Chipstrat Chat with Jared O’Leary, CEO of SirenOpt
PhD to Entrepreneur, Non-Destructive Metrology, Entering Established Industries, Berkeley Skydeck, Advice for PhD Students, Physics-Informed Machine Learning
I like to pay attention to startups because they highlight customer problems incumbents don’t (or can’t) solve. Today, we’re speaking with Jared O’Leary, a Berkeley PhD and co-founder of SirenOpt, about a novel metrology platform that impacts semiconductor and battery manufacturing.
We cover
The transition from academic research to a venture-backed startup
Technical details of cold atmospheric plasma for non-destructive metrology
Challenges of selling hardware to high-volume manufacturers
SirenOpt’s strategy to enter established industries
Advice for PhD students considering entrepreneurship
I also threw a curveball to Jared at the end with a question about the usefulness of LLMs for his business, and he has an interesting take that aligns with Nvidia’s vision for physics-informed (Modulus) and physics-aware (Cosmos) neural networks.
If you’re new to the topic of metrology, here’s a brief introduction:
Also, SirenOpt is hiring EEs right now.
I’ve included the video, audio, and transcript below. Enjoy!
Introduction to Jared O’Leary and SirenOpt
Austin Lyons: Hello, Chipstrat listeners. Today we’re talking with Jared O’Leary, co-founder and CEO of SirenOpt, a company transforming manufacturing intelligence with a novel form of metrology that uses cold atmospheric plasma, sensors, and machine learning. SirenOpt is helping manufacturers measure the unmeasurable, reducing waste, improving product quality, and redefining what’s possible in advanced materials production.
So hello, Jared. Welcome. First question for you, will you tell our listeners a little bit about you?
Jared O’Leary: Yeah, sure. I co-founded SirenOpt in August of 2022 with my former PhD advisor based on my PhD research at UC Berkeley in the Chemical Engineering Department. Before that, I did my undergrad at Stanford, also in chemical engineering. In between Stanford and Berkeley, I worked in industry for a few years.
PhD Research and Development of Plasma Sensing Technology
AL: Nice, nice. Okay, cool. So, how did you land on plasma sensing during PhD work?
JO: My PhD broadly focused on characterizing, modeling, and controlling what can be labeled as intrinsically stochastic systems with nonlinear dynamics. So basically systems that evolve probabilistically and nonlinearly.
My PhD was almost entirely computational. It was really about developing computational tools to characterize, model, and control these types of systems. We looked at a lot of different systems during my PhD, including systems for semiconductor manufacturing, including systems for colloidal self-assembly–if anyone’s ever heard of that.
Towards the end of my PhD, I started briefly looking into these systems of cold atmospheric plasmas. We were interested in characterizing plasma material interactions. As a second step, trying to develop advanced feedback control algorithms to control those plasma material interactions.
We were looking at cold atmospheric pressure plasmas in the context of treatment–treating skin tumors and things of that sort.
Understanding Cold Atmospheric Plasma and Its Applications
JO: I think at this point it’s probably reasonable for me to define what a cold atmospheric plasma is. Cold atmospheric plasma is a weakly ionized gas that sits at about 40 degrees Celsius at atmospheric pressure.
Plasmas in general are considered the fourth state of matter. So, lightning is plasma, and stars are made of plasmas, etc.
We were looking at a cold atmospheric pressure version of that.
So anyways, when we were looking at these cold atmospheric plasmas in these treatment applications, we realized two things. One, the feedback control algorithms that we developed could create a very, very consistent plasma. And second, we realized we could create and maintain a very consistent plasma at a very, very low energy, which allowed these plasmas to interact with materials in a non-destructive manner.
Once we made that realization, we knew that there was a lot of information in the resulting plasma material interactions. We knew we had the machine learning expertise to extract that useful information from those interactions. So we thought, hey, maybe this could be a sensor. This could be a non-destructive metrology instrument.
Because of our familiarity with controlling advanced processes and advanced manufacturing processes, we intimately understood how often a lack of information limits these advanced processes. We thought, hey, this could be a sensor or metrology tool.
We knew, broadly, this might be important. So, let’s just run a few experiments and see what we see. And then we ran a few just quick little dumb proof of concept experiments that showed, OK, I think this actually could be a sensor.
After doing that, we then had to ask the question, would anyone use this? How valuable is this? What types of industries would maybe want to use the future version of this type of tool?
Luckily, we’re in the Bay Area and there are lots of battery startup companies and all sorts of interesting research around. We talked to a few companies who had a lot of manufacturing challenges that we felt we could solve.
That was enough for my former PhD advisor and me to say, hey, we like working with each other. Let’s try and start this company.
AL: I love it. You’re in the right area [Bay Area] for that.
So it sounds like you started on the computational side and then eventually started experimenting with it. Could we actually physically make a system to do this? Then you started looking at what the applications are.
You said you landed on cold atmospheric plasma. Not super high temp, not lightning. And atmospheric—not super low or high pressure, but just normal pressures. So I assume that that means it’s amenable to manufacturing because you don’t have to have a high power or expensive system.
JO: Yeah. Let’s be precise here. The primary benefits of the plasma being cold–which we defined as below 40 degrees Celsius— and atmospheric pressure is that this type of system can easily be placed in a lot of different types of manufacturing systems.
So when you say this thing is not necessarily expensive–well, let’s talk about what expensive really means: this does not require some extremely fancy setup. We believe can be pretty easily integrated into high-volume manufacturing workflows.
JO: Another benefit of the low energy or lack of temperature, coldness for lack of a better term, is that the plasmas can be very gentle. They can interact with materials non-destructively.
AL: That’s right–non-destructive. Presumably, that means you could do inline sensing.
JO: Exactly, exactly. And that’s the goal of where we’re going here.
The two key aspects non-destructive: Can you measure fast enough? And can you actually be integrated into a high-volume manufacturing workflow in a reasonably easy way? We believe our system meets those key requirements.
Technical Details: Sensors and Data Processing
AL: Nice. For a given industry, if your system is inline–you’ve got the plasma and you’re gently interacting with the material and measuring some interaction–what measurements are you making?
JO: Great question. So let’s let’s answer this in two stages.
So the first stage – there’s these plasma material interactions and and they’re informative.
Well, okay, how do we record them? What is that? What does that mean?
So, the types of interactions that we have or induce are synergistic thermal, electrical, and chemical plasma material interactions. You actually have something very interesting in our system: as the plasma interacts with the material, the material responds to the plasma, and the plasma responds to the material. So you end up with a lot going on.
Our system has detectors that are directly meant to record these synergistic thermal, electrical, and chemical interactions. So, our system contains an optical emission spectrometer. It includes a high-resolution thermal camera. And it contains a variety of electrical probes.
The detectors it contains right now are by no means the final set of detectors that this platform will ever have. An active area of research for us is seeing which new or additional detectors we can add. But at least for right now, we have these three classes of main detectors, and the platform records all this raw data; right now it’s 213,000 data points per measurement.
And that raw data itself can be physically meaningful and can be informative. Certain peaks in the optical emission spectrum mean certain things physically. Certain electrical signals mean certain things physically. If you look at the thermal image and how heat diffuses over the thermal image, that means something physically.
For advanced manufacturing processes, a lot of people care about specific properties.
AL: So, let me reflect back to you and see if I got it right.
So you’ve got the plasma, and it interacts with the material. You’ve got sensors that measure the optical emission, so the light, the spectrum that’s given off as you interact with that material. And the thermal characteristics—how hot is it or how does heat dissipate? Then there was electrical, so is there an induced change in the current or voltage.
And then you take all of that data and you feed it into a machine learning model, which has been trained on specific properties.
So it probably can say: yes, this looks like silicon dioxide with some dopant in it or something. You can infer things from it.
JO: Exactly. Yeah, and it’s not to say we don’t use the raw data. We also use the raw data that’s helpful. Still, it’s simultaneously providing all sorts of raw data that people can use and then also providing these critical material property measurements.
Advantages of Plasma-Based Metrology
AL: So you can get information about the material's properties and then the benefit is multiple characteristics at once. So you can measure everything at once.
Ultimately then, the benefit for the customer, is it a cost optimization thing? Instead of needing five different tools, you can have one tool?
JO: It’s not really that. Our core value proposition is measuring certain types of properties inline or non-destructively that otherwise cannot be measured either non-destructively or in line.
All these properties that I talk about are being measured in manufacturing workflows already. But they’re being measured by taking a few samples off the line occasionally and then getting those property measurements.
AL: So it’s every five minutes we pull something off and we destructively measure it. And okay, by the way, this one was bad, but the last one wasn’t. So, for the last five minutes, was everything bad? You don’t know.
JO: Yeah. Exactly. Then basically this better real-time information allows manufacturers to have more sophisticated and tighter quality and process control.
I think it’s very easy to think of that in terms of a yield improvement. But when you can have tighter quality and process control, you can also make a higher-performing product. So it’s not necessarily about reducing scrap rate, which I think a lot of people when they first hear about this, they think that’s our core value prop, and I mean that’s part of it. But the type of value that I’m most excited about providing is allowing people to really make higher-performing products.
Early Customer Engagement and Market Positioning
AL: So tell me, as you came out of your PhD, you started this company with your professor and you guys raised some money, I’m assuming you built some prototypes. How are the conversations with early customers?
JO: I would describe the conversations in the following way: We’re never trying to compete with any existing metrology instrument. We’re always trying to provide measurements in certain spots of manufacturing workflows where there basically are zero measurements.
So we would always open our conversations up with introducing the technology and then say, hey, what do you wish you could measure that you currently cannot? And that would get to a decent brainstorming conversation.
I think we had a reasonable amount of technical credibility, but also not all the technical credibility in the world—lots of PhD students work on weird stuff that can never work in industry.
But the core benefit of making a metrology tool was that it’s very easy to try it out and test it.
So I think very early on, I was always extremely careful to say: the platform can do what we have shown it exactly can do. And to always lead those early customer conversations to some demo where we say, “Ship us some samples. We’ll run them under our metrology tool. We’ll collect the raw data. We will make a detailed report about the performance. And then we can take it from there.”
It’s not like testing the viability or proving the concept for this is particularly sophisticated. In a lot of cases, it’s really just sending us some samples in an envelope – seriously. We’ll see the results. For certain customers with very specific samples, very sensitive sample — we say, fine, just come and run it on our campus. That’s also fine with us.
Despite maybe early and fair skepticism, it was always very easy to prove the concept of the system.
AL: That makes sense. I like how you position what you guys do as “We’re gonna unlock something new for you. We’re not trying to come in and take away business or anything. So, just give us a chance, let us measure the thing and let’s go from there.”
JO: Thanks. Exactly. Because I think it’s very important to position ourselves this way.
There’s no benefit for us trying to be a slightly better version of any existing sensor. And, even if we are a slightly better version of an existing sensor, well, that existing company already has a relationship with the other company—there’s just so much risk involved in switching.
So we’re always just trying to provide new types of measurements.
From Alpha to Beta: Scaling the Product
AL: Give us an update on just where you guys are at as a company. You’ve raised some money, hired some folks.
JO: We raised some money, we hired some folks. Yeah, I mean, I think that’s a fair summary. I think at this point, we’ve now raised about $9.7 million from venture. We most recently closed a $6.6 million seed round in July. We now have, I think, a whopping 19 full-time employees.
We are basically creating iteratively better versions of our product and releasing those to paying customers. So at the beginning of 2024, we released our alpha units, which were offline units, which are beneficial because even if something sits outside of the manufacturing line, just being able to measure multiple property measurements at once with full uniformity maps, non-destructively, still provides value. And it’s a stepping stone to our inline tool.
AL: Builds trust with your company too.
JO: No one will put a tool inside their manufacturing line unless they’ve used the offline tool for six months to a year.
Yeah, so we released these alpha offline units early last year. And now we are on the precipice–we think as early as later this month, early next month–to release beta versions of those units which are significantly, significantly improved and have a lot more functionality. And we plan to ship these alpha units—we basically made five of them. We put one in the academic lab. We had one do demos in-house. We put three at paying client sites.
Now, with these beta units, we’re building about 25 of them, and we expect to deploy around 20 to different places in the US, Canada, Europe, and Asia. So we’re really excited about that.
A lot of our early traction was primarily focused in the battery space. And that’s because battery manufacturing is really hard. And it’s new. So the argument of, hey, do you want more information resonates.
But now as we’ve gained more credibility, we’ve been able to enter other spaces as well, including aerospace, energy generation, semiconductor, and various types of electronics applications.
So mostly just excited to see how this platform can work for these various other types of industries.
AL: Nice. Totally. Yeah. Congrats on the beta. That’s exciting.
JO: Yeah, it’s been taking a while. I thought it was gonna be done much sooner, gonna be honest.
AL: Especially if you’re building physical systems, I’m sure that a lot can go wrong.
Entering Mature Industries
AL: You talked about the battery industry, and that makes a lot of sense. It’s an easy place to start because it’s new, and people are hungry for more information.
With some of those more mature domains like semiconductors or aerospace, how do you get your foot in the door there?
JO: That’s a great question. I would say our original foot in the door in these more established industries has always come from some warm introduction from someone that we know, either that’s an advisor that we have or a friend of an advisor, someone that we’ve met.
But even in those industries, particularly in semiconductors, we’re still starting with the newer versions of those technologies. For example, in semiconductors, the vast majority of work has been done with gallium nitride and silicon carbide wafers. No one has let us touch patterned silicon.
That’s not to say we won’t get there, but in all of these things, you see this trend where you have to start with these next generation newer types of processes where there isn’t a lot of information. Then you gain credibility and work your way up.
Even in batteries we started with next-generation batteries. And primarily, next-generation batteries being made by startup companies who were very new.
I think this general trend I’m seeing in these industries: you start with something new where people don’t have a lot of information, then you gain credibility and then you work your way in.
AL: That makes a ton of sense. And I’ve seen that in my day-to-day life where companies come and they want to sell you on something, but what you already have in the process and the product you’ve built, it’s already working, it’s already out in the field, you’re already maintaining it, you don’t want to touch it.
So it makes a ton of sense to say, well, for that next-generation thing you’re building, what’s something awesome that you would want to do but can’t? For example, from a different lens, if an AI ASIC chip is trying to come and say, buy our chip instead of using Nvidia, we’d be like “Dude, we’re using Nvidia.”
But if they’re like, “What if we could give you 100x the frame rate or 100x the inference or way lower latency on the next project?” That resonates with me. Then we might try it, because we haven’t built the next thing yet. So why not experiment?
JO: Exactly. And if we prove that, then we’re credible there, and then there’s ideally opportunities for, the more traditional processes.
AL: That makes sense, that’s good.
Berkeley Skydeck and Fundraising Lessons
AL: I know that the Berkeley Skydeck accelerator is a part of your story. Tell our listeners, what is that? How did you get hooked up there? What’s the benefit?
JO: I’m a big advocate of Berkeley Skydeck. It was really, really, helpful to us.
So, my PhD advisor, when we decided to start our company, we then asked the question everyone asks which is “Well, how do you fund this?” And broadly, how do we even run the company, etc?
And so we applied to Berkeley Skydeck. I think we just heard about it on a mailing list or something. We weren’t very familiar with it. We weren’t very familiar with accelerators at all.
And we luckily got in shortly after I finished my PhD, and they were critical for us in so many ways. Not that I have a great idea how to run a company now, but I especially didn’t back then. And so there was a lot of stuff they showed us, just from very basic things about logistically and operationally what you should do.
But also, they set us up with great advisors who were extremely helpful.
And then third, an important benefit of Berkeley Skydeck, a unique benefit is they really do just connect you to a lot of people. All of our initial clients really were just directly from Berkeley Skydeck warm intros–every single one. In addition to all that, it’s an accelerator that wants to have its companies raise a seed round and raise money after, and they’re phenomenally helpful with raising money.
Again, it’s not that we’re the most successful company at raising money. But it’s a complicated and sophisticated landscape, and they really helped us at least get our feet wet with that and provide us a lot of great advice through that as well.
Had we not gotten into Berkeley Skydeck, we would frankly be on a very different trajectory. Probably much slower trajectory.
AL: It sounds like they helped give you funding, teach you how to run a company, connect you with customers, connect you with your next round of funding, and help navigate that.
JO: Yeah, it’s connecting with the next round of funding, but it’s also really helping us navigate that. I think there’s a lot of nuance in understanding what to ask for. That’s not a trivial problem. I think a lot of people, or at least the perception I hear from maybe people just starting companies, they think, “Okay, I have an idea and now I can go raise a seed round.” I mean, maybe if you have the best idea ever. But we had to raise a series of SAFEs over an eight-month period before we got enough traction to justify a seed round then,
And there’s a lot of strategy in that. Who do you bring to the table? How much do you ask for? How do you best take advantage of the relationships with your investors and use their strengths to help you?
Advice for PhD Students Considering Building a Startup
AL: So now you are an example of someone who’s transitioned from academia to industry, starting your own company.
If anyone’s listening to this who is currently a PhD student or thinking about that path, do you have any advice on transitioning?
JO: Let’s just talk about it from a PhD perspective. I think the whole point of a PhD is to provide you the tools to solve very open-ended problems. So, I actually think that getting a PhD is very helpful for starting a company in the sense that you really are thrown into something that you really don’t know anything about. And you at least have some tools to attack or at least not be afraid of these very, very open-ended problems.
The big differences to prepare for, and I guess the advice that I would give: think about whether this lifestyle, for lack of a better term, is for you.
In a PhD, you get to focus on your research and you get to look at problems very in-depth and rigorously, but you don’t have to move that quickly. In industry, you have to move much faster.
Another big one for me: I was totally unprepared for the amount of external communication that starting our company involved. In my PhD, I basically mostly only talked to my PhD professor. And now I’m talking to God knows how many people a day. Building up the stamina for those types of conversations or understanding how to use those conversations or how to get the most out of those conversations is complicated.
So the first thing I would say is prepare yourself and think about what your life is like if you’re talking to tens of people a day. And all those conversations can be important.
I would say, and I think actually PhD students are pretty well prepared for startup life. The benefit of getting a PhD is you live off a PhD stipend for five or six years. So you’re already trained to not think, “I’m gonna make it big any time soon”.
I would also give the advice as well that the patience and persistence required for a PhD also seems very well suited for this type of lifestyle to me. So I would not start this and think, okay, well now I don’t have to be patient. No. Those skills are important.
AL: I can resonate with the open-ended comment you made: it’s a startup, everything is open-ended, the world is your oyster, but you gotta go figure it out and do it. That’s the PhD in a sense too. It’s solving open-ended things and going and building whatever you need to build, writing papers, discovering what you need to discover to someday graduate.
JO: Yes. I think people talk about in startups all the time how certain successful companies have pivoted and they’ve worked on something else. Well, that pivoting action happens in research constantly. I mean, you’re constantly looking at one thing, and you think, wow, this is interesting, let me go down this route. So this is an example of something I think PhD students are well prepared for.
AL: Sure, totally, yes. I also appreciated when you said they have patience and persistence, grit, and resilience. As a PhD, you’re doing that for five years – figuring things out, trying to make it happen, living off of your stipend. And then startup life, it’s that same thing. You got some good tools in your toolbox.
The communication piece is interesting too. I think you’re hitting on–in your PhD, you’re probably mostly scientist, academic researcher. And now all of a sudden there’s this other vector that you have to become proficient on, which is dealing with humans and communicating.
JO: Yeah, and communicating concisely and clearly to all sorts of different audiences.
And each of these audiences requires a different type of communication. There’s also interesting aspects too, where these individual audiences broadly require different types of communications, but then there’s also the best way you can communicate within those broad categories. The way I communicate with a customer, I’m not going to use the same words as someone who’s been in that industry for 30 years and is trying to sell them something, because we’re looked at differently.
So there’s a lot of nuance, I think, in making yourself understandable, for lack of a better term, to a lot of different audiences. And that’s hard. I’m still not particularly good at it, but it’s a constant challenge and something I hope I can improve at over time.
AL: Sure. Well, hey, you’re self-aware. It sounds like you’re working on it. That’s great. You’re in a good place.
Physics-Informed Machine Learning and LLMs
AL: Okay, so a non-SirenOpt, just random question for you. I have debates with some of my friends about how much people are actually using ChatGPT or not.
In your day-to-day life building a company, where do LLMs and AI come in for you?
JO: I hate LLMs.
AL: Really? Interesting. Tell me more.
JO: I have to be honest, I’ll be totally honest—maybe that was too strong. But here’s what I’ll say.
For SirenOpt, do we use LLMs? No.
Are there certain types of neural network-based architectures that we have looked at and will look at? Yes.
But in general, where I stand philosophically is I really dislike the trend of just building larger and larger black-box models that we throw a lot of data at and don’t really understand the output. We should not be using machine learning to replace all the physics and domain knowledge that we have developed over time. We should use it to augment the physics and domain knowledge we’ve built over time.
I think there’s a reason why LLMs and ChatGPT are constantly being applied to extremely low-stakes problems like writing emails or making a new image or answering a prompt correctly 25% of the time. There’s a reason why we don’t have some big large language model running a factory. And the fact is, when the stakes are high, you need to be able to guarantee certain types of performance and understand certain types of generalizability.
To take this further, a lot of my PhD research actually was broadly in what can be considered physics-informed machine learning, where I was interested in taking maybe some limited physics knowledge that we had and augmenting that knowledge with machine learning tools.
AL: So in your products, you will always have a physics-informed model that starts with the actual physics and is trained on data and then can make predictions, but it’s informed by reality, not just what it inferred from all the YouTube videos out there or something.
JO: Yeah, exactly.
I mean, even if we define “what is machine learning?” Well, machine learning is any data-driven model. A line of best fit is a machine learning model. Although our goal is to put progressively more physics in our platform over time, we’re very likely to always have some data-driven element.
The main point here about how I feel about ChatGPT or how much do we use it: These LLMs seem to be good to just provide a lot of information quickly that is broadly similar to what you might want to look into. But I don’t see large language models ever solving extremely complex problems that are important.
One last example I’ll give is self-driving cars. Self-driving cars have all the training data in the world. I mean, there’s no more training data you could give any type of system. And Waymo still blocks traffic three times a week in San Francisco.
So there’s just—I think we’re going to—the way forward is not to build progressively larger neural network models. It is to see how we can augment what we already know about the world with these powerful tools.
AL: Gotcha, nice. Thank you for your hot take, I appreciate it. I put you on the hot seat there!
SirenOpt is Hiring!
AL: Do you have anything else top of mind?
JO: We do have some open positions right now here at SirenOpt. So if you are an electrical engineer with some high-voltage experience or a firmware engineer who knows about FPGAs and microcontrollers, please go to our website and apply.
SirenOpt.com. It’s easy. Or on LinkedIn, or email me.
AL: Alright, that was awesome. We covered a ton of ground. This is super informative. I look forward to seeing how things go for you.