Beginning the morning in the metrology session, Yasuhiro Shirasaki of Hitachi High Technologies used the electrical behavior of a SEM to look at more than just images. As Voltage Contrast mode shows, the electrons from a SEM beam can probe the electrical properties of a point on the wafer. Here, a measurement of charge (as indicated by the energy of the secondaries) versus time was used to measure gate leakage: charge up a transistor and see how long it takes to dissipate. Toshimasa Kameda, also of Hitachi, looked at different SEM voltages and signals to try to investigate profile changes during self-aligned quadruple patterning (SAQP). A voltage of 5keV maximized sensitivity of the linescan to top profile asymmetry (by comparing the linescan midpoint between left and right edges at a threshold of 0.5 versus 0.9). Differences in the pattern depth of the different spaces was estimated using the space width divided by the graylevel of the space at a voltage of 300 V. I suspect, however, that both of these metrics are sensitive to a variety of factors, not just the SAQP profile shape.
Pushkar Sathe of NIST created synthetic SEM images that were then measured to discern the sensitivity of LER measurement to SEM Noise, SEM contrast, and feature geometry. There was nothing about this study that I liked. The synthetic SEM images were exceptionally simplified and thus not representative of real SEMs. The “LER” was in fact just a single jog in a line of various of amplitudes and lengths. And finally, the measurement of LER used image processing techniques that did not represent anyone’s best practices for LER measurement. I don’t think his results are useful.
I always try to make a point of attending talks by Ryosuke Kizu of the National Institute of Advanced Industrial Science and Technology (Japan), since they are always full of careful work and good science. The same was true this year despite it being one of too many talks using machine learning for metrology. Kizu’s goal was an ambitious one: get a good LER measurement from a single noisy image with 12 features. Did he succeed? As is usually the case with new machine learning studies, the answer is maybe. He defined three loss functions that were intended specifically for this problem, trained on a modest number of image (508), and showed decent results. However, success in the lab and success in the fab can be very different, and much more testing will be required to see if a model trained on the past can adequately evaluate an uncertain future. Machine learning is very good at interpolation, but not so good at extrapolation.
As an aside, I’m happy to report that Kizu described his approach as Deep Learning rather than AI. It has become trendy to relabel all machine learning approaches as “artificial intelligence” in order to capture a bit of the current hype and euphoria around AI. I don’t like it. If you used machine learning, call it that.
EunKyeong Jong of SK Hynix, in a talk with Applied Materials, looked at contact hole shape metrics in addition to CD in order to characterize stochastics. I’ve been promoting this concept for many years, so I am glad to see it catching on. The two metrics were called striation and triangularity, though they were not defined in the talk so I had a hard time interpreting results.
Shubhankar Das of imec gave one of many talks at this conference pushing the limits of tip-to-tip spacing using high-NA EUV single patterning and dry metal oxide resists. It is important to know how far one can push that CD without the use of directional etch, since directional etchers are very expensive! I liked one graph of his in particular, showing a nice parabolic response of spacewidth roughness to focus. There have been many talks at this conference (including by Fractilia) that indicate stochastics metrics can be better at detecting a focus drift than CD.
I learned of a new high-NA EUV stitching technique from Natalia Davydova of ASML called Block and Route. The EDA (electronic design automation) step of floor planning can be used to place the IP blocks in a chip so that the (now usually jagged) stitching region falls between these major functional blocks. This means that stitching only happens at later metal layers (wiring the functional blocks together) where small stitching errors have less impact.
I went to only one talk in the Novel Patterning Technologies conference all week, a consequence of too many parallel sessions. The last talk of that conference was by Bodil Holst of Lace Lithography (Norway), and I made a point of seeing in because Dr. Holst had reached out to me last year on her topic of metastable atom lithography. (Full disclosure – I have no financial interests in Lace Lithography, but I gave some informal advice to them about the talk. I hope the advice was worth the price – free.) This new lithography company is nothing if not ambitious. Using metastable neutral atoms of Helium (energy = 20 keV, wavelength = 0.1 nm) they demonstrated the first printing results of their prototype lithography tool. That wavelength and energy are quite nice for exposing a monolayer of resist, but the challenges are immense. How do you pattern transfer a monolayer of resist (even more challenging than the top surface imaging approaches used 30 years ago)? The mask is a silicon nitride stencil membrane, with well-known problems of manufacturability and stability (though their holography-inspired nearfield imaging approach allows for struts to be placed within the pattern). Overlay has yet to be addressed. Still, it was fun to see such an audacious attempt to move the needle on resolution by a very large amount.
In the afternoon I saw some of the talks on high-NA EUV readiness for high volume manufacturing. The bottom line: very good progress. In the last few months the first EXE:5200B was qualified, ASML’s target model for production. HVM qualification, however, is still ongoing. Marie Krysak of Intel discussed that company’s experience at replacing a three-mask SALELE (self-aligned litho-etch-litho-etch) process at 0.33 NA with a single-mask high-NA EUV print. One quote: “Random variability has replaced overlay as the largest component of total EPE budget.” She also mentioned an oft-neglected benefit of reducing line/space roughness: reduced false defects during optical defect inspection.
The last talks of the week were in the metrology conference. KLA and imec gave a talk about using a calibrated stochastics lithography simulator (PROLITH), accelerated with machine learning, to predict contact hole defectivity. One interesting outcome when simulating defectivity through focus was that best focus (minimum defectivity) was not the same for missing holes as for merged holes.
Elisa Novelli of IBM gave a talk (I am a co-author) on the importance and difficulty of measuring small contact holes. A square array of 45 nm pitch holes through dose produced a very wide range of hole sizes. Two CD-SEMs from different manufacturers were used to measure those wafers (using the manufacturer’s BKM), and then MetroLER measured hole CDs using the same sets of images. Predictably, none of the four sets of results matched very well (though MetroLER matched the two CD-SEMs the best). But one CD-SEM failed almost completely to measure holes below about 12 – 13 nm in diameter when the pixel size was 0.5 nm. This prompted a pixel size study that included trying to understand the influence of sample damage. The results were very interesting, but the message I got was very clear: Your current approach for measuring contact holes may not work as we push CDs lower (for example, with high-NA EUV). Don’t take your current metrology for granted.
Philipp Wieser of Brookhaven National Lab. looked at the measurement of resist line/space patterns using CD-SAXS and quantified the damage to the resist caused by the x-rays. Linewidth changed by 5% during the measurement and LER increased, requiring efforts to reduce the x-ray dose. Miki Isawa of Hitachi High Technologies used a combination of secondary electron (SE) and backscattered electron (BSE) images from a CD-SEM at 300 or 500V to try to detect if a contact hole is scummed. When the resist is on a spin-on glass underlayer, the BSE images clearly showed when a hole was sufficiently scummed. I doubt the same will be true for an organic underlayer, but combining SE and BSE images is an interesting option for at least some applications since those BSE images essentially come for free.
With the conference over, I can look back with some small amount of perspective. Two themes stand out to me:
- It has become accepted wisdom that scaling is now limited by EPE, and that the largest component of EPE is stochastics
- Resolution transitions (to high-NA EUV, for example) are vastly more complicated and are happening more slowly each time.
Item 2) is partially a result of item 1). The other major lesson is that AI is a huge boon for the industry and represents a tool that all companies are trying to figure out how to use to address items 1) and 2). None of this is easy, but all of it is fun (at least for someone with a twisted sense of fun like me).