# Colloquium 10/8/26 Dr. Umberto Celano

**Channel:** Department of Physics - Arizona State University
**Source:** https://www.youtube.com/watch?v=jo4IZO9gE_Y
**Transcript page:** https://www.withtranscript.ai/video/jo4IZO9gE_Y

## Chapters

- 0:01 — Introduction: ASU, Semiconductors, and Industry Context
- 9:56 — E-beam Induced Conductive AFM Without Back Contact
- 22:07 — Non-contact Modes and Integrated MEMS AFM Vision
- 27:11 — Automated 2D Material Analysis with Machine Learning
- 36:17 — Conclusions and Acknowledgments

## Transcript

**[0:01] Speaker A:** Okay, thanks for the introduction. Very kind. And it's my pleasure to be here with you and and share some of the results generated here at ASU for for most of them in the last 3 years with my group. Um Indeed, a lot of time spent in my life doing semiconductor metrology and and basically which in fact I like to often define as glorified material characterization. Effectively, it's just a lot of material characterization that we do with in mind accessing some specific information that are useful for your device um um physics understanding and as much as possible postmortem analysis and things like that. So, this is a slide that you have certainly seen but it's relevant to set up the stage here for First of all, why did I move here in first place? So, in other words, what was very attractive to me and of ASU and and the whole Phoenix area with the huge concentration of fabs.

**[1:05] Speaker A:** For the work that I do, fabs are key cuz actually there is no relevance for the work that I do without a fab. So, in other words, we are literally intertwined with our work to what happens on wafer, on chips, and and of course as the chips are manufacturing processes develop. So, this one doesn't need a lot of introduction but clearly you see that this is really a hot spot for for chip manufacturing. So, we have with with the establishment of TSMC and Intel, this is clearly a no way back event. So, these are, you know, hundreds of billions of dollar investment and after you do that, you don't change your mind very very quickly about that. So, this place and this university somehow related to that, it's certainly going to be at the epicenter of semiconductor manufacturing, semiconductor technology, and therefore also metrology, design, you name it.

**[1:55] Speaker A:** And on top of that, so we have this this beautiful building here MTW. I'm sure you're familiar with that. That's the sort of off-site campus where there are a few labs including my own labs and and there is also 300 mm clean room in some parts share with with real tool manufacturers such as applied material. So, that gives also a little bit more relevance to some of the activities that we are doing cuz there's a sort of hope to shorten the lab to fab transition in some cases and to work together with some of these manufacturers. So, in in a nutshell, what does my group do? It's it's very really really easy to to describe when you look at any of the road maps that could be proposed by your favorite IDMs which stands for integrated device manufacturing. So, in other words, the TSMC of the world, the Intel of the world, they like to project a little bit what their need and what's going to be needed in order to fulfill what they call as PPAC which

**[2:57] Speaker A:** stands for uh performance power area and cost paradigm. So, which essentially means if we can accomplish these kind of technology improvements in this amount of time, then we're going to be able to deliver this amount of performance according to what we have been doing for the last 50-60 years at this point. And what we really do is simply saying, "Okay, let's assume all this is happening. We don't care too much about the time cadence for that, but if all that happens, then what are the gaps and challenges for the material characterization?" So, if you get manufacturer a C fat which is literally two transistors one on top of each other, therefore kind of capping in half in a single shot the real estate occupation on your chip for for for one CMOS inverter which is taking two transistors, then what do you need? What kind of techniques are needed there to study carrier profiling, strain, or or simply magnetic property, conductive properties, and whatnot?

**[3:53] Speaker A:** So, if you look at the left side, you will see some of the main innovations that is currently um, you know, investigated and somehow happening also at the level of the uh, major manufacturer and these are effectively the integration of gate-all-around transistor. That means going from uh, FinFET, which is already the step that you might have heard after planar MOSFET, uh, abandoning that architecture to get an architecture which is based on nanosheets. So, that's the ultimate type of electrostatic control, as you can imagine, cuz now effectively you have some some very Uh, we're going to show some images probably, but so you have uh, the ability to wrap around your silicon channel four side of metal gate for electrostatic control. Uh, very exciting is the ability to maybe um, abandon silicon entirely at some point and use 2D materials. Again, you got to also understand these companies need to fill up a relatively long time gap over here.

**[4:52] Speaker A:** So, since they don't want to give out what they're really going to do, at some point they need to throw there some like fantasy futuristic scenarios. For example, when I was younger, there there would have been a lot of carbon nanotubes much closer in time. Now, clearly the carbon nanotube they've been pushed another 20 years. So, they maybe never going to happen. That doesn't really matter and frankly from our perspective, we don't care. Carbon nanotubes as much as 2D material are greatly very exciting material system to study, a lot of interesting physics, a lot of interesting challenges for the material standpoint. So, we love them whether they would be manufactured or not. But so, that's really in a nutshell the way my group work and and again, this is what I was referring to. We definitely can identify a couple of big big transition um, inflection points that are happening just together this year, by the way.

**[5:45] Speaker A:** On the left, uh, you heard all the incredible um, success of AI and and somehow related to that the the chips that are powering AI. So, these are very special chips in the sense that for the first time in history, now we are dealing with hundreds of dies that are all stuck together. So, for example, Blackwell, which is already something that you would consider fairly old uh in the space of uh Nvidia uh line of chips, it is having more than 100 dies. Okay, so die Each die is effectively something that you have to manufacture on a wafer and then sit on the package. Therefore, there are literally hundreds of chips in one of these uh Nvidia chips. They are based on FinFET technology and they're fairly large. They fit in the palm of your hand. And from the standpoint of metrology, for example, this is bringing a lot of innovation at the level of how do we assemble these very heterogeneously integrated systems.

**[6:37] Speaker A:** So, a lot is happening in packaging. We mentioned Shield. Um a lot is happening in the space of large field of view access. If you want to look inside your sample, how do you prepare a sample where now you need to expose tens and tens of mm squared compared to what you would historically be doing now up opening up few hundreds of micrometer squared. On the right side, there is the other very exciting inflection point. So, as we um in- produced on the left, these are the type of transistor we have been using from 2011 pretty much up until last year. These are called FinFET. Now, you don't need to uh understand much more than these, but seeing that the metal gate, it is wrapping your fin on three sides. So, that's a little slab of silicon which is protruding out of the plane and you have a metal gate which is touching on three sides and therefore providing electrostatic control on the current inside the fin on three sides.

**[7:30] Speaker A:** That was considered a in great innovation in 2011 when it was introduced. It has been scaled actually quite aggressively for many many generations and now it's been abandoned in favor of gate-all-around uh released in January

**[7:43] Speaker B:** for the first time in high-volume manufacturing for logic. In all fairness, released in October 2023 from Samsung and some of their watches. Uh the technology here it's based on a little channels and you see a cross-section cut of that transistor. And the the current is flowing in this little channel and the channel is wrapped on four sides by your metal gate. So, this is the ultimate type of electrostatic control that you can get on those channels when you want to turn it on and off by applying a bias to them. Importantly, the other incredible new architecture which has been launched in the same type of product, the product we are talking about is the 18A from Intel. It's the backside power delivery network shown over here. So, this chip it's effectively containing wires made of copper on both side of the chip. So, if you did any sort of study related to solid state electronics or or or CMOS manufacturing, for for long time we just counted on the substrate to be there, a bulky nice piece of silicon that we could

**[8:48] Speaker B:** use in many ways including for metrology that was transparent to IR light and we could access or perturbate our transistors in many way. Now, we have copper wires on both sides. So, apart from the incredible challenge in term of the manufacturing, this is providing and introducing quite a lot of challenges also for the metrology. So, these lines which are a little bit larger as you can see they are delivering mostly the power while the the wires that are still remaining on the top part are delivering the signal. So, the stuff that would deliver clock and turn on and off the gates. But, very important is the fact that right now your silicon it's literally a very teeny tiny amount of material that remains in your in your chip. So, you are not wrong by telling that less than 10% of the whole chip is silicon.

**[9:34] Speaker B:** The rest is silicon and copper. So, these two things together are creating a lot of interesting opportunities out there. One thing that we are doing in my group is indeed trying to leverage some of the atomic force microscopies and some of those methods which are based on scanning probes to generate data that are valuable for further analysis or material engineering.

**[9:56] Speaker B:** So, in here I want to give a um quick overview of this electron beam induced conductivity FM somehow the work we are doing to to make uh scale up for electrical AFM. So, this is going to be very clear, um I guess, in a second. What do I mean with scaling up? Now, this slide sounds very crowded, and again, you don't need to understand each and every image there, but it's just to give a little bit of a breakdown of how many techniques do we have out there from classic AFM just showing morphology to conductive AFM showing conductive defect to actually SMM, which is using nanoscale microwave imaging to generate maps of of difference in the impedance of a 3 GHz microwave that we're shining on the sample, down to the stuff that we use when the chip is fully processed. So, for example, the scanning capacitance, which is mapping out the carrier that you have on the silicon, the SSRM, scanning spreading resistance microscopy, often used to look at the 2D profiles and 3D profiles of carrier distribution in in

**[11:04] Speaker B:** your device. So, this is just to launch uh and to explain the impact of this technique. And so, everybody somehow uh after this slide would think that if you go into a fab and you look at one of these machines that are costing a couple of million dollar and are heavily automated, that are devoted for very high throughput wafer characterization using AFM, should be doing a ton of these electrical modes cuz there is value and they are valuable and they generate information as you characterize your wafer and so on. However, that's not the reality, and as we are um let's say, as of today, roughly like 0% of the cool stuff I showed in the previous slide can be done in fab and can be done in line on wafer. So, then you might ask yourself, is there any physical limits, you know, it's it's too big, it's too large, not possible for my scanner?

**[11:56] Speaker B:** Not really. The The real blocker is something as silly as what you see over here, and it's the fact that most of those techniques that I showed you, they are two probes technique. Very, very simple two probes experiment. One One probe must apply bias, the other must collect some charge, and therefore, nobody wanted to break your wafer, nobody wants to put some dirty silver paste, contaminant silver into the rest of the manufacturing flow. So, that stupid back contact, it's kind of hampering the use of a lot of electrical modes in a fab, okay? So, that's really the premise of the work here. And so, we decided to explore alternatives to basically get rid of this. We want to get rid of that little black back contact and and still being able to generate some sort of electrical effect and contrast. So, to do that, over the years, we came out with the idea of taking a self-sensing probe.

**[12:46] Speaker B:** This is an interesting solution. I mean, of course, it's a very well-established technology for doing atomic force microscopy. The main difference is that this one here does not require your laser optical detection system. So, the cantilever is somehow a smart cantilever. It has embedded a little piezoresistive element. And so, as we bend the cantilever, the cantilever itself can explain what goes on in term of strain and stress. And so, you can use that to remove your laser. That simplifies the setup. And with that a simplified setup, we decided to sit this guy into a FIB-SEM chamber. Could be your favorite CD-SEM. It could be anywhere where there is an electron beam. And so, we decided to try to shine an electron beam in proximity of the probe using the probe for draining charge and using the electron beam to provide and supply this charge. So, completely getting rid of the back contact as a physical object, which we discussed before it's very much not accepted in your inline manufacturing.

**[13:43] Speaker B:** So, that's how it looks like. This is our Helios 5 in here. This is a light scope 2.0, one of the tool capable of offering this kind of self-sensing AFM. And so, the real schematic you can consider is shown over here. E-beam is impinging not far. Your probe offers a path to ground. We just collect these these charges as we scan around. So, technically speaking, we just created the very same circuitry that we use for classic conductive AFM. But with something that we often refer to as remote contact. Now, the idea of this electron beam impinging is now your remote contact, let's say. It turns out that one of the first area where we applied that it's like 2D and 2D materials. For those of you working with 2D materials, it's a very much kind of perfect playground for this type of technique. In fact, whether you're working on flakes, whether you're working on on partially grown layers, whether you're working on transfer film, or whether you're working on the ideal target for for this 2D material, which is area selected growth only

**[14:45] Speaker B:** where you want to make your device. All these situations do not offer a connected path to ground. So, in other words, you would have trouble to put your back contact there pretty much no matter what, okay? And and the flake is the one that I'm going to introduce first, which is also the very simple one to understand. On top of that, we do know that there's quite some value in doing conductive AFM for 2D materials. We have experienced detecting individual defect vacancies, grain boundaries, and therefore we said, "Okay, let's see what happens on 2D materials." And so, to make this experiment very very safe, we decided to isolate entirely the sample from our chuck. We also decided to take a flake which is sitting on an insulator. So, technically there is really not a path for for your charges to go to ground. And as you would expect, there is no way you get contrast on your conductive AFM image because simply charges cannot be collected in any possible way and bias cannot be applied in any possible way.

**[15:41] Speaker B:** But, in the same area, if you start applying an e-beam on the 2D material and you just drain your charges through your probe, here you go, you get some contrast, you get some current. So, effectively you are doing C-AFM without a back contact. So, clearly we did quite a lot of work to then trying to engineer this. And the next intelligent questions you might ask is how close can I get with my e-beam to the area of interest? And here we can get in the ballpark of 10 micrometer. Actually, over the time we we managed to get a little bit closer, but then the real question is also what are you measuring? Can it be damaged by getting too close? Can you get too much parasitic current into the probe? So, there might be some limits there. But, the other interesting uh observation is that your current is kind of linearly scaling with the beam current from the electron source in the SEM, which makes total sense.

**[16:29] Speaker B:** So, we are kind of acting as a sort of current injected conductive AFM. The more charge I'm pushing in, the more I'm picking up down. And and there- therefore, you can get also way for tuning your signal to noise ratio by cranking up or down the the brilliance of your electron source over there. Some of the images that you can generate are on area selected MOCVD. Again, the question here is once I grow this 2D material just in this selected squares, how it's going to look like electrically? We do know, as I said, that electrically we can pick up individual defect grain boundaries. We have a whole block of this presentation on that. So, here I just want to show you that actually this approach work on isolated 2D material. It works on um pattern wafer. So, this is the work done in collaboration with STMicroelectronics.

**[17:19] Speaker B:** They very kindly created a pattern wafer pretty dummy, just some some silicon dioxide which has been patterned for a whole 300 mm wafer, and then molybdenum disulfide grown everywhere on the 300 mm wafer. So, here you can see that basically these two images, one it's acquired with your classic system using back contact, the other one is acquired with our system using an electron beam type of back contact or remote contact. They are pretty similar, pretty comparable, especially in term of the amount of current that you can sense. The community of 3-5 people, it's another community that would love this. They can check their defects without breaking their wafer. They can check their with defect in line. So, here now all these images are taking with E-beam CAFM. So, you can see that we can respect a very decent resolution in the ballpark of 10-15 nm for lateral features such as step terraces or or trading dislocation.

**[18:16] Speaker B:** This is an image for gallium nitride. The samples are coming out of NIST, uh, where where eventually we are comparing ECCI, which is an electron contrast channeling method. So, here what we we leverage is the the fact that the presence of a defect will produce some variation in the backscattered electrons from your SEM. So, that would be the essence of your ECCI. And and we find a very good correlation when we study that with CFAM, or better said, e-beam CFAM. I just want to point out that simply by walking around your probe in and out of the area of the e-beam impinging, now you have also a simple way to create a small TLM structure, a transfer length method structure. If I go far away and my sheet resistance start kicking in, I can compare two material and be able to quickly assess which one of the two has a higher sheet resistance, simply because that will, basically, impact on

**[19:14] Speaker B:** the amount of current that I can collect inside my probe. Now, one legacy application for CFAM, conductive AFM, is the study of thin dielectrics. Thin dielectrics, they tend to break, they tend to have, you know, stress-induced leakage current and a lot of these problems that reflect into issue with reliability of your transistors. So, historically, uh, one very big area you will find hundreds of paper where people are taking thin films, applying a bias, measuring with tip and sample, and generate maps where they can trying to pinpoint presence of conductive paths. So, here these black points are little region on the thin film dielectric where there is enough, uh, density of defect to sustain a little bit of losses in the tip-sample junction. So, this is effectively, um, a something you want to at least inspect, possibly mitigate, if you want to use this dielectric for memory or logic application.

**[20:09] Speaker B:** And basically, we managed to replicate something very similar with e-beam CFAM, which means that effectively we can probe, uh, also when when you have, um, the electron beam forced to go through the oxide, getting it to the the back contact, and then generate a voltage across the tip and sample. For for those of you who might be kind of expert of the community here, I just want to emphasize that when we do the measurements with your classic back contact, being the film relatively thick, this is 20 nanometer thick titanium oxide, which for a probe which is 50 nanometer square is effectively a very thick dielectric, by the way, we just cannot push a lot of current. So, we are in the ballpark of hundreds of picoamp when we apply 10 V. But in case of our e-beam experiment, we easily go a factor 10 above that number, which is clearly suggesting the effective voltage drop between tip and sample in our experiment is way, way higher, probably a factor 10 higher

**[21:04] Speaker B:** than this 10 V that we are applying over here. This is not a parameter we just can control because as a matter of fact, it will be anything that builds up spontaneously, but that's some kind of an interesting point to take away. Finally, the failure analysis space, so you want to be really leveraging the ability of your FIB to do milling, and so in this case, for example, you can start from a fully fabricated chip, again, create a very large area that you are delayering with laser, go inside your FIB, let the FIB do the the FIB job, which is exposing layer down to the area of interest. You can land your probe, and at this point, you can simply apply the e-beam instead of making a back contact, and generate contrast. So, this is a case where we are looking into some vertical structure, these are vertical channel of 3D NAND devices.

**[21:51] Speaker B:** And as you manufacture this 3D NAND, it's very important to understand the amount of current which is flowing uh in the vertical channel, whether any of the process condition can change that, if there is a failure, and all of that can be done uh without back contact using our approach.

**[22:07] Speaker B:** Very quickly, the stuff I discussed before um or up until now, it's all in contact. So, it's a two probe experiment, bias and and current sensing apparatus, and everything is happening in contact. So, provided you can impinge some charge, then you might expect to find something also on the other side. Now, we decided to explore if there is any chance to do something also with the with the no contact and and somehow the first one in the dump is is probably STM. Now, apologies if anybody here has ever done STM because these are very poor quality STM images. The reason is that the system is not supposed to do STM, it's still an AFM. The scanner is not as good and not as precise. However, you can clearly see that we can generate an STM kind of contrast with a gap of 10 nm um or something in the ball park of uh yeah, 10 nm um in case of copper and graphene uh

**[22:59] Speaker B:** coated on top of copper. And finally, here at the uh bottom animation, you'll see something called EFM. That stands for electrostatic force microscopy. That technique, it's a two-pass method. First pass is mapping morphology, then the second pass literally excites a resonance in your cantilever and passes on the same place with a little bit of a bias applied so that there would be some long-range electrostatic forces that will be very sensitive to location where you have either different materials or different doping. And here I show that when we do that kind of EFM pass on one of our 2D material coated sample, we do see the turning on and off of our e-beam, which is a clear indication that somehow you can sense or provide some far-field electrostatically driven contrast. So, with that in mind, uh you can consider putting all this together. So, at the end of the day, um what this stuff offers is that you can in single in a single um a measurement environment offer SEM resolution, EDX sensitivity, conductive FM resolution for electric defect, and of course, charge

**[24:06] Speaker B:** um state resolution in this case if you're looking for charge grain boundaries or things like that. All in the same place, all at once, all in one system, possibly without breaking the vacuum. And finally, um although the EFM experiment is a very weak proof that that we have some sort of time-dependent um you know, ability to apply a signal, we are exploring a couple of other options because actually applying a time dependent signal, as you see over here, it's very much key to enable a lot of other modes of operation. So, you might be familiar with KPFM, Kelvin Probe Force Microscopy, PFM, Piezoresponse Force Microscopy, scanning capacitance, all of them they do depend on the application of an AC bias. So, we would like to show or try to prove that we can apply some sort of time dependent signal in our FIB-SEM experiment so that we can mimic also all these other modes.

**[25:00] Speaker B:** We're not yet there and we are exploring a few options. In the crowd, there are a couple of students that are uh you know, endorse of the responsibility of delivering that, but we have some reassuring um data. So, for example, here we have a side with gold, a side with aluminum nitride. Our probe is sitting at the certain gap from the surface doing EFM, so it's actually excited in resonance. And you can clearly see that as we have our electron beam uh shuffled around and basically uh wiggle from from the gold to the aluminum uh gallium nitride layer, we can couple some kind of time dependent signal in our phase shift. So, again, at some point, if that's going to work, you can think about having a locking amplifier attached at that frequency and recover your type of e-beam induced AC modulation. So, it's quite far from where we are in term of the final demonstration of that, but this is one idea and that's what we're doing right now.

**[25:58] Speaker B:** Finally, I like to show this um before moving to the next topic, and that is the concept of integrating an whole AFM, an entire AFM into one MEMS. So, in fact, you know, when I show this kind of presentation, I often um want to deliver a message, and my message is not to consider integrating FIB-SEM and AFM into a 100, you know, $2 million machine, but rather if you can make an AFM like this one uh basically all integrated into a MEMS, that means your AFM is getting extremely cheap at that point. It's almost a disposable chip. Then you can think about integrating these things and doing all the modes that I described because this is the same technology. It's basically self-sensing in every tool which we have out there doing some sort of electron beam related measurement and there are plenty of them. So in our industry everything moves with SEM inspection, CD SEM, voltage contrast SEM and FIB and SEMs are everywhere.

**[26:54] Speaker B:** So my view and the reason why we're doing this this work is that effectively we want to make these two things basically so inexpensive and so integrated that you can do pretty much all I've shown you at a fraction of the cost of today in every tool which has an e-beam.

**[27:11] Speaker B:** Okay, I'm going to go quickly here and and jump in some of the application related to reducing time to data for 2D materials. In particular, I think I wanted to show some of the stuff we're doing to get rid of that back contact and now I will go back to data acquired in both ways with and without the back contact. Now the bottom line here is how do we extract information from these 2D materials? Again, it's sort of crowded slide and and here my objective is again something related to transferring a message of the value that these techniques can add. If you're interested in characterizing 2D materials, probably like STM and SPM in general scanning probe microscopies, they have no competitions out there. Whether you're interested in looking at the nucleation, shape and size of the quality of your growth, presence of cracks and defects, presence of electric defect or presence of moire pattern if you're interested into you know, the twist angle

**[28:12] Speaker B:** of of bilayer that you are putting one on top of each other. You can get them with friction, with microwave, with current, with capacitance. In here we show that very often, again by virtue of the fact that we are dominated by the spreading current, if we have a defect in one of your buried layer, let's assume you have a tri-layer 2D material and you have a defect in the middle layer, you can get sensitive to that as well. So, it's a little bit like a very special material system where we can detect features that are also buried just by virtue of their uh impact on the spreading resistance. So, I often like this to show this slide to say that, you know, there's clearly value in generating images like these. I mean, you can quite like quite definitely say there is information there and uh quite definitely use that to qualify the growth of your 2D material.

**[29:00] Speaker B:** Indeed, the image on the right side, we um often used that in the past to study what happens after a growth and transfer. So, am I losing any current? Am I introducing any defect just because of the transfer? So, I think there is no debate about the value of this technique, but there is quite some debate about the interpretation of this technique. So, I like to show this slide that says that eventually you can see pretty much everything you want in there, and extracting information in a very systematic way it's basically impossible. And and honestly, every time I see these images, I think about some sort of you know, catering um uh contemporary art. And I'm sure that uh once you see these, probably you just can't unsee these anymore. It will be forever with you on every 2D material image. But, so, we did a lot of work to understand uh okay, let me try to really simplify this problem in term of the tip sample junction.

**[29:51] Speaker B:** And so, what is the transport mechanisms that is dominating the formation of contrast? We we we did this to learn that as a matter of fact, if you are sensing on a 2D material where your back contact is pretty close, most likely you will be dominated by both sheet resistance and back contact resistance. So, basically resulting in a in a level of current which is also showing some kind of gradient. Because as I get closer to my back contact, I will measure more current. The opposite is true is if I'm very far, and therefore I'm effectively limited by my uh back contact resistance. Finally, the third option here it's when your back contact is literally just under your 2D material. And in this case, technically you can think about your transport to be completely dominated at the tip sample junction and effectively you can think about your CFM almost like an STM.

**[30:42] Speaker B:** So, CFM will generate contrast based on density of states. Once you understand that, then you might think, "Okay, let me get the human out of the loop. Let's make this in a very systematic way." That's the work we did basically more than a year ago and the idea was, "Let me set up some rules, treat this data set as statistics ensemble of pixels and apply a bunch of filters, apply a bunch of computer vision algorithm so that we can make a labeling free and and rule-based extraction of information. So, I get a general image that is containing information about how many defects do I have per centimeter square, how many grain boundaries do I have versus the total area, what how many layers are closed are on that image, how many islands do I have and so on." And so, we we came out with this, which is a very nice exercise in term of showing a defect classifier.

**[31:33] Speaker B:** These images are generally acquired with our system that works in high vacuum. That's also why you will probably see that we can hit very high resolution. So, we can measure between 2 to 3 nanometer pretty routinely when measuring in high vacuum. And and so, thanks to that, we basically go under the whole flow of defect classifier defining where is coverage, where are individual defect, where are extended defect. In some cases, you can even generate maps that are transferred to the next guy or somebody you want to make some kind of transport simulation to basically say, "Okay, I have this amount of defect, they are positioned in this place. Now, let me sit the source and drain here and see how my current would flow in that material if I want to use it as a channel." And that's of course something we did in a very much like old-fashioned way, but then we started asking ourselves, "Can we do anything better with machine learning?"

**[32:25] Speaker B:** The very first attempts were pretty disappointing. So, we worked on a very simple way using a pixel level training for our machine learning algorithm and you could see that even though we could hit a pretty good accuracy on the same image that we use for the training, generally we were pretty inaccurate every time we were providing a new image. And so, we decided to merge the two efforts, so use a lot of the rule-based understanding, so that would be all the the beautification that we can do on the image in order to make the image more digestible. And finally, we we throw that into a ML algorithm, which is made in this case custom, just to generate this kind of radar chart that at the end can spit in a few uh seconds the same amount of work that I would do in tens and tens of minutes.

**[33:10] Speaker B:** The work um the results of that is shown over here. So, we can achieve 99% accuracy compared to what is manually labeled. So, therefore, we can trust this very much on our data analysis. Now, the flow is shown over here. It's a little bit more complex than before. I just want to emphasize this block. Of course, there is a flow which goes for the training, and then there is a flow that goes for the execution when you give it an unknown image or when you're treating data. But, this part is what probably matters the most. So, here we replaced our pixel-to-pixel um training type into a component um segmentation type. So, now we have 36 features that are effectively used when it comes to classifying the components in our image. So, in other words, each of the components that could be one of your green things or the red things will undergo a bunch of this classification, most of which are very dumb, have to do with the shape, uh

**[34:07] Speaker B:** eccentricity. Is that thing closer to a sphere or to an elongated object? Some of which are also funnier, like the the skeletonized, which somehow very quickly tells you if you're talking about a single defect or an elongated defect. Or some are also fairly sophisticated, like in this case where we are trying to detect parsing our image north, south, southwest, and and all the direction looking for our back contact, cuz we learned that if there is a back contact, there is also a gradient in our current as I get closer to the back contact. And so, with that in mind, um we we basically created this flow. By the way, the flow leverages also morphology. Let's not forget we have access to morphology. And so, in many case, we can distinguish non-conductive defect, whether they're real non-conductive defect coming from the material or coming from some sort of absorbed junk on top of the material just because of the presence of morphology.

**[35:02] Speaker B:** That works at different scales for different materials, whether you know them before or you train your algorithm on that. Of course, they must belong to the class of 2D materials. Here you see some atmospheric pressure CVD for molybdenum disulfide at different uh landscape, 1 by 1 micrometer, 500 by 500 nanometer, down to 200 to 200 nanometer for a MoS2 CVD case of 2D material, in this case molybdenum disulfide every time. And And you can see for all of them, we can pick up and extract and compare grain boundaries, clustered defect, point defects, and and basically most of these uh type of defects. Finally, this is where you get really your time uh savior. Uh you can quickly throw and fraction of second later, you generate statistical data sets from tens of images. So, again, it doesn't matter how good you are, it doesn't matter how sophisticated is your data analysis with gradient, with MATLAB, with Python.

**[35:55] Speaker B:** So, this process uh of extracting this data uh it's going to take bunch of minutes per image no matter what. In this way now, it takes less than a second per image. You can give a lot of images to our system and get access to this kind of um comparison. And again, there you can extract, as I said, coverage, defect area, the amount of extended defect, amount of individual defect.

**[36:17] Speaker B:** All right. So, I don't think uh I I I spoke enough. I want to just um quickly give two sides of this conclusion. The one on the E-beam. So, the E-beam is a very exciting area in our opinion cuz break free from this concept of having a back contact, it allows a lot of the existing know-how about electrical AFM to be transferred into in-line AFM potentially. Or, as I said, if we can make very inexpensive self-sensing AFM, it can allow to get virtually everywhere this type of sensing. Um a lot of that has to do with 2D 3 5 and of course failure analysis, which is what we do. And of course now a lot of the work must be done in the space of probes, in the space of developing also the non-contact mode, and we're working on that actually. We have at least a couple of PhD student working on this uh subject right now.

**[37:05] Speaker B:** On the ML and AI, uh that's with us and will be forever with us in the space of metrology. We do see a lot of value in term of reducing time to data. We do see a lot of value in extracting um information from our layer of interest. We do have I mean we we we presented here a lot of work on 2D materials, but in the background we're doing quite a lot of work also on magnetic sensing. Uh for example, here using SMBM, which is scanning nitrogen vacancy magnetometry. We do a lot of work on scanning thermal microscopy, so that's for picking up thermal feature. And of course, we love our CMOS. So, we do a lot of work also on SSRM. And of course, everything that give us the structural resolution such as TEM and STEM and the stuff that provides value. I am mostly talking for uh a living, but the doers are in this slide.

**[37:57] Speaker B:** These are the guys to acknowledge and uh those that are making all the work. Um And that's pretty much all I have for you. So, let me thank them and uh thank you guys for your attention. Thank you.
