In 2006 ASML delivered its first two alpha tools. Yet chip manufacturers will probably only start using EUV in production in 2018 or 2019. Why is it taking so long? The two biggest stumbling blocks explained: the source and the vacuum.
Jos Benschop’s career at ASML coincides almost completely with the journey EUV lithography has made from scientific experiment to production-ready technology. Benschop joined the machinery manufacturer in 1997, the year in which EUV LLC was founded, and all his years of effort still haven’t been rewarded with mass-produced wafers.
Yet Benschop has never stopped believing in EUV. ‘I’ve never had the idea that we were doing the impossible. The work of EUV LLC convinced me that the physics presented no fundamental obstacles, and that hasn’t changed since.’
In contrast, his colleague Hans Meiling, ASML’s vice president of EUV service and product marketing, has to admit that his faith was tested. ‘One low point in our EUV journey was about two and a half years ago, when TSMC gave a presentation that was taken somewhat out of context. People got the idea that there was still much too much to be done.’
A year later, the tide turned. Development efforts began to gain speed, ASML was reporting decent progress every few months, and since then the chill seems to have left the air. Customers have committed to using EUV in production in 2018 and 2019. Only for the most complex chip layers, it’s true, and only if ASML manages to achieve further improvements in the meantime, but still: the contrast with a few years back is significant.
It’s taken a long, long time to reach this spot of smooth sailing. Once ASML took the plunge and decided to pursue EUV, in just a few years’ time it created two alpha demo tools, which were begun in 2006 at Imec in Leuven and at the College of Nanoscale Science and Engineering (CNSE) in Albany. That was a reasonable time span in every sense of the term, but both Benschop and Meiling admit that ASML severely underestimated the next step, to production. It was the engineering, not the physics, that turned out to be the greatest stumbling block.
The light source was hurdle number one. ‘In itself, it isn’t even that hard to generate a large amount of EUV light in a laboratory setting; the difficulty is creating a source that does it day in, day out. People sometimes say, ‘Those are just engineering challenges.’ Well, you can forget the ‘just’ in this case! We’ve had to make a few serious inventions,’ Benschop says.
To produce EUV light on Earth, you need a synchrotron, but those are expensive. The only alternative is to create a very hot plasma, in which the cloud of electrons surrounding an atom has been largely stripped away. The electrons that remain are those close to the nucleus, with energy level transitions that correspond to EUV light.
This plasma state can be created in two ways: using a pulse of electric current, or using a laser. At first the electric pulse seemed preferable because it was simpler and cheaper to accomplish. Indeed, the alpha demo tools at Imec and CNSE used discharge-produced plasma sources, made by Xtreme Technologies in Aachen. Cymer in San Diego, a subsidiary of ASML since 2013, also initially focused on DPP technology.
But then the requirements changed. ‘We assumed that 40 watts of light was enough for production. But the resist turned out to be less sensitive than we’d hoped: we’d counted on 2 to 3 millijoules per square centimetre, but it was closer to 20. That meant we had to generate greater power,’ Benschop explains.
‘But then issues crop up, using DPP sources. It becomes hard to cool the machine sufficiently, because the plasma is close to several components. And the DPP plasma itself is relatively large, on the order of millimetres. That makes it harder to funnel all the light you create through the optics.’
So Cymer switched to laser-produced plasma technology, whereby a powerful laser pulse transforms fifty thousand micrometre-sized droplets of tin per second into a plasma and the emitted light is bundled by a concave mirror and focused to a point. The plasma is much closer to a point source and there’s more space around it, which reduces the cooling issues. Xtreme kept betting on a DPP variant, but gave up the fight when ASML acquired Cymer.
LPP has its own challenges. Even just reliably ejecting fifty thousand droplets of molten metal is complex. Hitting the falling droplets is doable, but how much EUV light that generates is another thing. To convert as much laser energy as possible into EUV radiation, the droplet first has to be hit with a relatively weak laser pulse to flatten it into a pancake, before the second laser instantaneously vaporizes and ionizes it. Add to that the fact that gas flows and shock waves disrupt the droplets’ trajectory.
‘Hitting the droplets is fundamentally a problem that’s a perfect fit for our competency: a mechatronic feedback system. But it really is an entirely different animal than positioning a wafer. We understand it now, but it’s taken a lot of time and effort to get it working,’ Meiling admits.
Keeping the source optics clean has proven so difficult that ASML still doesn’t fully understand all that’s involved. ‘Five years ago we thought we knew how it worked, but we had a lot less data back then,’ says Meiling. ‘Only now can we say that we understand it reasonably well – but not entirely. Converting scientific knowledge into a product that always works, twenty-four hours a day: that’s an engineering feat of massive proportions.’
The difficulty is that the spent tin shouldn’t precipitate onto the mirror, because that would decrease reflectance, which leads to undesirable machine downtime to clean or replace the collector. Thanks to a curtain of hydrogen and other tricks, ASML has been able to keep the collector clean for three months, but it has to do even better. ‘Replacing the collector once a year is currently an attractive goal,’ says Meiling. ‘For the customer, by the way, predictability is as important as availability: better a lower availability combined with schedulable downtime, than going for maximum availability and having the sources break down at unplanned times more often.’
To date, chip makers have concerned themselves primarily with process development, whose few hundred wafers per day don’t demand that the machine operate at top productivity. Next year, however, they’ll start ramping up to production. That means the average number of wafers per day must increase. Customers have already been able to achieve the desired fifteen hundred wafers per day over a longer time span, but that number must be systematically achieved. And availability is also a concern; next year it must rise from 80 to 90 per cent.
A second major stumbling block has turned out to be the required vacuum in the machine – all gases absorb too much EUV light. Working in a vacuum is an art form, where other rules hold sway, certainly when fairly aggressive electromagnetic radiation is also roaming the space. The materials used must inherently adsorb few gases on their surfaces, because those gases desorb in a vacuum. Plastics often contain solvent residues and other volatile compounds that ‘outgas’, and over time they become brittle. Grease – a proven aid in making connections airtight – and glue cannot be used in the ultra-clean vacuum of a scanner. Purity is a must, because any debris can damage the six EUV mirrors and the photomask.
‘Before we started developing EUV, we barely had any experience with vacuum, and though parts can generally keep functioning according to the same principles, they have to be implemented differently. The choice of materials is very important. Now we understand vacuum quite well, but it’s taken time to train our engineers. The alignment system in our XT and NXT scanners is now in our EUV systems, too, for example: conceptually the same but completely redesigned,’ says Meiling.
Benschop adds, ‘These are the kinds of problems that you never for a second think are unsolvable, but it takes everything you’ve got to get it all on track.’
And sometimes you see the problem, Benschop says, but you attack it too aggressively. ‘We posed pretty strict outgassing specifications on the resist. Then you need devices to measure that. There are four of them in the whole world, and there’s always someone who wants to test the same pot of resist on all four.’ The outcome? All four results were different.
‘So then we took a good look at our assumptions, and we dialled back our requirements,’ Benschop says. ‘Suddenly our suppliers could make a more sensitive resist, and we could make do with less light on the wafer. Without any problems!’
Essentially, Benschop says, ASML was driving development with the hand brake on. ‘I sometimes forget to take off the hand brake, and then it starts to stink when you drive. Once we’d taken the hand brake off, we were suddenly driving a lot faster. There was every reason to worry about an outgassing resist, but we were maybe too cautious.’
Around the turn of the century, the semiconductor industry thought it would need a successor to 193-nm lithography in 2005 or 2006 for the production of 90-nm and certainly 65-nm chips. That turned out to be an overly pessimistic estimate: immersion lithography – which hadn’t even been well explored back then – turned out to make more things possible than everyone had expected. This technology pushed the need for EUV back to the 32-nm node, situated in 2010.
ASML long maintained that it would meet that deadline. In 2008 it said that volume production would be ready in 2010. In 2010 that estimate became 2012. Meanwhile, chip manufacturers had no choice but to construct the densest layers in two or more exposures – a time-consuming and expensive manoeuvre. The number of layers that need multiple exposures grows as chip generations pass, as does the number of exposures per layer.
That’s nothing less than an existential threat to the industry, because at some point developing the next generation is no longer cost-effective. Nonetheless – and this must be emphasized – the damage done by the multipatterning era is limited. Thanks in part to ASML’s release of ever-faster immersion scanners, the pace of scaling has only been slightly delayed.
And yet: ASML didn’t have a good handle on EUV. ‘We were too aggressive with our timing,’ both Benschop and Meiling now admit. That damaged the company’s credibility and the industry’s faith in EUV. ‘We kept saying at conferences that next year we’d have a source with this many or that many watts, and then it didn’t happen. People started to laugh about it, and I completely understand that,’ Benschop says.
In the last few years, ASML has set and communicated more realistic goals – and these are being met. Thanks not only to evolving insight, but also to serious repositioning. In 2012 the company took full control of solving the source issue by acquiring Cymer, and that same year Intel, Samsung and TSMC bought a stake in ASML to cofund EUV R&D. It was a move to organize heavy-duty commitment, which now seems to be paying dividends as one objective after another gets met. EUV production is dawning on the horizon, even if not all the experts are convinced yet.
Benschop and Meiling shrug at those last remnants of scepticism. All that worries them is what their customers and the rest of the supply chain think – and they’ve been mobilized. The doubt that plagued Meiling two and a half years ago has even made way for bravado. ‘If it were easy, someone else could do it, too,’ he winks.