Maskless lithography

[gbbinbum]

I was reading about Multibeam's multi electron beam lithography system, and it got me wondering: what is high-throughput maskless lithography good for? Multibeam apparently is applying it to IC security, but I wonder if it could become a cheap alternative to optical lithography.

 

[nghanayem]

Right now it is used for research and development (both in industry and academia) as they don’t need high volumes and a couple of e-beam tools are far cheaper than even a single DUV tool. Interestingly I saw an old paper that was comparing having a bank of 10 e beam tools squeezed into an easily serviceable array that was about the footprint of a low NA EUV tool. I think they said throughout for this setup could be like 180 wph. And there lies the crux of the problem. It is very hard to add more beams to a tool as they can interfere with each other, kind of permanently dooming the tech to low throughputs. Maybe if we need CDs lower than what high NA can offer and hyper NA proves to be impractical, then E beam litho can be a thing.

I’m sure Fred also knows of some other issues that I am unaware of that cause e-beam litho to have a rough time actually achieving these high resolutions.

 

[gbbinbum]

Right now it is used for research and development (both in industry and academia) as they don’t need high volumes and a couple of e-beam tools are far cheaper than even a single DUV tool. Interestingly I saw an old paper that was comparing having a bank of 10 e beam tools squeezed into an easily serviceable array that was about the footprint of a low NA EUV tool. I think they said throughout for this setup could be like 180 wph. And there lies the crux of the problem. It is very hard to add more beams to a tool as they can interfere with each other, kind of permanently dooming the tech to low throughputs. Maybe if we need CDs lower than what high NA can offer and hyper NA proves to be impractical, then E beam litho can be a thing.

I’m sure Fred also knows of some other issues that I am unaware of that cause e-beam litho to have a rough time actually achieving these high resolutions.

I see, so if they had higher throughput they could be used as an EUV alternative and even a high NA/hyper NA alternative? And the advantage is price because masks are expensive?

 

[Fred Chen]

The original attraction of maskless was to address the cases where there were too few wafers per mask. Then the expensive mask sets would be avoided since they would not be worth it.

Now they have maskless lithography systems for producing the masks themselves, like the IMS Multibeam tool: https://www.ims.co.at/en/products/

There are some alternative maskless tools with competitive resolution but they still have throughput limitations, e.g., Heidelberg NanoFrazor: https://heidelberg-instruments.com/product/nanofrazor-explore A concern here is the wear of the tip.

Throughput is very important. In case the time is many hours (as is almost always the case), then the e-beam conditions can drift over time.

Shot noise is well-known with shrinking beam spot sizes in electron-beam lithography and furthermore severely trades off against resist heating: https://design2silicon.com/wp-content/uploads/2020/08/021012_1.pdf

 

[staf]

I saw an old paper that was comparing having a bank of 10 e beam tools squeezed into an easily serviceable array that was about the footprint of a low NA EUV tool. I think they said throughout for this setup could be like 180 wph.

Yes, I also remember this paper, I think it was from TSMC. If I remember correctly this was based on the MAPPER lithography machines that were on development at that time. The plan was to use an array of hundreds (if not thousands) tips as e-beam generators. It was fascinating technology but I think in the end one of the showstoppers they stumbled upon was the afterglow of emitted electrons when a beam was turned off. Also ASML thought MAPPER would never be able to do the synchronisation between the electron beams and the scanning of the wafers.
I know ASML has also done development on optical maskless lithography using arrays of small tiltable mirrors.
But in the end the industry decided to go full in on EUV.

And there lies the crux of the problem. It is very hard to add more beams to a tool as they can interfere with each other, kind of permanently dooming the tech to low throughputs. Maybe if we need CDs lower than what high NA can offer and hyper NA proves to be impractical, then E beam litho can be a thing.

The problem is that electrons are charged particles and particles with same charge repel each other. Increasing throughput means increasing density of charged particles increasing the repelling forces. There is thus a relationship between throughput and achievable resolution with e-beam writers.

 

[Paul2]

I-line microled matrixes are there, surprisingly efficient, and they indeed been used a lot already for non-litho tasks.

Check out these fellows: jb-display.com . If micro-led matrixes would be able to do maskless 180nm, it would already be a world-changing thing, and if they manage to make a matrix on aluminium nitride to get to lower wavelengths, I think even 130nm is not out of question.

Given that the image would be changeable in real-time, and greyscale, that may be exploited to squeeze a bit more NA.

 

[Fred Chen]

I-line microled matrixes are there, surprisingly efficient, and they indeed been used a lot already for non-litho tasks.

Check out these fellows: jb-display.com . If micro-led matrixes would be able to do maskless 180nm, it would already be a world-changing thing, and if they manage to make a matrix on aluminium nitride to get to lower wavelengths, I think even 130nm is not out of question.

Given that the image would be changeable in real-time, and greyscale, that may be exploited to squeeze a bit more NA.

Micronic uses DUV-based projection maskless, but its resolution is not competitive with wafer or e-beam unfortunately. They need to shorten the wavelength to ArF with higher NA.

 

[Paul2]

Micronic uses DUV-based projection maskless, but its resolution is not competitive with wafer or e-beam unfortunately. They need to shorten the wavelength to ArF with higher NA.

I looked up Mycronic, and they only had DLP systems, but no mention of micro-led

 

[Paul2]

Nevertheless, it's impressive what they managed to achieve with 450nm light source 2 decades ago.

Sci-Hub | Resolution extensions in the Sigma 7000 imaging pattern generator. 22nd Annual BACUS Symposium on Photomask Technology | 10.1117/12.468623

The techniques used to get acceptable image out of artefact producing DLP matrix could be used on a micro-led matrix as well.

 

[Tanj]

Check out these fellows: jb-display.com .

Those are actually pretty slow, 240 Hz. TI make a DLP used for 3D printing which run around 30 kHz with a 60 gigapixel/s throughput. That is used with i-line for the 3-D printers which cure surfaces pulled out of the liquid.

If micro-led matrixes would be able to do maskless 180nm, it would already be a world-changing thing, and if they manage to make a matrix on aluminium nitride to get to lower wavelengths, I think even 130nm is not out of question.

I'm curious, what would that world-changing market be? I've talked to some litho folks about digital masks and they don't get excited at all. But maybe they are missing the point?

FWIW, I think there are ways to get down to around 50nm squares at 50 WPH, using solid state DUV sources and a MEMS modulator.

 

[Fred Chen]

Yes, I also remember this paper, I think it was from TSMC. If I remember correctly this was based on the MAPPER lithography machines that were on development at that time. The plan was to use an array of hundreds (if not thousands) tips as e-beam generators. It was fascinating technology but I think in the end one of the showstoppers they stumbled upon was the afterglow of emitted electrons when a beam was turned off. Also ASML thought MAPPER would never be able to do the synchronisation between the electron beams and the scanning of the wafers.
I know ASML has also done development on optical maskless lithography using arrays of small tiltable mirrors.
But in the end the industry decided to go full in on EUV.

The problem is that electrons are charged particles and particles with same charge repel each other. Increasing throughput means increasing density of charged particles increasing the repelling forces. There is thus a relationship between throughput and achievable resolution with e-beam writers.

It could be this paper: https://www.spiedigitallibrary.org/...formance-of-production-worthy-multiple-e-beam maskless-lithography/10.1117/12.848319.full They claimed 30 nm hp hole resolution with CAR.

There is still significant scattering with 5 keV electrons. And presumably the same stochastic effects found with EUV also apply. The shot noise alone would require a prohibitive dose.

 

[Paul2]

Those are actually pretty slow, 240 Hz. TI make a DLP used for 3D printing which run around 30 kHz with a 60 gigapixel/s throughput. That is used with i-line for the 3-D printers which cure surfaces pulled out of the liquid.

I'm curious, what would that world-changing market be? I've talked to some litho folks about digital masks and they don't get excited at all. But maybe they are missing the point?

FWIW, I think there are ways to get down to around 50nm squares at 50 WPH, using solid state DUV sources and a MEMS modulator.

Custom ASICs in production numbers without cost, and lead times of mask shops will allow non-million dollar companies to access IC manufacturing for the first time.

Things currently served by many passive components, individual FETs, and 74XXX to do thing like sequencing power-on, reset, or interlocks on a laptop motherboards can for example be replaced by cheap custom made IC, even if they will be 180nm, or 130nm.

Signal switching, RF switching, power ICs sized to individual use cases, etc.

You can make per-chip mask ROMs, doing away with huge i-fuses.

 

[SPQR54]

One of the issue with multibeams technology is that wafer throughput is also linked to the bandwidth of the transmission of data to the writing head. When you workout the numbers you see why is used for masks but not for wafers. Also, as one of my more experienced colleague told me many years ago, litho is not about patterning but about putting a pattern on top of another. Resolution without alignment and registration is useless and nobody has spent enough in developing alignment systems for e-beam platforms comparable to optical one. Look at registrations specs for any e-beam tool.

For low end the issue is that, at the margin levels they operate nobody is going to spend 1$ to change any part of a well proven, fully amortised, nearly-perfect yielding old process or invest in a new tool to capture small markets.

 

[Tanj]

One of the issue with multibeams technology is that wafer throughput is also linked to the bandwidth of the transmission of data to the writing head. When you workout the numbers you see why is used for masks but not for wafers.

If you can image a *shaped pixel, a 20nm square for Manhattan pattern, you need about 250 Gpix per cm2, each being 1 bit. 20nm squares are *roughly equivalent to what an EUV machine can do. There are 700 cm2 per 300mm wafer so if the threshold for production throughput is 50 WPH in a specialty line, or maybe in any line if they are cheap enough to buy multiple, that would require about 3 Tbps. A modern HBM3 runs around 8 Tbps for one stack. An 8.5cm2 reticle equivalent requires just under 300GB of data, so a more practical data source would be 160 LPDDR5x chips, allowing for some real ECC, and you could get those on a several DDIMMs. Not really that difficult.

*shaped pixels require new thinking. Past approaches were unworkable.
*roughly - true squares in what are inherently multiple exposure so the design rules are easier, but sometimes EUV could do specialized tricks a Manhattan pattern cannot. You might want a second ebeam machine that does round pixels when you are exposing vias.

Neither of these say it is impossible. Just hasn't been done right yet.

 

[Fred Chen]

Also, as one of my more experienced colleague told me many years ago, litho is not about patterning but about putting a pattern on top of another. Resolution without alignment and registration is useless and nobody has spent enough in developing alignment systems for e-beam platforms comparable to optical one. Look at registrations specs for any e-beam tool.

For low end the issue is that, at the margin levels they operate nobody is going to spend 1$ to change any part of a well proven, fully amortised, nearly-perfect yielding old process or invest in a new tool to capture small markets.

E-beam alignment is not shabby (already used for some advanced masks) but is made more difficult by throughput. Since the e-beam write times are long, this adds extra opportunities for drift which makes alignment correction more difficult. Another possibility would be to port the wafer scanner aligner parts over to the maskless systems. Not impossible but is it worth the disruption. Multibeam might reduce the write time enough to make it worthwhile.

Yes certainly the massive bandwidth for controlling so many beams is daunting.

 

[Fred Chen]

If you can image a *shaped pixel, a 20nm square for Manhattan pattern, you need about 250 Gpix per cm2, each being 1 bit. 20nm squares are *roughly equivalent to what an EUV machine can do. There are 700 cm2 per 300mm wafer so if the threshold for production throughput is 50 WPH in a specialty line, or maybe in any line if they are cheap enough to buy multiple, that would require about 3 Tbps. A modern HBM3 runs around 8 Tbps for one stack. An 8.5cm2 reticle equivalent requires just under 300GB of data, so a more practical data source would be 160 LPDDR5x chips, allowing for some real ECC, and you could get those on a several DDIMMs. Not really that difficult.

*shaped pixels require new thinking. Past approaches were unworkable.
*roughly - true squares in what are inherently multiple exposure so the design rules are easier, but sometimes EUV could do specialized tricks a Manhattan pattern cannot. You might want a second ebeam machine that does round pixels when you are exposing vias.

Neither of these say it is impossible. Just hasn't been done right yet.

It is hard to get squares with electrons, including secondary electrons. Overlap of two exposed rectangles is the way to avoid roundness.

 

[SPQR54]

Agreed that is doable, my point is that requires a lot of engineering resources to do it, as there are a lot of challenges that make it a risky development. And this for an even smaller market that existed before EUV.
I have not really looked into it since the demise of Mapper but I am not aware of e-beam machines at better than 25nm alignment 3sigma (it can be just ignorance on my part). At the time there where contacts with scanner manufacturers for the stage, as suggested by Fred, but they were not fruitful. Maybe things have changed but I have not seen either Nikon or Canon going that way recently and neither NU-Flare nor Leica seems interested in the high-throughput direct write market. IMS, the only ones using multibeams to my knowledge, has very early gone towards mask writers because the alignment requirement are low and the issue of registration can be managed better. KLA is sitting on a few patents that are relevant but never went beyond looking into inspection using electrons.

The problem with the throughput of data is not only on the memory side, but also on the transmission side. You have to send that flow from the memory to the writing head. For electrons you need mechanical shutters (dense multi-beams makes electrostatic deviation unworkable) to which send the appropriate on/off signal. So now you send the Tbps along wires and then, in real time you convert them in the signal to move the shutters synchronizing the beam emitters, the shutters and the alignment over the wafer. Again, doable but not without a lot of work in development and engineering. By the way, you have to balance writing head size vs max chip size to avoid again alignment and registration issues.

Mapper went bust trying and, if memory serves me well, they spent over 150M€ trying. At the time EUV was not there yet, so potential market was larger but nobody really wanted to invest or be seen working on the technology. I suspect that the fact of being seen working on anything else than EUV was considered as jeopardising that effort and none of the player could afford it. EUV 1st generation production worthy machine probably costed about 3-4B$ over 15 years to be developed. Maybe with 500M$ a multi e-beam machine today can be done, but who is going to invest those?

 

[Fred Chen]

...neither NU-Flare nor Leica seems interested in the high-throughput direct write market. IMS, the only ones using multibeams to my knowledge, has very early gone towards mask writers because the alignment requirement are low and the issue of registration can be managed better.

I would have thought NuFlare would give it a shot, if not already. The interesting thing about IMS is that it's part of Intel.