Why EUV lithography blasts tin droplets with lasers

The most demanding layers of leading-edge chips are patterned by a machine that, tens of thousands of times per second, vaporizes a falling droplet of molten tin with a high-power laser. The setup is absurd — and nothing else makes 13.5nm light in a production tool.

The picture version

Six pictures for a reader who has never wondered how a chip is printed. The prose below fills in the seams the pictures skip.

Why it exists

You’ve held a phone up to a projector screen and seen how the image blurs if the projector is out of focus, or watched a cheap laser pointer make a fuzzy dot instead of a sharp one. Chip-making is that same problem, run at an absurd extreme: a machine projects the picture of a circuit through a mask onto a silicon wafer, like a slide projector throwing an image on a wall — except the “picture” has features smaller than a virus. (Where the projector picture breaks: a slide projector magnifies, and a lithography scanner shrinks — the mask is about four times larger than the image it prints. Everything else about it is worse than a projector in every way you’d expect and several you wouldn’t.) Everything strange about EUV lithography follows from one number: the wavelength it prints with. Finer lines demand shorter light, so the industry had to move — but how far wasn’t dictated by the transistors. 13.5 nanometres is a point the industry landed on, and every absurdity in the machine is the price of that number.

A recent phone’s CPU is built on what the industry calls a 5, 4, or 3 nanometer node — marketing names by now rather than literal measurements, but the real features on the chip are still smaller than a single virus particle. The catch is the wavelength. To resolve features that small, the light has to be comparable to or smaller than the features. DUV at 193nm — the workhorse for two decades — can be pushed a remarkably long way with multi-patterning (7nm-class logic was shipped on it), but each extra patterning pass costs another mask, another exposure, and more yield risk. The next step down on the spectrum that anyone could plausibly engineer was extreme ultraviolet at 13.5nm: roughly fourteen times shorter than DUV.

Then the problems start. There is no practical laser that emits coherent 13.5nm light — research-scale EUV and soft-X-ray sources exist, but nothing you could put in a production tool. There is no transparent material at 13.5nm — every glass, every plastic, every gas, even air, absorbs it. Lenses are out. Even the mirrors don’t work as ordinary mirrors; aluminium reflects almost nothing in this band. And whatever source you build has to throw enough photons per second to expose hundreds of wafers an hour, because chips have to be cheap.

The solution that ASML and its supplier Cymer landed on, after roughly two decades of development, is the kind of thing you’d write off as a joke if a sci-fi novel proposed it. Drop a tiny droplet of molten tin into a vacuum chamber. Hit it with a low-power laser pulse to flatten it into a pancake. Hit the pancake with the main CO₂ pulse, far harder. The tin vaporizes into a plasma so hot it glows in the EUV band — including, importantly, a useful peak right around 13.5nm. Catch that light with a stack of multilayer mirrors, bounce it through the system, and project the image. Repeat the whole sequence roughly fifty thousand times per second — the figure ASML and Cymer materials consistently quote, though the exact rate varies by system generation.

That’s how the most demanding layers of leading-edge logic are patterned today. Less critical layers on the same chip still go through DUV, because an EUV exposure is expensive and you spend it only where you must.

Why it matters now

Without EUV, leading-edge logic would lean far harder on DUV multipatterning, at worse cost, cycle time, and yield. The roadmaps that take logic from 7nm down to 5, 3, and 2nm — the nodes that current AI accelerators, modern phone SoCs, and the latest CPUs ride on — all assume EUV. The performance and density story behind everything from GPU training clusters to the M-series Mac in front of you is downstream of EUV being a working production technology, not just a research project.

It also matters geopolitically. ASML, a Dutch company, is the only manufacturer of EUV scanners in the world. Each machine is a bus-sized vacuum system that costs in the low hundreds of millions of dollars and ships disassembled across many freight containers and multiple cargo aircraft. Export restrictions on these machines are now a primary tool in the U.S.–China semiconductor dispute precisely because there’s no second source: without an EUV scanner you can still make advanced logic on DUV with enough multi-patterning, but you pay for it in masks, cycle time, and yield — which is another way of saying you can’t compete on cost at the leading edge.

The short answer

EUV = 13.5nm light + tin-droplet plasma source + mirror-only optics in vacuum

Picture to keep: a slide projector where the bulb is a droplet of metal being vaporized by a laser fifty thousand times a second, every lens has been replaced by a mirror because glass would swallow the light, and the whole thing sits in a vacuum because air would swallow it too.

The tin droplet exists because no laser produces 13.5nm light directly — but a hot, dense plasma does, and tin happens to have a strong emission line right at the wavelength the multilayer mirrors are tuned for. The mirrors-only optics exist because no material is transparent at 13.5nm. The vacuum exists because air absorbs it. Each strange ingredient is forced by the wavelength, and the wavelength is forced by the feature size.

How it works

Follow one droplet of tin from the reservoir to the wafer, and every absurd part of the machine turns out to be the fix for the previous part’s failure.

The source: every obvious approach is unavailable

Naive attempt: buy a 13.5nm laser. There isn’t one. Lasing requires a population inversion in some medium with a transition at your wavelength, and nobody has built a practical one there. Synchrotrons and free-electron lasers produce EUV, but neither fits in a fab. So there is no source you can simply buy and bolt on.

Fix: don’t emit it, glow it. Anything hot enough radiates across a broad spectrum, and a plasma hot enough radiates in the EUV band. Tin is chosen because its ionized states emit strongly right around 13.5nm — which is also where the mirrors (below) can be tuned to reflect. The two choices are locked together: the wavelength is a compromise between “what a plasma emits brightly” and “what a mirror can reflect at all.”

But: how do you make a plasma that hot, that often, and that cheaply? You need it 50,000 times a second, from a target that has to be replaced every shot, without wrecking the chamber. That constraint is what produces the falling-droplet design.

A reservoir of molten tin sits at the top of the source vessel and drips droplets — each on the order of tens of micrometers across — into a vacuum chamber. The droplets fall in a precisely timed stream toward a focal point.

When a droplet reaches the right position, two things happen in fast succession:

1. A pre-pulse from a smaller laser hits the droplet and flattens it into a thin disk roughly perpendicular to the main beam. This dramatically increases the surface area the main pulse will see — without this step, the main pulse couldn’t deposit enough energy fast enough to make a useful plasma.
2. The main pulse — a multi-kilowatt CO₂ laser firing in nanosecond pulses — hits the disk and ionizes it into a plasma at temperatures comparable to the surface of the sun. The plasma’s emission spectrum has a strong peak in the EUV band; among the photons it throws off, a useful number land near 13.5nm.

This sequence repeats about 50,000 times per second. (I’ve seen this figure cited consistently in ASML and Cymer materials; the exact rep rate depends on the system generation.) The 13.5nm photons are collected by a “collector mirror” at the back of the source chamber, focused into a beam, and passed into the rest of the optical system.

The byproducts are a problem. Each pulse spits tin debris into the chamber — droplets that didn’t fully vaporize, plasma ions, neutral tin atoms — that would coat the collector mirror and ruin its reflectivity within hours. Mitigating this is its own field: hydrogen gas flow to sweep ions, magnetic and electrostatic fields to deflect them, sacrificial layers, and constant cleaning routines.

The optics: each fix creates the next problem

Now you have 13.5nm photons. Naive attempt: focus them with a lens, the way every other projector does. Dead on arrival — every transparent material absorbs at that wavelength. Glass, plastic, gas, air. There is no window.

Fix: use mirrors instead. Also dead on arrival: at 13.5nm the reflectivity of bulk metals is near zero. An aluminium mirror is, at this wavelength, roughly a black surface.

Real fix: build a mirror out of interference. The trick is a multilayer Bragg reflector: alternating layers of molybdenum and silicon, each only a few nanometers thick, deposited with sub-angstrom precision. Light reflects weakly from each boundary; if you size the layers right, those weak reflections add constructively for 13.5nm photons hitting at the right angle. A well-made Mo/Si multilayer mirror reflects roughly 70% — which sounds fine until you remember that an EUV scanner uses around 10 mirrors in series, and 0.7^10 is about 3%. Most of the 13.5nm light you generated never reaches the wafer. Brute power at the source compensates.

The whole light path runs in vacuum. Even nitrogen absorbs EUV strongly enough to matter; air is opaque.

The mask: the same problem, one more time

Naive attempt: use a normal photomask — a transparent plate with an opaque pattern on it, the way lithography has always worked. But we already established there is no transparent material at 13.5nm, so light cannot pass through a mask any more than it can pass through a lens.

Fix: make the mask a mirror too. The mask is itself a reflective Mo/Si multilayer with the chip pattern laid down in an absorbing material on top. EUV bounces off the unpatterned regions and is swallowed by the patterned ones, so the reflection is the image. That image is projected onto the wafer through a stack of demagnifying mirrors (typically 4× reduction), where a photoresist sensitive to EUV photons records it. From there, normal lithography resumes: develop the resist, etch, deposit, repeat.

Which also explains a headache with no clean fix: because the mask is reflective, the light has to arrive at an angle and leave at another, and any speck of contamination on it is baked into every wafer it prints. Masks for conventional lithography are protected by a transparent cover (“pellicle”), and EUV pellicles do exist — but they have to be ultra-thin membranes that the light passes through twice, so every one of them costs you some of the source power you worked so hard to generate.

Why this is so hard to replicate

If you wanted to start a competing EUV scanner company today, you’d need to independently solve: a stable >200 W EUV plasma source with manageable debris, multilayer mirror coatings to ASML’s tolerances, a vacuum optical column that maintains nanometer-level alignment under thermal load, a wafer stage that positions to a fraction of a nanometer while accelerating at multiple g’s, a mask handling system that keeps reflective masks defect-free in vacuum, and the systems integration to make all of it work as a production tool with high enough throughput and uptime that fabs will actually buy it. ASML had two decades of head start, billions of dollars of R&D, and Cymer (the source supplier they eventually acquired) doing nothing else. There’s a reason there isn’t a second source.

You started with EUV = 13.5nm light + tin-droplet plasma source + mirror-only optics in vacuum. What did this post add? — + every term after the first is a forced consequence of the first. Nobody chose molten tin, or megawatt CO₂ lasers, or ten mirrors that throw away 97% of your light. Those all fell out of the wavelength. And the wavelength itself wasn’t free either: scaling demanded something far shorter, and the standard account is that 13.5nm is simply where a workable mirror coating and a workable plasma source happened to overlap. Pick a point on the spectrum and physics hands you the rest of the bill.

What isn’t public

The “50,000 droplets per second” figure comes from public ASML and Cymer materials — it’s the right order of magnitude and the number you’ll see quoted, but specific systems run at different rates and ASML doesn’t publish detailed specs for current generations. The same caveat applies to source power (often cited around 250 W for production systems) and mirror reflectivity (~70% per mirror is the textbook figure).

I’m also glossing over High-NA EUV, the next-generation system that increases the numerical aperture for finer resolution. ASML delivered its first High-NA system in late 2023 for process development, with high-volume production use expected around the middle of the decade. ASML’s own announcements are the place to check which system went where and when. The basic source mechanism is the same, the optics are bigger and even more demanding.

Famous related terms

• DUV — DUV ≈ 193nm light + immersion fluid + multi-patterning — the previous generation; still used for non-critical layers even on EUV-patterned chips because it’s cheaper.
• Multi-patterning — multi-patterning = expose the same wafer multiple times with offset masks to print finer features than the wavelength allows — the trick that kept DUV alive past its theoretical limit, and what EUV partly displaced.
• High-NA EUV — High-NA EUV ≈ EUV + bigger mirrors + smaller field — the next generation, ~2nm-class nodes, shipping in the mid-2020s.
• Power wall — adjacent story: chip speed hit a thermal wall, chip density hit a wavelength wall, both around the same era.
• Why GPUs took over AI — the chips EUV makes possible are most of why frontier AI training is feasible at all.

Going deeper

• The SPIE Advanced Lithography + Patterning conference proceedings, where ASML, Cymer, imec and the fabs publish their own source and optics papers — for “what are the real source powers, mirror reflectivities, and throughput numbers?”, answered by the people who measured them rather than by a press release.
• Chris Mack’s Field Guide to Optical Lithography — for “why does wavelength set resolution in the first place?”, the optics that everything above is an extreme case of.
• Rabbit hole: Chris Miller’s Chip War (2022) — for “how did one Dutch company end up as the single point of failure for advanced computing?”, the industrial history rather than the physics.