SemiWiki.com - Podcast EP367: How Keysight is Accelerating RF and Microwave Design with Matthew Ozalas

Episode Date: September 18, 2026

Daniel is joined by Matthew Ozalas from Keysight. With an experience of more than 20 years, Matthew brings a plethora of knowledge in complex RF designs using CAD tools.   Matthew has also authored several application notes on RF and Microwave design and development using Keysight ADS with his recent ones based on “Artificial … Read More

Hello, my name is Daniel Nenny, founder of Semaywiki, the Open Forum for Semiconductor professionals. Welcome to the Semiconductor Insiders podcast series. My guest today is Matthew Ozzalus from Keysight. With an experience of more than 20 years, Matthew brings out a plethora of knowledge in complex RF designs using CAD tools. Matthew has also authored several application notes on RF and microwave design and development using KeySID ADS, with his recent ones based on artificial neural network.

simulation models for RF power amplifiers. And he has produced several tutorials on YouTube regarding complex RF designs. Welcome back to the podcast, Matthew. Thank you, Daniel. It's nice to be back. So Matthew, paint us a picture. What does a typical day look like for a high frequency circuit designer today? How might it look five years from now? Yeah, that's a really good question. And things are changing very quickly. And I think if you look at design today, it's not all that different from the way design was, maybe 20 years ago in the sense that people

have a set of tools, they pull them up. They're these user interfaces. They've got like components, they mix little schematics. They run EM simulations. They have all these different pieces and steps that they do. And so they're sort of highly customized and highly specialized. And so at this point, the only thing that's really changed, I think, in the last 20 years has been the complexity of the products they're working on.

So essentially, it just means they have to simulate more stuff and they've got more complicated models. So, you know, if you look at an RF designer, I mean, for the most part, they're usually a circuit designer. So they have, they're piecing together these circuits with schematics. They're simulation engineer because they've got to run these different simulations and understand these types of technologies, things like, harmonic balance, how does it work? You know, how do you make it converge well? If it doesn't converge, how can you figure the problem out? They're also a data analyst. They're generating lots and lots of data. They've got to make all kinds of plots and draw conclusions and analyses from these plots. They've got to run optimizations. They run electromagnetic simulations. Then they build

layouts. So there's all these kind of different steps in the design process that get pulled together. And these engineers really are able to go in and perform just a plethora of different tasks and make them all work together. So that's kind of the highly specialized nature of RF microab design. Now, the challenge is today, if you really look at what designers are doing, a lot of it is just jumping between different tools, reviewing results. There's a lot of iteration in the workflow, and it's very knowledge intensive and customized. So what do I mean by that? It's repetitive, essentially. We did a study with RIC designers inside of KewSight, and we found that roughly 40% of the work

that they do is what we call rote tasks. It's like clicking around in the tool and doing things that aren't necessarily building circuits. It's like putting a component down and changing the value of something on a component or drawing a, you know, drawing a repetitive structure in the layout. So over the next couple of years, we're really expecting that to transform. And we're already seeing that happening. And the two things that are coming together really are kind of this, this 6G piece,

which is moving engineers and wireless technology up into really high frequencies, like the terrahertz frequency range. So, you know, 100 gigahertz plus frequency ranges. And so up there, a lot of the things that we got away with down at the lower frequency ranges like where we're at even today for our commercial cellular and things like that. That stuff can't be taken for granted any more. Multiphysics is absolutely critical. And then, of course, everyone's talking about AI.

So that converges on the other side. And that brings the potential to change the work. So nobody knows, of course, what's going to happen. But we see a couple of steps along the way. It's kind of like the driverless car. There were five different levels to get to autonomous vehicles. We kind of have a similar view. And I think the industry has taken a similar view to that where start out in the existing state where people are driving everything manually through graphical user interfaces.

And eventually we step in slowly. We add automation. We start to bring in like co-pilot assistance. And I think an engineering role is going to transform more into being like you're going to have a team of maybe you call it assistants or agents who are going to go out and maybe do research for you or maybe perform some of these routine tasks. that aren't so much engineering. And over time, we get to a point where engineers are thinking at a higher level. The optimist in me wants to really think that the engineer gets to be created in the future. They get to just harness their creativity. And a lot of the things that are routine and tedious today and manual go away. And so they're not really being held back from what they want to create.

So that's kind of the way we can imagine it. And from a just like a workflow perspective, If you think about design, there's already automation that comes in. That's kind of like, okay, I still have my tasks, but maybe they're faster. We get to a co-pilot, and then you start to ask, and maybe AI will make suggestions for you. As we get into agents, you might delegate part of your work, or maybe you want an AI agent to go and do a research study for you, like maybe something you might ask an intern or a new engineer to try to do. And after a while, you start to get teams of these that come together. And eventually we see that organizations and groups is almost becoming like this learning system,

this larger, bigger learning system, whereas you're doing design, you have access to all of the information from all of the engineers and everything that's been learned in your organization and beyond. You have that at your fingertips. So that's the potential future we see. That's a great picture. So most of my experience has been in digital design, and today it's highly automated. RF and microwave design seems to have lag behind in automation. And this is a big question.

So why is that? Yeah, that is a good question. RF is fundamentally a little bit different than digital design. So I come from the opposite side, and I don't, I have to admit, I don't know a ton about digital design. But I think there are some things that digital design benefits from. You've got standardization, so you have standard cells. And then you have abstraction. so you can abstract those cells into essentially code, right?

And so you get to a coding problem and then you can repeat it, and then you can reuse it. You can kind of scale that code to massive things. And I think the benefit of that is since you're in a domain where, and this is probably my naive view as an RF designer, but it's that, I think I'm jealous of how the digital pieces kind of work like they, like you expect for the most part. When you design this big architecture and there's all these smaller blocks, these little transistors, and they're switching, you can kind of predict the net effect of that at a high level. And for RF microwave, I mean, that's not really the case. The transistor is not as well tamed as it is for digital. So it's highly physics driven and the physics depends on the device. So if you look at an RF design, maybe there'll be a few transistors as opposed to digital where there's millions and millions.

transistors. So for that, we're trying to hop on a bronco. We have these, we have these high frequency devices that just have so much gain and so much power that they're, they're hard to tame. And so we have to, we have to consider things like electromagnetic effects. Everything is a little antenna on the circuit. And then when we go to high frequencies, I mean, one fundamental piece is the antenna size goes down at frequency, when you go up in frequency. So like if you, you know, I go out to the coast, I'm here in California. So, and they have these giant antennas there for communicating across the ocean. Those are relatively quite low frequency. And so they need big, huge antennas to produce that signal. But when you get up to even the millimeter wave frequencies like

where we're at already for 5G, the antenna shrink to something that just fits into, you know, almost the chip size. And then when we start going up towards terrahertz, they become like the structures on the chip, like the metal and the transistor structures, start to actually become antennas. And so you have all these transmitted receive paths happening all over the chips, which is very, very difficult to manage. Add into that packaging, we have more complex systems. Everything gets packaged together. And we have antennas everywhere, and we're cramming more stuff together, and they're all interacting. And then to make matters worse, these simulations aren't perfect. The simulations are quite difficult. So when you, modeling these things is challenging, and you go into the lab

and you measure it, and it doesn't agree with what you predicted, because you have essentially a flawed or a simplified model of reality. And when you come into the lab, something slightly different. It's like if you have a phone and you put it close to your head, the antenna pattern changes on the phone because it's close to your head. So anything around you now in the physical world can impact the performance of the part. And you can't predict all of that in simulation. So it's very challenging.

And as a result of that, the workflows that people build are like so customized. You know, you have this view of the RF and micro of engineer as almost an artist. And so they're custom and they're ad hoc. I was designing an amplifier for an example this morning. And you just kind of go back and forth. There's not there's not like this predicted repeatable pattern. You go, okay, that over here that I changed this component and it changed.

And then I'm going to go over to this other part of the circuit and change this other thing and see what happens on the other side. Okay, now I got that tune. Now I'll go back. And so it's iterative and it's also organization specific. Every company that you work for has a different set of tools and a different way to do design. And so that's kind of a tricky part. The way I see it, I mean, I'm a, I work at KSight, so I'm sure I'm biased because we're an instrument company, but

I think when we look at instrumentation, I think it gives me hope going forward. Like instrumentation is something that's been around for a long time. It's been around longer than simulation. Like we measured things before we simulated them. And I think that if we look at instrumentation, we can we can start to maybe get at parts of this, right? Like measurement has sequences. Measurement is the sequence that we repeat. Measurement is something where we can take different devices and we can put it onto a measurement bench and we can measure a part, move it around, take another part, measure that

with using the same set of hardware. So we can do it on multiple devices. And we've also automated the process. If you look at these test systems that KSight, our customers are using, they're highly automated. There's a lot of different instruments working together. all of those instruments are working together and they all got this common way they can speak to a computer. Interesting. So going back to the day in the life of a designer that you described,

can you give us an example of how part of that workflow can be captured and reused because reusability is a big deal here. And where does AI first start to enter the picture? Is this the entry point? Yeah, sure. Let's stick a little bit on the instrument side and think about one of the things that on instruments we do is there's programs and we have this of graphical programming idea where we set up like sequential. So HPV was the one I first used. I started my career as a test engineer and I used HPV and it's like the set of blocks and you connect the blocks together. It's kind of like this visual flowchart. And then a lot of people know LabView, another program that's kind of like that. You have these blocks and this is how we orchestrate and automate test sequences, right? So we've been looking at doing that for design,

well. And so the example, let's go back to the power amplifier example. I'm pretty familiar with that and I was working on one this morning. So it's fresh in my mind. But let's say that an engineer wants to select the best transistor for a power amplifier, right? Today, the process, you know, assuming you have a transistor that you can get into the intrinsic parts of it, you might like run a DCIV sweep. Maybe you perform a load pole and you you look at the waveforms on it. You follow this process, like you draw a load line on the DC curves, and then you try to present that load line to the intrinsic node of the device outside, you know, through the parasitics, which warps the effect. And then you tune up a matching network to try to give you that result, and you hook them together, and you kind of compare the results manually. So this is a multi-step process. We're running DC simulations here because we have to get the IV curves. We're running harmonic balance simulation which is a large signal like nonlinear simulation where we put a tone through that's so that we can get the compression and then the characteristic and the nonlinear effects we're running

small signal simulations so we can do stability so if you look at what if you're thinking about this in a ui or like like a program you know we have like 10 different you know test benches and then we're going to run we're going to run like a load pull we're going to change like each harmonic we're we're going to tune the load of that over the entire range in the Smith chart. So the point I have is that this is a very tedious and repetitive process, and it's like a fine-tuning thing. You know, people are going in and literally we have knobs on the simulation tools, and we, you know, like you fine-tune this knob, you dial it in,

then you move to the next one, you dial that one in. But this starts to look like a test sequence a little bit when you think about it. And so what we've been able to do is take that process with all of those different simulations and put them together in blocks. So like one block might be the DC simulation. So the first thing you do is a sweep on the DCIV and you generate the curves. And then you take those curves and you draw a load line using like a function, you know, Python function or something. And then you take the results of that and you move on to your load characterization process. So you take that as an input to the next step and you operate.

on the next step and your optimization goal is the load that you derived earlier, both for the fundamental and harmonics. And then you sort of iterate that through the harmonics. My point is that you could take this multi-step multi-simulation process and you can turn it into a sequence, like a block sequence, where each block is a simulation and the results of the simulation change the way you do the next simulation until eventually you can engineer something. You can do like a literally engineer the waveforms inside of the intrinsic node of these transistors. And so that's very interesting because once you capture it, this, you kind of have a little piece of knowledge. And I think RF engineering, it's all about these little pieces of knowledge. Once you have these little pieces of knowledge, you can start to reuse them and repeat the process and make it more straightforward. And so just to draw AI into this a little bit, this is something called a skill.

in AI. So if you have your engineering process, the sequence of steps that you take to do a design, you can start to teach AI about those steps. And you can teach it things called skills. And the skills are nothing more than the things in the box. Oh, this is how to run a DC simulation. Okay, remember that. You know, this is how to do a load pull simulate. Remember that. Now let's connect the skills together. Now you can start to have agents use these skills to walk through the process. So that's maybe the first, or I guess the transition that you can make between getting your engineering process and starting to bring AI in slowly, right? So once you've captured these individual engineering workflows, what's the next step? Do they simply connect together or is something more required?

Yeah, so each little bit is maybe a small piece. So even with the waveform engineering example I gave you, right, that's like a small little piece of the overall process. That's not the whole engineering process by any means, but it's a small step that you've captured. So you might say, well, big deal. You know, what does it matter? I still have a hundred other things to do. I have to go design a matching network. I have to make a layout, et cetera, et cetera.

But the point really is that once you have that little piece, you can start to iterate through that piece. So before we captured that set of the workflow, and by the way, if you read Microwave Journal, I had an article with one of my colleagues back in the July issue. where we talked about this exact problem. But once you have that, he's captured. Now you can start to iterate. So the manual process, right, it takes, let's say, 10 simulations and one designer walking through all those simulations, it takes like half a day.

So you get a transistor, if you're really good, and if you know what you're doing, it takes half a day, right? So you get a transistor, you go through all these steps. You finally get it right. Now you have the specs for your matching network, and then you'll go design a matching network. Well, if we've captured that process, now you can start to think about scale. And you can do things that once you get to scale, you can do things that one designer could never do in a day anymore, right? Because you could never, for example, I might get a library.

I work with modolithics parts. They have Corvo gallium nitride transistors in them. One library might have like 30 or 40 different transistors in it. And I try to choose the best one. I just like look at the data sheets and I go, okay, I use judgment to try to pick the best one. But I really have no idea. I can't test, I can't try waveform engineering on all 40 devices in my library. Well, now that we've captured this into a step, it's just a for loop.

If you think about a sequenced diagram like in a test and measurement, well, now we just connect the output back to the input and iterate, right? That's a four loop. And so you imagine you start to put four loops in there. Now let's do a four loop and every time we run this, we're going to swap the transistor. So we're going to run this exact same process, but on a different transistor, every single. single time. Now you can go through your entire library, and this is like scaled automation, AI turns out to be really good at writing code and scaling automation. You can go through all your

library and you can say all the devices in my library, go do a waveform engineered design on these and tell me which one is the best. And you know, why stop there? Let's, companies have multiple technologies. They have access to it. You say, we have our own fab or when we can use parts from other fabs or transistors from other fabs. So maybe you start to say, you know what, I have four different technologies. Each of those technologies comes in a different PDK process design kit. It has a library with 40 parts. Now we're talking hundreds and hundreds of parts, right? Go through all my technologies, all my parts, and give me five really good candidates to start my design. So that's the type of scale that you can achieve. And it's just not something you can do with a single experience,

even experienced engineer. It just takes too long. So that's where you can start to coordinate those workflows and get them pulled together. And that involves the step that we call it is orchestration. So now we're controlling different technologies or orchestrating technology. We also have to think about orchestrating simulations because that's just one piece, but we have to orchestrate electromagnetic simulation too. So we have different tools and we have to get them to work together. I kind of think about it like a conductor, right? A conductor has a lot of different pieces. Maybe there's different tools, different technologies, and they have to get them all to work together in sequence and harmony to build something useful.

Right. So let's say we have a perfectly captured and orchestrated workflow. Does that automatically make engineering faster? Not necessarily. It's the prerequisite, I think, because we're talking about how to apply AI in practical ways here. So capturing and orchestrating are the necessary steps. You have have to be able to capture your engineering process. Otherwise, you're just throwing data at AI, which RF engineers are the most skeptical people. Because nothing ever works, right? Like when you go into the lab, your part,

everyone who's designed hardware at high frequencies, your part never works. And so that leads to skepticism. And so you can't just take all of your data from the lab and throw it at some large language model and expect to get some kind of reasonable design, or you can't throw all your models at a large language model.

And it's going to come up with perfect designed for you. That's like a thousand monkeys or a million monkeys on typewriters. Uh, analogy, right? So you have to get that engineering process first. And then once you grab hold of the engineering process, that you're like, you know, jumping on the Bronco a little bit, you have to be able to control and then steer it. And that's the orchestration piece. And then from there, you can start to make it go fast again, right? Now that you have control, make it go fast. And that's really the key. So we have the consistency and scalability, we still have speed challenges. So if your simulation, like think about

an electromagnetic simulation, that may take a couple of hours to overnight, that can slow you down. Even if you have these really powerful tools at your disposal, if every time your automated flow has to go off and run a simulation, and that takes, I don't know, a day to run, well, you're back to square one. You can't really move that fast, even though you have, you know, you have all of this power. So we have to be able to make that accessible. And that's where things like, surrogate models come in to play, right? So surrogate models are abstractions of electromagnetic models. Essentially, we run a bunch of electromagnetic simulations up front. We kind of, I call it paying the toll. And then from there, we can fit an artificial neural network, that data. Now we can get a really

fast result. We can fit an artificial neural network to a physical structure and then predict the results of that structure. Now we can move, again, right back at the speed of that AI wants to move. So one of the things probably the most famous example of surrogates in our industry is I don't know if you've seen these weird QR code like structures or there's these little like blobby pieces of metal and with RF engineering right it's just about metal coupling so people have invented these QR code looking things and they're just metal on or off it's like a it's like a binary and we can engineer those then if you look at enough of these you'll eventually get all kinds of different responses and then you could fit that to a neural network and predict whatever a response

you have in the you know the frequency domain for example you can backfit that then using a neural network and say well that one came from this sequence of metal and no metal that gives me that result so we can use things like that to create structures that would have never been possible to actually design using data you can also by the way use that on just i think a lot of engineers are skeptical about that it's a little controversial i make the point that you can do this with inductors you know and then and structures that we already know how how the physics physics work, right, and how they work. So you can take all of your physical pieces in your library and convert these to surrogate

models and now you can optimize very, very quickly and take advantage of that AI piece. There's these different approaches to accelerate reinforcement learning as another way to accelerate. We can have optimizers that learn from the last step and we don't have to relearn everything over again. Like when you start an optimizer, you start an optimization sequence. doesn't ever remember what the last optimization you ran was. So it always starts from scratch. Well, reinforcement learning lets you take the knowledge that came from the last optimization run or any optimization runs and apply it forward.

So there's these ways we can make the actual analysis piece faster. So that's kind of the piece. So the three steps that we think about, right, we think about capture, that's get the engineering process, orchestrate, get everything working together in sequence. That's, by the way, that's like instruments too. You have to orchestrate lots of instruments. And then once you have those two, you can accelerate process forward. Matthew, last question and probably the most important one. Is anyone actually building and using these kind of AI-driven engineering workflows today?

Absolutely. Yeah, there's a lot of teams that are implementing pieces of this journey. So the point I want to make is when you talk about five years from now, everything's going to be different or we're going to have all these agents running around, it implies like maybe there's this step function that somehow we're going to, you know, one day we're going to come into work and everything is going to be different than the day before. And I don't think that's really actually going to be how it happens. It's just one of these slow changes where over time you get a little bit more automation into your flow. You start to capture a little bit of expertise, add some more orchestration in there. You add the more

introduce AI assistance and they start small and they get more and more sophisticated. And then you adopt some of these accelerated technologies like surrogate models or like reinforcement learning. And so we're absolutely seeing that stuff happening. But the first step, I mean, what do we see? Well, Python-driven automation, that's something that we've been working on at KSight for a while. A lot of our customers are embracing that piece of it. So with all of our tools, we're building in Python APIs, and those are just like command layers. It's actually very similar to Skippy commands for instrumentation, where you can control an instrument with it, kind of taking them. approach as well with our tools. And then AI is really good at writing scripts with this Python

automation. And that leads us naturally to AI assisted engineering and this design exploration. So the people that win in this space, it won't necessarily be, I don't think anyway, the companies that have these like huge AI models and anything. I think it's actually going to be the companies that can do these steps, capture, orchestrate, accelerate the most effectively. And so one example, it's kind of funny. Normally I work with pretty big companies and the hardware space has always been a place where there was these big incumbents because of the capital costs, right? So there's not a lot of startups in the hardware industry. Recently we're seeing in RF microwave a lot of different startups and one really interesting one that I met with recently is a company

called Spear Semi. And they are doing switchable filters and they actually have taped out design and gotten them back and had them working. And so these things are generated by AI. Essentially, what they do is they've automated their entire engineering process. So they can generate these layouts and designs just using code. So once they have that, then they can start plugging in like optimization tools to be able to physically optimize the structures. And essentially, they can generate thousands of designs because they're running at the scale

I was talking about. In the time, it takes a human to do one design. And so they'll generate several thousands. A lot of those designs that are generated are, yes, they meet all the rules and the requirements, but they don't perform so well as you'd expect. But what's really interesting is a couple of the designs, when you get to that kind of number, actually do perform well because when you look at them, they take advantage of coupling. Humans, we want to avoid coupling, right?

That's like if you're an RF engineer, that's your worst nightmare. It's two things. So you make everything spaced apart to avoid and minimize coupling. what we're seeing with these designs is they actually take some advantage of a coupling. So they'll place things close together and the coupling just works. And so they've done tapeouts of switchable filters. They have measured results. And they've gotten, I think they say they have something like 30% reduction in tunable filter area and 6db improvement in isolation on their switches.

So it's quite impressive. Hey, thank you, Matthew. It's always a pleasure. I look forward to your next visit and have a great day. Thanks so much, Daniel. That concludes our podcast. Thank you all for listening and have a great day.