picoradians

8 views
Skip to first unread message

Dave Typinski

unread,
Jan 6, 2026, 2:10:45 AMJan 6
to howthi...@googlegroups.com
Veritasium made an excellent piece about how ASML's EUV lithography chip fab
machines work.

Soft X-rays, picoradian pointing accuracy and the flattest mirrors in the solar
system.

The engineering behind these machines is mind blowing.

https://www.youtube.com/watch?v=MiUHjLxm3V0

--
Dave


Paul Koning

unread,
Jan 6, 2026, 10:54:35 AMJan 6
to howthi...@googlegroups.com
I've studied bits of their work for a while now. It started when I heard from one of their technologists, who was a student of my father in the early 1980s.

The analogy I've come up with is that their accuracy amounts to shooting a rifle at the moon and hitting a man-sized target there.

The accuracies they deal with imply very accurate temperature control as well (millikelvins). And part of what makes it mind-blowing is that it is also done at high speed; I forgot the precise numbers but positioning involves accelerations of well above 1g, executed several times per second. That's a key property because these are manufacturing production machines, so the job is not just accuracy but accuracy at speed.

BTW, the mirrors are made by Carl Zeiss.

paul
> --
> You received this message because you are subscribed to the Google Groups "howthingswork" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to howthingswor...@googlegroups.com.
> To view this discussion visit https://groups.google.com/d/msgid/howthingswork/695CB573.9020200%40typnet.net.

Dave Typinski

unread,
Jan 6, 2026, 12:14:11 PMJan 6
to howthi...@googlegroups.com
Indeed, with such accuracy and precision, /everything/ is a thermometer.

According to that video: hitting a dime-size target on the Moon while bits of
the machine undergo 20 gees acceleration and physically oscillate forth and back
across several cm many (tens? hundreds?) times per second.

They figured out how to get a chaotic stream of molten tin droplets to
self-organize, on the fly, into a regular stream of droplets -- all while moving
at 250 km/hr and at 50,000 droplets per second -- and the laser that excites the
tin drops to emit X-rays never misses a single droplet.

They back-calculate the X-ray diffraction pattern produced by the reticle (the
X-ray "mask") and build opposite errors into the reticle so that the image
formed is more accurate.

Yep, Zeiss makes the mirrors -- with surfaces that are smooth to around four
silicon atom diameters across a 50(?) cm mirror.

Their machine can produce up to 100 layers on the same die with the total
stack-up of positioning errors for all layers combined held to a single
nanometer -- the diameter of a couple atoms -- all while the mask is undergoing
the oscillation and accelerations mentioned above.

Most importantly -- as you said -- that machine can do /all/ of that 24x7x365,
producing enough (hundreds?) of high-end CPU chips per hour to make its half a
$billion cost economically viable.
--
Dave

Paul Koning

unread,
Jan 6, 2026, 4:33:55 PMJan 6
to howthi...@googlegroups.com


> On Jan 6, 2026, at 12:14 PM, Dave Typinski <dav...@typnet.net> wrote:
>
> Indeed, with such accuracy and precision, /everything/ is a thermometer.
>
> According to that video: hitting a dime-size target on the Moon while bits of the machine undergo 20 gees acceleration and physically oscillate forth and back across several cm many (tens? hundreds?) times per second.
>
> They figured out how to get a chaotic stream of molten tin droplets to self-organize, on the fly, into a regular stream of droplets -- all while moving at 250 km/hr and at 50,000 droplets per second -- and the laser that excites the tin drops to emit X-rays never misses a single droplet.
>
> They back-calculate the X-ray diffraction pattern produced by the reticle (the X-ray "mask") and build opposite errors into the reticle so that the image formed is more accurate.
>
> Yep, Zeiss makes the mirrors -- with surfaces that are smooth to around four silicon atom diameters across a 50(?) cm mirror.
>
> Their machine can produce up to 100 layers on the same die with the total stack-up of positioning errors for all layers combined held to a single nanometer -- the diameter of a couple atoms -- all while the mask is undergoing the oscillation and accelerations mentioned above.
>
> Most importantly -- as you said -- that machine can do /all/ of that 24x7x365, producing enough (hundreds?) of high-end CPU chips per hour to make its half a $billion cost economically viable.
> --
> Dave

I'm not sure the accuracy numbers are quite THAT good. I remember looking for specs and finding stuff like 10 nm accuracy -- that is across machines, so you can make a wafer using several machines and things stack up that accurately. For today's wafer sizes (300 mm) that translates to 3 parts in 10^7, so I round it down to one part in 10^8. The moon is 3*10^8 meters away, so that in turn means an actually slightly larger than man size target.

As for oscillating, it's the wafer that moves, not the mask.

The TCE of silicon is about 3*10-6/K so holding things to one part in 10^8 means temperature control to 3 mK, or actually better of course because you want temperature issues to be just part of the total error.

Oh yes, part of the fun is that all the machinery is in a hard vacuum, and as we know temperature management in vacuum is not a pleasant problem at all.

It's interesting that ASML is the only outfit making EUV machines. It's fairly clear that they have a natural monopoly, 30+ years of built-up expertise that simply no one else can replicate. In particular, it's such an inherently hard problem (with such a large number of elements to it) that the usual IP thieves haven't gotten anywhere and I expect aren't likely to anytime soon.

paul

Paul Koning

unread,
Jan 9, 2026, 4:21:40 PMJan 9
to howthi...@googlegroups.com


> On Jan 6, 2026, at 12:14 PM, Dave Typinski <dav...@typnet.net> wrote:
>
> ...
> Yep, Zeiss makes the mirrors -- with surfaces that are smooth to around four silicon atom diameters across a 50(?) cm mirror.

I wonder if the mirrors are the hardest part of the machine. One additional complication on the mirrors is that they are, I'm pretty sure "grazing incidence" -- the light strikes it at a steep slant because X-rays go right through a mirror if they hit straight on. And while EUV as used in these machines is technically not quite X-rays, it's close enough that considerations get similar.

You can see this in the diagram of the optical path, which shows up in some of the graphics on the ASML website.

BTW, there was a fun article in the Wall Street Journal a year or so ago about a young woman who is an ASML field service technician. It discussed her training (9 months or so I think) and the fact that being reasonably slender and flexible is a job requirement because some of the service operations require climbing inside the machinery.

paul

Paul Koning

unread,
Jan 9, 2026, 5:28:20 PMJan 9
to howthi...@googlegroups.com
I'm starting to watch it. Nice!

Just got past the part where they talk about synchrotrons as x-ray sources. I think IBM actually built one for lithography purposes, in the 1990s or thereabouts.

A high school classmate of mine did his Ph.D. on x-ray diffraction crystallography (of proteins, which are difficult subjects) where the x-ray source was a synchrotron at a lab in England.

paul

> On Jan 6, 2026, at 2:10 AM, Dave Typinski <dav...@typnet.net> wrote:
>

Dave Typinski

unread,
Jan 9, 2026, 5:34:29 PMJan 9
to howthi...@googlegroups.com
On 1/9/26 16:21, Paul Koning wrote:
>
>> On Jan 6, 2026, at 12:14 PM, Dave Typinski <dav...@typnet.net> wrote:
>>
>> ...
>> Yep, Zeiss makes the mirrors -- with surfaces that are smooth to around four silicon atom diameters across a 50(?) cm mirror.
>
> I wonder if the mirrors are the hardest part of the machine. One additional complication on the mirrors is that they are, I'm pretty sure "grazing incidence" -- the light strikes it at a steep slant because X-rays go right through a mirror if they hit straight on. And while EUV as used in these machines is technically not quite X-rays, it's close enough that considerations get similar.

Looks like the angle of incidence isn't that large if the diagram here is to be
believed. I'm not sure it's accurate -- the video mentions six mirrors, but the
diagram shows eleven.
https://www.asml.com/en/technology/lithography-principles/lenses-and-mirrors

Creating all 100+ alternating layers of Mo and Si each to uniform thicknesses
such that the ~22 PHz emission reflected from each layer boundary is all emitted
from the stack in phase sounds quite challenging.
--
Dave


Bob Reite

unread,
Jan 9, 2026, 7:54:25 PMJan 9
to howthi...@googlegroups.com
Go to https://www.youtube.com/watch?v=MiUHjLxm3V0 for a detailed
presentation of the serous engineering challenges of building a EUV machine.
--
Robert D. Reite
Telecentral Electronics, Inc.

Paul Koning

unread,
Jan 10, 2026, 4:18:31 PMJan 10
to howthi...@googlegroups.com
My bad memory, I actually saw that diagram and somehow remembered it as grazing incidence mirrors. I do know that they have been used in X-ray telescopes (which are carried on satellites to avoid the absorption by air).

I know about multilayer dielectric mirrors, they are commonly used in HeNe lasers to get reflection coefficients higher than what metallic mirrors can offer. The reason is that the gain of HeNe lasers is rather low, so single mode versions (which have to be short, 150 mm or so max) are rather marginal and need reflection coefficients around 99% or so.

But doing multilayer mirrors at X-ray wavelenths is yet another mindboggling achievement. I liked the bit about sputtering on a layer, then zapping it with an ion beam to rattle the atoms in place and make the whole thing smooth. Needing to worry about atomic size unevenness is a rather amazing notion.

paul


Paul Koning

unread,
Jan 10, 2026, 4:28:16 PMJan 10
to howthi...@googlegroups.com


> On Jan 6, 2026, at 2:10 AM, Dave Typinski <dav...@typnet.net> wrote:
>
At 32:03 they show a physics paper that was relevant to the design, from 1950. One of the authors is John von Neumann. I assume that's the guy best known for creating some of the first computers.

paul

Paul Koning

unread,
Jan 11, 2026, 1:24:40 PMJan 11
to howthi...@googlegroups.com
I sent that link to a former colleague whose daughter is engaged to an ASML worker who is about to head off to Holland for a two year stint there.  He sent her another link with a different but also quite well done description.


paul

On Jan 6, 2026, at 2:10 AM, Dave Typinski <dav...@typnet.net> wrote:

Paul Koning

unread,
Jan 12, 2026, 8:32:58 PMJan 12
to howthi...@googlegroups.com


> On Jan 9, 2026, at 5:34 PM, Dave Typinski <dav...@typnet.net> wrote:
>
> ...
> Creating all 100+ alternating layers of Mo and Si each to uniform thicknesses such that the ~22 PHz emission reflected from each layer boundary is all emitted from the stack in phase sounds quite challenging.

I've referred to the red HeNe line by its frequency (476 THz if I remember right) on a number of occasions, but this is the first time I've come across a petahertz.

The 476 THz came up in connection with work done by my father's PhD student, who worked on stabilized lasers and verified their long term stability by building several of them and capturing the difference frequency over the course of several days. The difference frequency that I remember was around 1 kHz, which translates to about 2 parts in 10^9, pretty amazing.

paul


Dave Typinski

unread,
Jan 12, 2026, 9:50:54 PMJan 12
to howthi...@googlegroups.com
On 1/12/26 20:32, Paul Koning wrote:
>
>
>> On Jan 9, 2026, at 5:34 PM, Dave Typinski <dav...@typnet.net> wrote:
>>
>> ...
>> Creating all 100+ alternating layers of Mo and Si each to uniform thicknesses such that the ~22 PHz emission reflected from each layer boundary is all emitted from the stack in phase sounds quite challenging.
>
> I've referred to the red HeNe line by its frequency (476 THz if I remember right) on a number of occasions, but this is the first time I've come across a petahertz.

It seems like science in general has slowly been moving from quoting wavelength
to quoting frequency. Still more than a few wavelength adherents, though. It's
good to have a passing familiarity with both.

Quoting in terms of frequency often makes sense for electromagnetic stuff.
While wavelength will change with the velocity of propagation, frequency remains
constant for most Earthly applications. Most people -- ASML included -- don't
need to account for changes in emission frequency due to gravitational time
dilation, relativistic Doppler shift, or the Hubble constant.
--
Dave


kins...@gmail.com

unread,
Jan 16, 2026, 7:03:42 AMJan 16
to howthingswork
I have gone through about half of that first video, and it is excellent.  I've taught all the basic optical topics and learned VLSI techniques when feature sizes were 2 microns square, which I thought was really small.  Nope.  One thing I really like about the graphics in this discussion is that it shows how basically crude the finished product is.  That is, those transistors and connection wires are pretty fuzzy, as one would expect given some thought.  

Wow.  Looks like the noble Krell designed that machine.
Reply all
Reply to author
Forward
0 new messages