resolution of 8 nm, and imec says the technology promises lines down to 8 nm. In 2024 imec printed lines down to 9.5 nm in one exposure, on the EXE:5000 in its joint lab with ASML.
Where lines blur into grey.
Light passing a fine pattern spreads out. The finer the pattern, the wider the spread. The lens catches only a cone of it, and what it misses is detail lost. imec calls the rule the Rayleigh equation: the finest half-pitch is k₁ × λ ÷ NA.
The surface is the light intensity across a row of lines. The flat plane is the resist's threshold. Where a ridge rises through it, a line prints. Shrink the lines and watch the ridges sink.
Simplified model. Scanners light the mask from an angle, so the lens catches two diffraction beams instead of three. That is what lets them work near k₁ = 0.25. Below it, no pattern is left in the light.
A film of water buys 45%.
NA = n × sin θ. In air n = 1, so NA can never reach 1, and in practice it stops near 0.93. The steepest rays leaving the last lens hit the glass surface and bounce back inside.
Fill the gap with ultrapure water and n becomes 1.44 at 193 nm. Those steep rays now reach the wafer and NA climbs to 1.35.
The catch is mechanical. Nikon moves the wafer under the lens at 500 mm a second or faster, and the water has to stay put and leave no spots on a surface that will be etched next.
Step, scan, repeat.
The scanner does not flash the whole field at once. A narrow slit of light stays still. The reticle and the wafer slide through it in opposite directions, the reticle four times faster. One field done, the wafer steps to the next and scans back the other way. Nikon gave its speeds in 2005: the wafer at more than 500 mm a second, the reticle at 2 m a second.
Before a wafer is exposed it is measured: a map of its height across the surface and the positions of its alignment marks, so each new layer lands on the one below. ASML's TWINSCAN scanners carry two wafer stages, so one wafer is exposed while the next is measured; Nikon's immersion scanner measures every wafer in an alignment station before exposure. Nikon specifies 2.1 nm or better between two of its immersion scanners, and ASML a matched-machine overlay of 1.3 nm for its own.
At 1× this scene finishes about 44 wafers an hour. For an immersion scanner, Nikon quotes at least 280 an hour at 96 fields a wafer, and ASML at least 310. For a dry ArF scanner, both quote at least 300.
Making light that air swallows.
At 13.5 nm no lamp or ordinary laser will do, and air, glass and water all absorb it. So the machine makes its own light, in a vacuum, from tin.
A tin droplet 20 to 30 µm across falls through the focus. A first laser pulse flattens it. A second, the main pulse from a carbon dioxide laser, turns it into a plasma at some 330,000 to 440,000 °C. The plasma glows at 13.5 nm, and a curved mirror, the collector, gathers that glow into one point on its way to the scanner.
One company builds EUV scanners: ASML, as the US Congressional Research Service and Georgetown's CSET write. In ZEISS's words, ASML developed this source, with a CO₂ laser from TRUMPF and a collector from ZEISS.
Real rate: 50,000 to 60,000 droplets a second in Michigan's paper, 50,000 in ASML's, ZEISS's and TRUMPF's; ASML gives its droplets as about 25 µm. ZEISS puts the plasma at nearly 220,000 °C, and ZEISS and ASML both call it about 40 times hotter than the sun's surface. The vessel is filled with hydrogen at low pressure: it hardly absorbs EUV, and it etches away as a gas the tin that lands on the mirrors.
Mirrors all the way down.
No lens passes 13.5 nm, so every optic is a mirror, the mask included. One surface reflects almost nothing, so each mirror is a stack of about 40 pairs of molybdenum and silicon layers, each pair about 7 nm thick, and the tiny reflections add up in step. Even then, a mirror returns about 70% at best.
Berkeley Lab counts about 11 reflections in a tool: 6 mirrors in the projection optics, about 4 in the illuminator, and the mask. At 70% each, 2.0% of the light arrives. The surfaces, coatings included, have to be controlled to 0.1 nm: about half the spacing of two silicon atoms.
ZEISS, which makes the mirrors in ASML's EUV scanners, puts the smoothness another way: enlarge one of its mirrors to the size of Germany and the largest bump would be 0.1 mm high.
When one pass is not enough.
Below k₁ = 0.25 the lens cannot separate the lines. So before EUV the industry split the pattern: print every other line, etch, then print the rest in the gaps. Two masks, two exposures, and in Nikon's words of 2008 the second must land within 3 to 4 nm of the first.
One exposure at this pitch: the lens cannot separate the lines, and the resist gets a grey smear.
Pass A prints every other line at a pitch the lens can resolve. Pass B fills the gaps. Together they reach half the pitch.
Self-aligned variants grow spacers on the sides of printed lines: imec describes one lithography step whose every line ends up as four. EUV took back many of those steps, and imec writes that High-NA EUV prints in one exposure what 0.33 NA EUV would need 3 to 4 masks for.
From disc to atom.
Slide from the wafer down to the spacing between two silicon atoms. Everything in between is what one scanner pass has to control.
Every chip in your pocket passed through a scanner several dozen times. Each pass: coat, measure, align, scan, develop, etch, strip.
Scanner makers on this record: ASML, Canon, Nikon and SMEE. Photoresist makers: Fujifilm, JSR, Qnity, Shin-Etsu, Sumitomo Chemical and TOK. Photomasks and blanks: DNP, HOYA, Photronics and Tekscend.
- Sculpting in atomswhat etch and deposition do with the stencil05
- The impossible machinewhy a scanner is fast and exact at once04
- Anatomy of a fabthe building the scanner stands in02