How did the ASML lithography machine reach a dead end step by step?

As the first person to serialize the "Lithography Machine War" series online and translate and publish the ASML biography "Lithography Giant", I have been avoiding this topic for the past two years. On the one hand, the threshold for in-depth technical discussions is getting higher and higher, and on the other hand, we are really afraid of convincing people with arguments.

Yesterday, ASML announced that co-presidents Wen Nink and Vanden Brink will retire next April, which suddenly made me want to make a small footnote to a little bit of history in the past.

Friends who have watched "Lithography Giant" must still remember Martin van den Blink, who has been the technical leader of ASML for thirty years.

In an interview last year, Vandenbrink hinted that the High-NA EUV lithography machine (NA=0.55) currently about to be shipped may be ASML's last generation product. Although the industry has begun to discuss Hyper-NA (hyper-NA, that is, NA>0.7), Lao Fan believes that its cost will be terrifying and impossible. The reason is probably because he can see that this generation of products has pushed partners to the limit. . (Note: NA simply describes how much light the system can collect and focus)

What about using light (rays) with a shorter wavelength than EUV? It was also rejected. Because the wavelength is further reduced, the adjustment of the reflection angle will cause the light loss to be unbearable. If the reflector on the optical path is increased many times, the lithography machine will become a big monster that is difficult to produce and transport.

Next, let’s review how ASML lithography machines reached a “dead end” step by step.

The full name of the word photolithography is Photolithography, or Lithography or Litho for short. The original meaning of Litho is a printing method that uses the principle of incompatibility of oil and water to separate text and blank space. Modern offset printing has exactly the same principle: the printing plate is placed on a roller, and the non-image and text parts on the roller are hydrophilic and the image and text part are oleophilic (ink). For color printing, the four colors of CMYK inks are loaded onto the cylinder in sequence, but it is obviously necessary to ensure that each overprint must be aligned.

Please remember the word overlay, which is also one of the keys to driving the photolithography machine crazy step by step. Overlay in photolithography generally refers to the accuracy of pattern alignment of different layers. The overlay accuracy of the printing machine is about 0.05mm, which is said to be enough to deceive the human eye. The overlay accuracy of the most advanced lithography machine is <1nm, which is a 50,000-fold difference.

The principles of early photolithography machines are indeed the same as printing. Patterned and non-patterned areas are separated by exposure and etching of photosensitive adhesive. Multiple exposures must be aligned, so it is called an aligner. In Taiwan, it is simply called it. exposure machine.

But if you really think that printing is that simple, you would be naive. The design of a printing press must not only carefully calculate the adsorption surface energy of the printing plate, but also consider the surface tension and rheological properties of water and ink. Moreover, the faster the printing speed, the higher the temperature, how will their viscosity and thickness change? How to ensure that different colors will not contaminate each other? Do different papers penetrate different inks differently?

We need to explain the difficulty of the photolithography machine, and we also need to understand how it handles various printing accuracy and material and temperature characteristics.

The lithography machine from Aligner to Stepper (stepper lithography machine) is an upgrade of MEMS. From exposing an entire wafer at one time to the optical head moving on the wafer step by step (Step and repeat) to expose a small square, then A simple mercury lamp was also used.

Stepper upgraded to Scanner (scanning lithography machine), because laser light sources are becoming more and more precious, and the optical head has changed from a square light field to a line scan (the same as the light principle of a copier). The technical difficulty of the Scanner's mechanical implementation increases exponentially, because this light needs to scan the mask and wafer simultaneously. Reticle is also called mask. It seems that the reticle can clearly indicate that it is not attached to the wafer.

As we all know, due to its excellent precision machinery and processing capabilities, Japan took the lead in various extremely sophisticated home appliances in the 1980s and 1990s, such as tape recorders, camcorders, Walkmans and Diskmans. Later, in order to install a coin-sized device on the first-generation iPod, Jobs For micro hard drives, you also have to go to Toshiba.

Therefore, in the era of Stepper and Scanner, Japan's Nikon and Canon relied on Japan's advantages in the precision instrument industry, and the two companies themselves were world-class in optics, quickly defeating the stupid and arrogant American companies.

However, with the advancement of Moore's Law, the size of chips is getting smaller and smaller, with dozens of layers exposed, and the challenge of alignment overlay is getting higher and higher. Japanese precision machining also has limits. Just as the Japanese consumer electronics industry suffered a disastrous failure after converting from analog to digital, software control is a dimensionality reduction attack on mechanical control. Modern lithography is an integration of Stepper + Scanner + high-precision measurement + computational lithography, etc., so it is called Lithography System (this is probably the most formal English name for lithography machines now).

We have repeatedly mentioned in previous articles that the software design of Japanese products is incredibly strange. For example, many years later, Sony Sharp's color TV remote control still has densely packed small buttons. I don’t know if Japanese culture pays more attention to tangible hardware. Anyway, Japanese software genes have always been underdeveloped.

Everyone on the Internet believes that immersion lithography is a key battle for Nikon to lose to ASML. I don’t think so. The immersion itself can be regarded as a mechanical implementation, it is not difficult, and Nikon did it very quickly.

What caused Nikon's collapse should be ASML's duplex platform TwinScan. TwinScan's lithography machine has more than one billion lines of code and countless high-precision sensors and controllers to cooperate with the software to achieve nanometer-level measurement and positioning. The Japanese have been unable to produce a duplex bench that can be accurately tested in advance. This has resulted in the production efficiency of its machines being significantly lower than ASML, and the natural error rate will be higher.

Another area that requires strong reliance on software is the so-called "computational lithography". For example, due to the physical and chemical properties of the photoresist itself and the refraction and diffraction of light itself, the denaturation pattern of the real colloid is not completely consistent with the template, so software modeling and correction become a killer. The key to industrial software is modeling, such as calculating in advance what brand of glue and what angle of light will produce the roughest lines, and then redesigning the mask, but obviously this requires a lot of historical data and algorithm libraries. The laser itself will also bring about changes in the temperature of the lens or liquid, which can be dynamically compensated through micro-mechanical linkage after calculation by the software.

Due to the Mate 60 series, everyone has been discussing DUV to produce 7nm chips recently. This thing is actually more complicated than using EUV at this stage: self-alignment is achieved by adding a CVD layer spacer, and four exposures are superimposed to complete the first layer. This overlay The accuracy needs to be around 1nm. Multiple exposures greatly increase the proportion of photolithography in the total cost: four expensive masks plus four immersion photolithography time costs. In the same way, defects will also multiply, leading to a decrease in yield.

I remember that Mr. Liang mentioned in his resignation letter in 2020 that it took more than three years to complete the magical triple jump from 28nm to 7nm research and development. This kind of jump is not due to differences in equipment, but because engineers from across the Taiwan Strait have made a miraculous transfer on the know-how of the wafer fab.

So far, we haven't started talking about EUV.

As everyone knows by now, EUV is an engineering fantasy spanning more than two decades. So, what did ASML do right to achieve such an impossible task?

This topic deserves a book. However, let’s imagine what you would do if you were the boss of a company. You would go around the world to find the top manufacturers in various fields to help you develop parts according to your high standards, right?

但是，一个机型一年卖几十台，十万个零件，核心部件全是非标定制，每个零件的订货数量少得可怜，供应商愿意么？

所以，不光得讲情怀，你要给足够的钱、足够的研发时间、足够的测试、足够的迭代改进时间…

这大概就是尼康碰到的问题。日系的供应商大多也是日系，如果裙带企业不愿意做，尼康只能降低spec要求。另外不像ASML没有退路背水一战，尼康本身还有相机、医疗仪器等大量其它产品，内部拖沓扯皮也会更容易发生。

其实日本研发人员在1980年代就开始研究EUV，同步加速器产生光源(清华方案的老祖宗)和实验室曝光纳米级线条30多年前就成功了。

氙气是后来大家一致认可的产生EUV光的方案，因为相对简单。在2000年前后有大量相关论文，包括英特尔在2004年安装的EUV实验装置也是用氙，尼康也押宝在氙，但最终还是无法解决转换效能低和污染问题。

ASML倒是老早就押宝在锡身上。锡并不是个很好的EUV方案，开始用激光击打固体锡总是产生大量碎片，而且锡片本身会阻挡掉大半宝贵的等离子体。ASML大概做了10年锡EUV，在2010年第一代EUV NXE3100上，可用功率也只有10W。这是什么意思呢？大概一小时只能生产几片晶圆，这种效率不会有人买单的。

同样，光讲情怀搞不定卡脖子的供应商。ASML在2012年走上绝路的标志，就是它把当年早些英特尔、台积电和三星购买其23%股份时承诺的研发投入17亿美元，自己再加了9亿一股脑全用在高溢价收购激光光源供应商Cymer身上了。

工程师们用激光轰击液体锡滴，但锡滴是球型的，激光接触面自然是很小的。为了尽可能达到尽可能高的转换效率，锡滴越小越好，而且最好激光击打到一个凹进去的形状里。这种想法提给工程师以后，很难想象得给他们喝多少鸡血才行。

反正最后的方案是这样的，锡滴一小滴一小滴滴下来，先用低能量激光把锡滴打变形出来凹饼状，再用高能量激光打在凹坑里产生宝贵的EUV。听起来是不是也不算难？问题是，液滴只有30微米大，每秒5万滴以时速近300迈喷出来，然后两枪激光必须每一次都要准确地第一枪打凹，第二枪打在凹坑里：每秒10万枪。这样的高效率，终于使得EUV光刻机的可用功率达到200多W，达到量产上百片晶圆目标。

在2004年ASML、尼康和佳能联合制定的EUV光源目标中，功率只有110W，可见当时大家期望都不高。但今天，ASML已经把目标定在450W了。

“我听说ASML对晶圆台启动移动的瞬间光子的浪费都感到可惜，因为EUV射线太宝贵了。为了保证产能，他们必须和时间赛跑，尽可能提高晶圆的移动速度，但台面飞快地加速和减速时，还不能产生一丝震动。” —-《和时间旅行者讨论半导体》

为了尽可能提高曝光效率，晶圆台的移动要越快越好，那么要多快呢？5个g的加速度，同时量测速度是一秒钟2万次，保证晶圆台飞一般地移动到正确的位置。那么问题来了，得配备什么样的传感器才能精准到这种程度呢？

ASML官方说，这些传感器的精度是60皮米，也就是0.06纳米。即使这样，ASML觉得还没做够，他们实现了7个g的晶圆台加速度，这样可以达到15秒处理一片晶圆，而在这15秒内要扫描曝光约100个地方。要知道晶圆台是托着12寸晶圆的大玩意，这么快的移动速度，怎么能不产生振动和热量呢？

ASML之前的TwinScan台是空气悬浮的，这样摩擦阻力可以很小。但随着芯片做到7nm以下，问题又来了，气悬的空气会随着晶圆台高速移动产生扰流，扰流会影响量测干涉仪的精度，这样就难以达到纳米级对准了。

怎么办呢？ASML咬牙把气悬浮改成了磁悬浮，这不就没空气了么，也避免了空气被加热的问题。但说起来容易，磁浮会带来超强磁场，副作用肯定也得解决。

那么，这样就可以了吗？

悲催的是，我们还没讨论最重要的光路设计呢。我想大家都看过EUV示意图。

有小伙伴问，既然EUV光线这么宝贵，为什么要反射这么多次？每次要损失近一半的光子呢，按一次50%损失反射6次就只剩2%不到了。

对，即使蔡司制作的这些反射镜是地球上最平整的平面(每个镜子有四五十层硅和钼交替的涂层，还得确保每层的厚度是EUV波长的一半)，仍然让EUV光损失惨重。

这里有好多讲究，一个是光不能随意拐弯，为了机器不是巨大塞不进飞机，光路设计要考虑空间。不考虑空间的体育记者手里的大炮相机和考虑空间的手机相机，差别是很大的。较大的入射角也是不行的，会导致更多的相差和损失。

EUV光子需要汇聚成线后先扫过光罩(也是镜子)，反射光需要缩小到1/4再扫过晶圆上橡皮大的曝光区(Field)，这个缩小过程更需要严格的对焦和光路设计，所以这些镜子并不是平面镜，而是带焦点的缩小镜。

只要是光学器件就会有缺陷，光路设计好则有可能补偿掉其中大部分。

听朋友传谣说，蔡司一开始是不想玩这个游戏的，一年生产几十套这个镜子，能赚几个钱？而且为了生产它们，需要几层楼高的超级真空腔和巨型机械手。更悲催的是做出来稍微有点瑕疵，ASML还不要。

这个谣言也许是真的，因为大约到了2015年ASML启动High-NA EUV项目时(当时Low-NA EUV还远未通过客户验证)，蔡司真的准备撂挑子了。当时还不富裕的ASML咬牙花10亿欧元买下蔡司半导体部1/4的股份，再加上承诺未来6年给半导体部拨款7.6亿欧元。

High-NA EUV系统已经是ASML能看到的一条绝路了。问题是，周围小伙伴们却认为那也许是一条死路。台积电和英特尔都对手里的ASML股票做了清仓式减持(《台积电等三巨头投资ASML的真相》)。

历史是必须要回看的，身在其中必然无法体会期间的奥秘。

第一台EUV跳票10年，确实是碰到的问题太多了。我们举个小例子：

现在的EUV是一台1.5兆瓦的功率巨兽，激光就像带着火把在森林里放火，每到一处产生的温度变化都不可避免导致器件变形变异，而在高端芯片上是错开1.5nm上下层就对不上了。

我们说过带有芯片图案的母版光罩也是镜子，这玩意大概30万美元一个，高能激光会导致光罩是有寿命的。别的镜子上有点瑕疵还好，大不了丢几个光子，而光罩上的瑕疵则直接导致芯片失效。还有一个问题是，小的杂质颗粒会掉到上面。

原来设计师的思路是，EUV光路是全真空的，根本不用考虑杂质的事。可现实情况是，鬼知道哪里来的肉眼根本看不到的小东西。晶圆厂通常只能在发现缺陷后，停机把光罩摘下来干洗或湿洗，反正是损失巨大。

有人说，不如给光罩贴个膜，发现问题撕了再贴一张不就好了。这个思路倒是一点都不蠢，居然和工程师想的一样。

But what kind of film can allow precious EUV light to enter and then be reflected back without loss? You must know that EUV can heat up more than 600 degrees when applied to it. We need to know that glass can absorb EUV, so we use mirrors. This film must be thin enough and strong enough to ensure smoothness.

It is impossible without light loss. Many manufacturers participated in the challenge but most eventually gave up. ASML conducted numerous experiments and finally chose a 50nm-thick polysilicon film, which is about 1/50,000 as thick as a women's facial mask, and can probably achieve only 10% EUV loss. In order to attract customers, ASML made this film automatic, automatically measures the impurities on the film, and then uses a robot to automatically block it onto the photomask.

Although they have strived for excellence, this 10% light loss also causes pain to the fab because it is likely to lead to a reduction in production capacity. Moreover, the lifespan of this film is only one or two days, so the wafer factory sometimes does not use it for small-size masks.

This incident shows from one side the tragedy of current lithography machines, that is, every small improvement requires a huge cost, and this cost and benefits often cannot be matched.

I don’t know how Intel’s CEO and CTO judged the situation. Even though 10nm was delayed for at least three years due to various obstacles, they still didn’t believe that EUV could be used. They had huge amounts of cash but missed the opportunity to seize Platform Semiconductor Manufacturing Co., Ltd.’s 7nm in one fell swoop.

Obviously, the smart people at Intel will not stumble in the same place twice. They decided to use High-NA EUV lithography machines earlier than TSMC and surpassed TSMC at the 1.8nm level (18A).

Unfortunately, ASML's High-NA has been delayed.

Although how many nanometer chips are now completely a marketing term, the increase in transistor density cannot be faked. Intel low-key no longer advertises that 18A is manufactured with High-NA, and can only silently use Low-NA multiple exposures whose yield is difficult to control.

In theory, High-NA is not as difficult as it was when EUV was first launched. Everything needs to be subverted and redone. So what is so difficult about it?

In order to collect more precious EUV, the most ideal situation for ASML is to greatly increase the area of ​​the photomask from 6 inches to 12 inches (the photomask is also a reflector), so that the production capacity (throughput) will be greater.

However, the wafer factory, mask factory, and testing equipment factory all voted against it, and they even disagreed with the increase to 7 inches. After all, they have to pay for such a consumable product as the mask.

But High-NA means a larger reflector, and the pressure is all left to ASML and then passed on to Zeiss. It is said that the reflector in the final optical path is more than 1 meter wide, twice as long as ordinary EUV. What's even more tragic is that the weight of such mirrors has increased sharply from 40 kilograms per piece for ordinary EUV to 360 kilograms per piece.

For such a heavy and large mirror, what kind of fixture can be used to keep the flattest surface in the world without deformation?

I don't know the specifics either. But looking at the press release, Zeiss used a huge robot to grab it in a huge vacuum chamber.

Let's go back to the most complex mirror, the reticle. Since the area is not allowed to increase, and the pattern resolution requires that the incident angle cannot be increased, we can only use mirrors with different reduction factors in the x and y axes (that is, one axis is (haha mirror), the scanning light field that finally reaches the wafer is half the size of ordinary EUV, which is considered a compromise to achieve 0.55NA.

It sounds good, but the difficulty of this high-precision magic mirror is obviously not on the same level as that of making a flat mirror. What’s more complicated is that the scanning light field is twice as small. How can the two light fields be spliced ​​at the nanometer level? The auxiliary measurement system needs to be overhauled again.

Another big trouble is that as the NA increases, the final focal depth of the light focused becomes shallower. Photoresist is a three-dimensional layer. Only with sufficient depth of focus can a photoresist of sufficient thickness absorb light energy and denature. The photoresist must be re-developed, and the flatness requirements of the wafer are higher than before (otherwise the shallow depth of focus cannot cover the surface fluctuations of the silicon wafer), which in turn involves a series of changes in the industrial chain.

High NA also requires that the laser power used to squeeze the toothpaste be increased again, and the laser that hits tin droplets is increased by more than 20% from 50,000 drops per second. However, the various energies emitted by this super-powerful beast also cause various problems that need to be solved, causing temperature deformation inside the entire machine.

In short, High-NA EUV is not a simple upgrade of ordinary EUV. This thing is almost a new design according to the extreme, so an "upgrade" took another ten years. Sadly, Hyper-NA EUV will break through every limit all over again. This is where Vandenbrink is "desperate".

ASML has produced approximately five to six thousand lithography machines in the past thirty years. What is extremely surprising is that 95% of the machines are still working normally in the wafer factory, including 1,800 units of the legendary old machine PAS5500 in "Lithography Giant". There are also countless second-hand machines imported from overseas running in China, and ASML also renovates and maintains a large number of old machines every year.

Such a business model sounds strange, right? If the old ones are not eliminated, to whom will the new ones be sold?

This is the power of the information age. The explosive growth of human information technology and storage demand for chips has given lithography machines room for continuous growth and development.

From this point of view, ASML is obviously the darling of the development of the times.

So, is it a good thing or a bad thing that its lithography machine has reached a dead end?

*Disclaimer: This article is original by the author. The content of the article is the personal opinion of the author. The reprinting by Semiconductor Industry Watch is only to convey a different point of view. It does not mean that Semiconductor Industry Watch agrees or supports the view. If you have any objections, please contact Semiconductor Industry Watch.

Today is the 3609th issue shared by "Semiconductor Industry Observation" with you. Welcome to pay attention.