The Data Science Renaissance

In this episode of the OnTrack Podcast, our host Zach Peterson sits down with Geoffrey Leeds, the founder of LEADS, LLC. If you've ever wondered how production managers react to integrating data science into production operations, or if the term "data-driven decision-making" has piqued your interest, you're in for a treat.

Geoffrey shares his journey from his time at Insulectro to venturing out to help companies embrace data science. The discussion dives deep into the challenges and opportunities of applying data science in manufacturing, with a focus on improving decision-making and efficiency.

Whether you're a seasoned professional in the field or just starting to explore the possibilities of data-driven decision-making, this episode offers valuable insights and perspectives. Join Geoffrey Leeds and Zach Peterson as they navigate the intersections of manufacturing, data science, and the exciting potential for innovation in the industry.

Don't miss out on this engaging and informative episode of the Altium OnTrack podcast! Subscribe now and stay tuned for more discussions with industry leaders and innovators.

Listen to the Episode

Download this episode.

Watch the Episode

Show Highlights

• Intro
• Return of Data Science
• Bad Data & Improper Recommendations 
• Automated Process Solutions
• Capturing Data
• IIoT
• Onshoring & the Winds of Change
• APlethora of Processes
• PCB Packaging 
• Manufacturing is Fun & Challenging Again
• New Data Solutions 

LInks and Resources

• Watch previous episode with Geoffrey Leeds: Materials Science and Manufacturing of Better PCB
• onnect with Geoffrey on LinkedIn
• Learn more about IPC-2591, Connected Factory Exchange (CFX)

Transcript

Zach Peterson:

How would a production manager or someone working on the floor, how would they react to the idea of bringing data science into production operations? Do they see that as being something that is, I guess, overly difficult for them? Because like you said, if they don't know the difference between mean and median, and then you say, data science, how do they react to that? Like, way over their head?

Geoffrey Leeds:

It's been something that I've encountering quite a bit of. And just like anyone starting their own business, there's unexpected challenges, right? You're gonna hit, you know, the grass is always greener on the other side until you start really dealing with how to, you know, put together an insurance policy and data protection, even internally, right? So when I got to actually start working with, you know, potential customers and clients, that was like the prime thing that I started hitting early on was, you know, misunderstanding and education of what these tools can really do. And truly defining what is the business problem from a production engineering standpoint.

Zach Peterson:

Hello, everyone and welcome to the Altium OnTrack Podcast. I'm your host, Zach Peterson. Today we're very excited to have Geoffrey Leeds back on the podcast. Geoffrey Leeds is the founder of Leeds Engineering and Data Science. We talked to him last year, right around the time of the Advanced Packaging Symposium, and this is gonna be a good time for us to catch up and see what he's doing today. Thanks for joining us, Geoffrey.

Geoffrey Leeds:

Happy to be here, Zach. Thanks for the invite, I appreciate it.

Zach Peterson:

So, since you and I last talked on the podcast, you've made a little bit of a career change, as I understand it. Why don't you tell us what you're doing now, 'cause I think it's pretty interesting and I think the audience will find it pretty interesting.

Geoffrey Leeds:

Yeah, appreciate it. So last year, I was working with Insulectro and since then I've decided to venture out and cash myself to the business wins, so to speak, and start my own company, Leeds Engineering and Data Science. And the idea behind it is that, we're looking, my partner and I are looking to go out and help companies understand, one, how to bring a true data science modern backend into their company, if need be. But also, two, providing our engineering services as consultants. It's very topical right now, considering the hype around artificial intelligence, and there's a lot of companies just trying to understand how do you define a problem such that bringing in a data science team makes sense to "deploy" AI. So that's what we're out here to try and help people understand and see where the world goes.

Zach Peterson:

You know, I feel like, ever since maybe 2019, 2020, I have not heard anybody talk about data science. Everybody's talked about blockchain or AI is kinda the real cutting-edge areas to work in. And it seems like right around the end, you know, right around the beginning or middle of 2018, data science, just kinda, the talk of data science just went off the map. What caused you to get into to doing this and what was the thought process behind this?

Geoffrey Leeds:

Yeah, so great question. And you're absolutely spot on with the point around data science fading to black. It really started for me, I've always been a thinker and an inventor. I still have an invention from when I was playing around with virtual reality and working on my own kinda input device, 'cause I was frustrated with some of them. But for me, it was seeing that, in the manufacturing space, especially coming from TCD manufacturing and being on the distributor side of it and really seeing what's going on, interfacing with designers, manufacturers, there's a big part of the puzzle that's missing in terms of the conversations around engineering solutions. And a lot of that comes down to being truly data-driven. And even in a manufacturing company, when I started out, as a process engineer, you're constantly focused on, you know, SPC, Statistical Process Control, right? And yet half the people in the plant, when I would ask 'em like, "What's the difference between mean and median," right? They kinda look at you and the eyes glaze over, right? And I remember this one time I was presenting a study that I had done and I had 60 samples and I had one sample that was, like, way out there. But because my sample size was significant, could you say, "Hey, this is an outlier. " And yet, the whole room was like, "Well, no, you have that one data point.

" And I'm like, "Okay, but do you understand statistics? " And that right there was the wall that I kept hitting, transitioning into being a distributor, working on a more macro scale. When I was working at, for a while, I was with Macdermid Enthone, doing process engineering essentially, or servicing their chemistries that we sold in that area. I'm working with different engineering teams across the board and realized that there's a pretty big gap when it comes to applying meaningful statistics and engineering practices in a manufacturing setting here in the U. S. And my cousin, who is my partner who's jumped into this venture with me, he's a statistician. He got his PhD and did his postdoc at the University of Chicago. Sorry, Bill, if I messed that up. I'm trying to remember his bio as well. But he jumped into the world of data science in 2012 when he came out to San Francisco. And at that time I was in San Jose, and he's been with some of the most incredible companies that have deployed data science in really meaningful ways.

The biggest one that I can rattle off is Global Climate Corp. They were one of the first "unicorns" in the Bay Area that sold, so it was over a billion dollars in valuation. And he led their data science team for a little bit after they were acquired by Monsanto. So from there, he's gone on to different companies. And again, he's always viewed it as data science 'cause that was where he came from, from that lens. And I would see, whenever we would get together how he talked about data and statistics and then seeing how it's applied in a really interesting way. They're going 10 tenths with statistics is really kinda the foundation of it, right, when we start talking about machine learning. And I always sat there thinking like, "There's gotta be a bridge between that and what I'm doing here, and process engineering and then manufacturing. " And circa 2019, when I started in Insulectro, we were trying to bring in distributors. It was actually, or sorry, suppliers. It was one of the first things that I tried to go do was find a, you know, an "AI" company to bring to our manufacturing partners.

Because I saw it as a potential way for us to actually meaningfully wrap around how to use statistics and say, "Look, if we take this concept of SPC and go one more level, we get access to tools and methodologies that could really," for lack of a better term here, I'm gonna borrow one from Tim Ferriss, force multiply our efforts without having to rely so much on how we analyze data or collect data and meaningfully use it in the manufacturing setting. And for about four years at Insulectro, I would find companies that weren't interested in jumping into this space. Now since then, there's been a number of companies that have realized that they need to bring machine learning in or machine vision is probably the first entry point you're gonna see it hit the manufacturing space. But it feels like a lot of people are trying to apply a solution without understanding the problem or looking at it from a truly data-driven way. And the core idea of data science is something called DDD, Data-Driven Decision making, which is a term that is blatantly misused. But if you can apply it properly, it is incredibly powerful. And that's what I wanna go out and do is try to find these small to mid-size firms and help 'em get started with it. Because once you get started down that path, you can avoid some of those pitfalls early on and get to a point where you can actually start bringing in a team to really deploy what data science is all about, which is quite simply just improved decision making, right? Improved intelligence. And how you apply that, is where we get something like an LLM, like ChatGPT, where people see this amazing "artificial intelligence" not realizing that the iceberg of ignorance is pretty substantial how to even get there. But it can be powerful, it can be used, but you just gotta approach it the right way and that's what we're here to help do.

Zach Peterson:

You know, it sounds like while you're at in Insulectro, you're trying to bring some DDD into the process of working with customers. Like, what was the idea to be able to recommend materials that could address common defects that may have been found from customers in certain situations? Maybe they're buying one material thinking, "This is gonna work best for the high voltage products we make." But they keep having defects. You now have a tool or a process by which you could possibly recommend an alternative.

Geoffrey Leeds:

That was the most lot, I mean, like, you're spot on with it. It was what kept me up at night when I was managing the flex product line was, "What if a product comes out that can suddenly point to our materials don't perform? " And what if it's been trained on bad data, right? Like the way it's giving a recommendation or prediction is incorrect, because it wasn't set up properly or it wasn't given enough context to understand what's going on. It's like when a new designer's trying to route high speed signals of a split ground plane, they don't know any better, but they just made an antenna, right? That's what kept me up. And having someone in the family who's been so intimately involved with it, I started poking around is into what is a solution to this idea of improper recommendations from a material-centric point of view. And that led me down the path of that iceberg of ignorance, of jumping below the water line, and suddenly realizing there's a lot more to it. And the more I sunk into it and credit to, you know, big thanks to Insulectro for letting me kinda go and investigate it, because we learned together. But it is, it's pretty substantial in terms of what you have to put together and how you have to pick apart that problem and the way you pick apart that problem was very different than anything I had encountered up until that point in my career. It was all manufacturing-centric, engineering-centric. We're talking about materials, we're talking about computer engineering, we're talking about electrical engineering.

And when you approach it from that data science perspective of trying to understand how to pick apart this problem of materials recommendation, which is something I would do all the time, and then trying to be truly data-driven with it, it is a very different problem. So how do you bring that into manufacturing on scale is not easy. There is no copy paste solution that's gonna work, it's going to be almost tailor-made. I suspect, over the next, probably five to 10 years, is this technology really starts to evolve, but it's always gonna require a certain amount of tuning. And while at Insulectro, that's when I kinda realized that this is what I really enjoy doing is really diving into these deep problems and figuring out how do you solve them? Because there's not enough people to figure this out, which is great, but also there's a lot of people who could be using this in a way that is incredibly beneficial. I think about 60% of what a process engineer does today from data collection and initial analysis could be completely automated with some of this, you know, the tools and skill sets that data science can bring to the party. But if you don't have your backend structured properly, if you haven't approached it from that mindset from the beginning, you're gonna end up with, you know, spending 50, a 100, $150,000 on solutions that don't have an ROI, that don't yield any fruit. And then your company could completely miss the boat on how to deploy something effectively. And I think one just has to look at TSMC who's got a pretty decent-sized data science team that doesn't publish a lot of research, and yet is at the cutting-edge of semiconductor manufacturing. They figured something out early on and that's what I wanna help do. But on a higher variety of scale, I should say.

Zach Peterson:

Sure, sure. And you mentioned that maybe 60% or so of some of those process operations could be automated through the use of some kinda platform or tool that helps automate, analyze the data, provide insights, and then, you know, the human makes the final decision as to what to do. That's really interesting, because I think part of that really hinges on a good UI for interacting with data in order to really get the insights quickly and make those decisions. And so I think that brings up another question is how would a production manager, someone working on the floor, how would they react to the idea of bringing data science into production operations? Do they see that as being something that is, I guess, overly difficult for them? Because like you said, if they don't know the difference between mean and median, then you say data science, how do they react to that? Like, way over their head?

Geoffrey Leeds:

It's been something that I've encountering quite a bit of. And just like anyone starting their own business, there's unexpected challenges, right? You're gonna hit, you know, the grass is always greener on the other side until you start really dealing with how to, you know, put together an insurance policy and data protection, even internally, right? So when I got to actually start working with, you know, potential customers and clients, that was like the prime thing that I started hitting early on was, you know, misunderstanding and education of what these tools can really do. And truly defining what is the business problem from a production engineering standpoint that we could use potentially, data science to help with. And from their perspective of, you know, if you're a production manager and you're trying to push product from, you know, one point to the other, a lot of these manufacturing execution systems in circuit board shops today are either still done on hand or in someone's head. So the idea of being able to come in and use something like data science, which is incredibly dependent on, you know, data in a computer, it's so over their visibility that to them it's just, you know, it's like, "Okay, you're promising me sky cake right now and I don't believe it," and rightly so, right? But how to talk to them and say, "Okay, what is a problem that you are dealing with that you just don't enjoy doing anymore," right? And it's sitting down and figuring out, "Okay, I have a target of how many boards I have to push out. " And they're thinking in their head, from their own context and what they've done, what they need to do and how to push it through a shop. And what they're really doing is behavioral analysis, right? In their own head, which is a whole field of study, I'm not gonna pretend to understand, but it is something that data science has tackled quite effectively.

Go look at any social media network, for example, and how they can predict who your friends are and what you're doing. And, oh, pretty good guess that if you were talking to Susan, you probably were talking about couches without even turning on your microphone, they can get a pretty good idea. So, you know, if you look at it from that lens, it's, "Okay, well, what tool would be helpful for you," right? And then working your way back through what is their current workflow, how are they collecting data? 'Cause if you do try to strap on any kind of, you know, let's just say modern Azure backend to what they're currently doing in their workflows, without really working with the people on the floor and breaking down the problem from their view set, you're not gonna get buy-in, right? Into the organization. And that's a huge problem across the board. Any change is always gonna cause problems. But specifically for those production managers, it's such a, you have to really get down into the details, into the nuance, to see and show them the value that without hearing them out and hearing what their problems are, to them, "Yeah, data science is, it's like, cool. ChatGPT, I can have a conversation with it, but how's it gonna help me plan when board A is coming out and going through, you know, sequence X, Y, Z? " It's a lot more challenging than even I thought, but it's a lot of, a lot of education.

Zach Peterson:

I think there are probably some production managers who are, I guess I don't wanna use the word "smarter," but maybe they understand that this is important, because they can link the analysis that's done to the value that's been created for other companies out there in the ether. And so they know it's important. When you say that there's value that can be created, they believe you, but then maybe to them they see a challenge in implementation. How do we get this into a process? 'Cause that's two things, right? There's the human process around it, but then, like you say, you have to actually capture data somehow. And so I think they start wondering, "Well, how do I even get data into this system in order to do any of this data science magic? And how do I do that when it's already tough enough to find business and keep the doors open and keep the lights on, you know, how much is this gonna cost me?" I mean, those questions start to come up, I think, and that's where the, you know, the rubber really meets the road.

Geoffrey Leeds:

Oh, yeah. It's, especially in manufacturing, where costs are, you know, they're an acute pain point, right? Their margins are thin enough as it is. It's a big reason why the manufacturing industry has gotten up and gone to, you know, the Pacific rim, especially for microelectronics. And that's where, I'd say data-driven decision making really comes to the forefront, that these manufacturing companies all say they're data-driven. And yet, when you walk in and you're seeing how they're doing some process by hand, and you ask the simple question, "Is that repeatable? " And they're like, "Oh, yeah, it is. " I'm like, "Okay, do you have a Gage R and R, or any kind of information that we could use? " And I've had one instance where the person say, "Oh yeah, we did a Gage R and R and they're good. " I'm like, "Oh, can I see the data? " Right? Just trust, but verify. It's been a great expression. I don't think Ronald Reagan coined it, but he definitely used it, brought it into its moment. And no matter how I asked the question, they didn't want to admit that they did do a Gage R and R. They don't have the data available for you to use it. And then for firms that do have a lot of data at their fingertips, they're not using it in an effective way. I had one customer who had beautiful data that was show highlighting a problem quite quickly and actually pointing to where the solution should be. But the way in which the team was approaching how to go handle that problem, it was suddenly the data isn't important, it's my opinion on the data. And then that sidetracked the whole investigation for like, six months, because it was people's opinions about what was going on, rather than saying, "Okay, let's break this problem down into a bunch of, into a series of small digestible blocks. " In each of those digestional blocks, you know, have been solved before or they haven't. And really parsing through the problem from that perspective and saying, "So where do we have data," right? "Okay, where don't we have data? " And then you enter into the data mining phase, which is often overlooked. And that's been codified, like, 20 years ago in a process called CRISP-DM, Cross-industry standard for data mining. And it's pretty efficient at figuring out how to go and collect data and tying in business perspective and needs in order to going and getting it.

So if you don't have that data, how much is it gonna cost to actually get it, and how do you start going through that iterative cycle? So there's methods to go and figure out those problems, but when you start looking at it from the cost perspective, that's something a lot of them don't know how to do. And that's an area where I get guidance from Bill, if he wants to look at it from like, "Okay, here's what it's gonna take for us to set up a cloud or modern architecture offsite or onsite. Here are some of the requirements we have. " And then you start looking at it from like a standard engineering practice. And when you can start feeding those variables in and plugging them into classical business, that's when a business will say, "Oh, oh, data's an asset. " Okay, so if data's an asset, then here is the infrastructure that I need to support that asset. Now we can start looking at how this ROI or what project we want to go do can actually affect our yield or our consumption of chemistry. There's a way to go from ethereal, we don't know what this looks like, to concrete results today, and it's been done time and time again. But for most companies that figure that out, it's just like when you figure out Micro-V reliability, how many of 'em are out there shouting from the rooftops, "Here is the process and here are the DOEs. " You need to do internally to guarantee. Not a lot of them, it becomes like a trade sequence. They don't wanna share that information, but the overall methodology of getting there is very well-understood.

And it's something that I struggled with and got to a point where I was able to start using this technology. And there's a way to help other manufacturers get there, but it really does start with, you know, to your point, taking a step back, saying, "Look, we don't have this data. This is kind of what we'd like to do. How do we even begin? " And that's exactly where we can come in and say, "All right, it's been done before. " Right? Like, "And here's how it was done in the past, but we might need to tweak it to your business case because every business is slightly different. " If every business was the same, it's like, "Okay, we already have a single winner takes all. " So there's gonna be no copy paste, but there is a way to go and do that. There are tools out there to help you, and we just need to get those tools injected into the manufacturing stream and it can be done. It's a long road and it's not cheap. And that's, I think the biggest thing people overlook is cost of data acquisition, data storage, model training, and then the whole other beast that is actually deploying and monitoring some of these predictive systems or collection systems you want to deploy.

Zach Peterson:

So on the collection side, there's already one buzzword that deals with this, and you probably know what it is. It's IIoT. I think someone might look at this and say, "Well, isn't this just IIoT?" One of the criticisms of IIoT that I've read in the past is that it's kind of a pipe dream. You have to deploy all these sensors everywhere, figure out how to make 'em all interoperable. I know there's the IPC-CFX standard. I'm not up on how that actually gets implemented in a practical sense. So I think that presents a whole other set of challenges just on the acquisition side. And as soon as someone says, "Well, how do we acquire all this data? We have production operations running 24/7, are you expecting me to pull someone off the line and just, like, manually key in data into some system and then voila, data science?"

Geoffrey Leeds:

Yeah, no, that's a, that's a really good point to bring up from the IoT perspective. So, you know, for big data companies like Google, like Facebook, like Amazon, their whole system is built on infrastructure that has ease of access to data. And their evaluations are in large part driven by their data assets that they have and how they utilize it. And unfortunately, for manufacturers who all claim to be data-driven, they are currently writing down data, at least the good ones are in logs, for example, and they're writing down, "This board went through at this time. " And all the process engineers and manufacturing people in the background are screaming, "Well, yeah, what about pencil whipping," which is where they just don't write down what the sensor in front of 'em is actually saying. They'll just write down what the person before them did on the piece of paper. And I've watched it before, it's wild. So that becomes like a data confidence, like how much do you trust this data kinda perspective. And again, that whole field of research and how to use that information. 20 years, 30 years of research on how to use that from a big data perspective, but in a manufacturing sense, very unknown. And when you look at IoT, right? You know, the joke is like the S in IoT stands for security, right? 'Cause these devices aren't inherently secure at a foundational level. And those problems though have not gone away, but they are being solved slowly but surely. I remember in 2010 when I took my first computer engineering course, my professor at the time was talking about how everyone's got the rage over IoT. But let's talk about these infrastructure problems. You're just gonna have all these devices, millions of them go to the internet and send a packet out at once. Like, how are you gonna deal with that with that networking issue?

Now, 2023, we have the ability to collect massive amounts of data. And that problem has been solved, pretty, pretty well at this point, I would say. Your cell phone is pumping out an enormous amount of data every second, and it's being collected on some kind of an interval. And all the big telecommunication companies know how to handle those issues. Unfortunately, on the board side, we've had to deal with them head on with a signal integrity. But when you start looking at how do you deploy an IoT sensor itself to the factory floor, there comes a whole host of issues with that. And those issues require you to have a network in place that can handle IoT data infrastructure. And you don't wanna put that on other networks, because all the IT people wanna have separate networks, which is a good thing. But now we start adding layers of complexity in terms of your security policy. So if anyone who's trying to go to CMMC right now, right? They're looking at the NIST 800-171 standard thinking, "How am I gonna implement a secure IoT collection system? " And fun fact, one more Google search and you'll see a NIST document for how to do secure IoT in smart manufacturing. But when you're struggling to just do the basics for, you know, CY, and how do you handle uncontrolled classified information or class, yeah, unclassified controlled information. It's like another layer of complexity. Is this a priority for us? So when you start adding on all the different problems, most of these manufacturing companies in the U. S. that say, do have a certain level of data intelligence, right?

Look at that iceberg and say, "There's no way this is gonna get deployed right away. " And likely won't be. If you don't see the value in collecting and using data that's at your fingertips, you know, no amount of "AI" showcase or, you know, ChatGPT, like showing you how to build a PCB from you just telling it, you know, "Here's what I want," right? No amount of that will get them to flip over to using data. So it presents a really interesting challenge that has to be driven from the top down, and really understand that you can use data effectively, but as a firm, everyone kinda has to get behind the fact and say, "Let's check our emotions at the door, and really say, 'Look, there's 20 years of tech companies that have been using data and data science to be wildly profitable and achieving massive amounts of success in what they do. '" And distribution companies like Walmart, like Target, have big data science teams to know things like, if a hurricane's coming, I don't just need to stock up on beer because people are gonna be stuck inside. They can predict what products people are going to buy and when and where it's gonna hit. So like one of Walmart's big things they revealed, I think it was circa 2011, was they started data mining. When a hurricane was coming through, what kind of products were being purchased in higher volume?

And they found all kinds of things. Then you have Target who knows that if they can get someone who is pregnant or about to be pregnant to start buying products from them, likely they'll be able to sell them everything else. So trying to show the business leaders how data science can help you increase sales is probably the easiest way to get people to accept data science. But then going down to the shop floor, it's gonna be, "Hey, this board you're trying to build, 30-layer board, all these different complexities and nuances to it. " If you run it within your set temperature windows, you're only gonna get 80% yield like you always have been. But if we do some unstructured, if we do some data mining and we start doing maybe some unstructured learning on all the available information, could we figure out how to do yield improvement? And if we get yield improvement to let's say 90%, how much does that help you in the manufacturing space? That's where they will say, "Okay, that sounds like a good plan," right? But now how do I go and deploy all these IoT sensors to get there?

Well, you gotta sit down and kinda map out the whole process and do data traces for all the information that's gonna be required. And unfortunately, there isn't a lot of research in this space, but that's also kind of the fun of going in and jumping into this world right now from my end is that, we get to go do that. It's very much still emerging in terms of how do you deploy IoT effectively. The CFX standard from IPC is an interesting solution. It's very focused on assembly. I'm on a team right now. We're actually looking at how to deploy potentially CFX or some of these other communication standards in a factory. But MQTT seems to be the lead right now, but again, it's got its own issues from a security standpoint. So it's, sorry for the ramble there, but it's not as easy as one may seem, but you can do it quickly and you can go by an Arduino today and hook up a couple sensors, and within 30 minutes, you can have that data available in an API call.

How to do that safely and securely is a whole another equation that I've been working on. And we've got some proof of concept demos, I've got this whole other project I'm working on with the DOD right now. We're pretty neck deep into it. And once we have a general solution out there to the public, I think that's when manufacturers in our ecosystem will start adopting it. But for the moment, most people are just buying like a Siemens or a Mitsubishi type IoT system that's already available or like a Rockwell Automation one, and they're just gonna pay handover fist for it, because they have a product that can afford that. But in our space, you're gonna have to be much more tactical. And what that means is you're gonna have to actually have some talent on the floor that understands the ins and outs of that ecosystem to deploy it effectively. So it's not solved. Well, there's some companies that do it decently, but there's not enough gray matter in the PCB world to be able to do that on scale at the moment.

Zach Peterson:

You know, you're talking about yield and the price of some of these solutions and, you know, just buying like a Mitsubishi solution or something like this. And I think that brings up a really important question, "Who's this all for?" Because let's go back to yield for just a second. Let's suppose you're a quick turn manufacturer, right? A quick turn manufacturer will be comfortable having a 90% or even an 80% yield. They'll just build in whatever the cost is for rework, because they're not marketing based on yield or they're not producing based on yield. They're producing based on, "We'll get you this thing in 48 hours if you really need it or if you're willing to throw enough money at it." So that's what they're selling. They're not selling necessarily the, you know, high yield. So for them, improving yield isn't really something they are necessarily gonna be driven by. And they're probably operating on smaller orders anyways. So why are they gonna invest in all of this in order to try and get some data insights if it's only gonna affect, you know, a small part of their business or if the need is constantly changing?

Geoffrey Leeds:

Yeah, no, it's a great question. And I would say, by and large, most of the board shops that I see here in the U. S. , a lot of them try to book the shop up to about a 100% capacity, maybe a little bit more. And then the idea is that when you do get a quick turn like that, it becomes the priority in the shop. So now that goes through all the processes first, you know, it didn't care about your production plan and what was gonna happen to it, right? We'll figure out that afterwards, how to meet those deliveries. This is the money maker, and it is, right? Most of the U. S. marketplace is quick turn focused.

And right now, from when I got into the industry in 2015 to, I would say about two, three years ago, right when COVID started to hit, that was the predominant mindset for most PCB manufacturers. Since then, the winds of change have hit pretty hard and now there's a huge focus on onshoring, right? And how do we, or friendshoring is another term that people use. So in the PCB world or centric to, like, the electronics manufacturing ecosystem, we now are being faced with, well, now we're starting to see some decent volumes really start to come on shore. I remember before COVID changed the world sitting down with a company that said, "Hey, we can't get anything from our fabricators out of China and this is February, and we need to get this material in the U. S. " And it was more material than I've sold and I would sell in a whole year. And they needed it in one order. And I was pretty baffled at that scale. And the U. S.

DOD has taken a pretty high note of this as well, right? Like, the CHIPS Act was really pushed in large part because of the fact that the supply chains were shown to be so weak because we have no capacity here in the U. S. to build volume at any kind of scale above the prototype quantities and maybe some small medical and aerospace that we're doing. And there's a huge demand from the commercial market segment and other market segments to actually push some decent volumes here that are changed to the rest of the world. But their risk management plans are now starting to change that equation where board shops no longer care about, "Hey, I'll just build, you know, I'll release 50 of these boards to the floor. " I'll only get 30 that come out great, but who caress, our margins on this are 60, 70% or whatever it might be, it's okay for us to eat that loss. That business equation has shifted entirely and that's a big reason why I have kinda jumped off on my own was because my background's in robotics and computer engineering, and every manufacturer that I would go into had no robotics or automation whatsoever. And yet we're pulling, you know, the next generation smartphone or high-speed server network coming out of some of these pieces of equipment from the 1980s. And I was actually having this discussion with some colleagues of mine a couple years ago, 'cause I was all about automation, you know, physical automation being the biggest way that we're gonna make changes. And there was some people saying, "Well, what about data automation?

" I was like, "Data automation, like who caress about data," right? Like, don't get me wrong, it's important, but for production, robotics is really gonna help us get there. But every board shop owner that I talked to or anyone who made those business decisions would say, "You know, the ROI has gotta be below three years. " And it's like, "Okay, well, that's not too difficult to get to. " And now that ROI has got even less because now there's not as much people out there anymore. I had one individual tell me that it's no longer whether or not I get the, you know, the cost of the ROI. I now have opportunity loss because if I don't have a person helping me run the line, then I'm not shipping boards out because there are so many more orders that are coming through. So that viewpoint, you know, like, the business problem you structured, in the U. S. has changed pretty rapidly in the last three years. And I've been very fortunate to be in a position where I gotta see that change across the landscape to where now yield is a big problem, especially for those that are trying to get into substrates right now.

Learning how to improve yield when you're doing like a 15 micron line in space redistribution layer and you're having poor adhesion, that's not an easy problem to solve. It's not trivial. There's a lot of factors you have to start considering and you gotta apply some real engineering brainpower to this. And some of these quick turn products that are coming out now have such just amazing density that we truly are getting to substrate like PCBs where research and development and process improvement is becoming the new bottleneck for them to be able to go after those high-yielding jobs. So from that perspective, that's where you're really gonna start seeing data science come in and say, "Yeah, we're not gonna be able to," you know, for you to say, "All right, let's take this quick turn job. You need to release 22 cores for this layer, 13 for that one to get, like, completely optimized. " You might be able to get there with enough time data and money spent, but if you've done your homework on the back end of it and you've improved your process, you can actually get to a point where you could predict the overall ability for you to build said job, or to figure out what data do I need to improve my process so I can make this job? And that's where being truly data-driven becomes the linchpin for a manufacturer to be able to go from 10-nanometer down to three nanometer, or one fab figured it out and they use all the tools at their fingertips, and one of 'em can't get the clean room, clean it off, right? That's where data becomes invaluable in terms of your R and D process. But if you aren't, if you haven't even started that path, it's gonna look like a mountain, right? But it's a, that overall perspective, it's weird for me to see it change in the last three years so rapidly, but that has been what I have seen, at least from my perspective.

Zach Peterson:

This is really interesting that you brought up packaging because whenever I talk to anybody about packaging and packaging capability in the U.S., they always seem to be talking about like the actual production equipment for packaging. 'Cause like you said, some of these companies are still operating with 1980s or 1990s maybe equipment, and, you know, they're not gonna go after super high-density packaging or substrates or anything like that, right? So the packaging capability is often talked about in terms of the actual production equipment, not in terms of the data and how it's leveraged for process improvement, but it sounds like one of those might be a little bit more important than the other, because you could throw all the money you want it, packaging equipment, right? To actually produce stuff. But if you constantly have yield of 50%, let's say, really terrible yield, then what's the point?

Geoffrey Leeds:

No, it's an acute issue. And in the packaging world, they're now that the expression is more than more, right? Because we've hit the limit on the IC side, right? Another fun fact, more in the 60s.

Zach Peterson:

The second more- The second more being Gordon Moore, yeah.

Geoffrey Leeds:

Yeah. And it's interesting though because the advanced packaging world, I don't wanna say I understand it right now because for the last year, the more and more I've dug into this ecosystem, I've learned more, but I feel like I understand it less, is probably the best way to describe it. Because the packaging world will use terms just like the PCB world, almost interchangeably and improperly, even though they do have more engineers available to them. So they'll figure it out. But the equipment manufacturers only know this one piece of the puzzle. And when you start looking at what is a circuit board, it's a composite, really it's a composite material that happens to be electrically conductive, right? That's really what you're making at the end of the day. You have to line up so many different processes properly in order to build the board. That one piece of equipment isn't just gonna be your end-all be-all savior, it'll solve a bottleneck for you for sure. Just like LDI systems solve a huge image for companies to start using lithog, go from lithography to laser developing. And while that was a big step in terms of density for us, well, can your Etcher etch those traces that you're imaging in?

Can you develop on them properly? Can you control your developer properly? All these other little side details that didn't necessarily rear their ugly heads. If you're doing five mil line in space, like, "Okay, that's pretty good," but when you go from five mil to five micron, right? That's very, very different. And suddenly, these small nuanced issues of a 60% break point versus a 45% break point can be the difference between residue being left over and causing issues with your RDL getting etched improperly and all kinds of weird being, void, nonsense, that if you have not done your homework and characterized your line and gone through it with a fine-tooth comb, you know, the ability for U. S. manufacturers to jump into the packaging world, I should say the substrate world 'cause that's really the piece of the pie they're trying to go after, you're gonna be unsuccessful. But if you do it properly, there will be some, let's say if you get down to two mil line in space, there'll be some substrates that use that size, right? As an interposer, and it's like, "Okay, you could build that. " But if you wanna take that next step, do you just build a whole new line and a whole new plant because you don't know where in your process you will be able to optimize and be able to handle this?

Nothing's in a clean room. So good luck even imaging anything below 25 microns, right? But that perspective is almost entirely lost in the U. S. right now, and they're trying to rediscover how to do that. And that's what the U. S. military, or I should say the Department of Defense is interested in, in helping out right now, which is one of the contracts that I've been retained on, is helping to understand how to bring something like MSAP capabilities to the U. S. PCB market to even get them to the point where they can even get within a Hare's Breath of let's say, two mil line in space 'cause a lot of them can't even get there. But the packaging world equipment sets are very good at small tasks, right?

Like, you can go find platers that cost 1. 5 to $2 million, that'll get you a flatness across your RDL of, you know, plus or minus, you know, 10% or less, and you're only plating 15 microns, which is pretty unbelievable. But that's $1. 5 million, all right? And you might've been, you have this acid copper plating bath that you put 50 grand into, right? So we're talking another order of magnitude to get there, but just because you have that piece of equipment capability doesn't mean you've done anything else in the process to be able to support and really get out of that piece of equipment what it can do. Even for some of the LDI systems that Orbit Tech is selling, that have the capability to get down to 25 micron line in space. I think a lot of us companies have them, but they're still kept out at three mils or 75 micron line space, because they can't get over the barrier required to get to the point where you can start shrinking your density and handling the equipment properly to get there. And no single piece of equipment will do that for you, unless it's all kind of like in one giant room, or, you know, like one giant piece of equipment. I guess you could have like a semi-truck with five or six of these things lined up. But it's gonna require some serious effort to even contemplate how to get into the interposer market.

And we're seeing it though, which is really cool. And some PCB board shops, they're figuring out that these package designers that are no longer dealing with the time domain, they're now dealing with signal integrity issues are comically similar to what we're dealing with on the circuit board. So how do we work together in this world? And that's where circuit board facilities can really help with the advanced packaging movement. But you've gotta get close to the density before you can even walk into the room and start talking about surface roughness and signal integrity and how you're, you know, back drilling your stubs. Yes, that's gonna cause an RF interferons, don't get me wrong. But when we start shrinking and starting putting down these five micron lines, five microns from each other, yeah, cross stock gets really, really weird when you start packaging everything that close together, which is why people start using Ansys HFSS to do their analysis and not just back of the Napkin Calculators. But to even get into those rooms you have to know what you're doing. Otherwise, it can be very expensive when we're standing at the MRB cart, trying to figure out what happened.

Zach Peterson:

Yeah, packaging stuff is fun and I'm lucky to have just started doing some of that myself. And it's interesting, because you just said, you know, there's the correspondence between now when you get more dense you start to notice things that the packaging folks have known for, who knows how long, right? 10 years, maybe even longer? Especially in terms of SI. When we were at PCB West, I was talking with Rick Hartley briefly and it was about stitching vias on differential pairs. And I didn't jump into to the packaging part of it, but once you start to look at what actually happens in the package on fine pitch BGAs, you start to realize, "Oh, that's why I need stitching vias on my differential pair." And no one, I mean no one in PCB land knows about that.

Geoffrey Leeds:

Yeah. A beam waveguide is a really novel concept to see and then to think about a circuit board and say, "Okay," right? When you start putting in those micro views there, you're essentially trying to create an RF cage, right? And depending on the frequency of what you're dealing with, it becomes a waveguide, because now it's just kinda guiding the waves, rather than truly carrying it. And that you have to change how you look at energy, but it really does change how you have to look at the circuit board because that cage has some very tight tolerances that if you screw up, you get leakage, you get noise, you get really erroneous electrical behavior. And for me it was, I was interning with NASA when I was in college and looking at some of these K and S-band transmitters at their Deep Space Communications network out in Goldstone, California. So their beam waveguides are tubes, these square tubes, and it threw my mind off. Because I was like, "How is a tube a waveguide? Like, what? " And when you look at a circuit board, yeah, fun fact, it's the same thing. It's just the energy's moving through different medium now, right?

Like you've got dielectrics, you've got conductors that's, you know, carrying a lot of the voltage. So it's very, very different. And knowing how many vias you can drill in this tight, small area and your ability to, you know, get that distance and precision down to where you can guarantee it, buys you space. 'Cause if you're doing two different, you know, stitched vias next to each other, you're offsetting them, you're eating up space that you could be routing with, right? But if you can't hit the tolerances to be able to put those vias in the spot you need them to control your noise, then you're not gonna be able to jump into those designs. I would say it's a fun problem to solve 'cause that to me, is absolutely fascinating. And it's interesting, especially when you see some of these signal integrity analysises that are done where they spend a lot of time and money on. But it's a, yeah, it's a cool world to be in for sure. Very different than what circuit board manufacturers are doing today when they see TDR, and within my rise time, are we good? Cool, ship it. So what did you find the most interesting about looking at some of the packaging world problems and seeing the overlap between what circuit board manufacturers need to overcome to get in there?

Zach Peterson:

What I actually found really interesting when I first started learning about packaging was really how similar packages look to, like, HDI circuit boards just on a smaller scale. And then when I started to think about the processing, I realized how critical it is to get these processes correct just to get down to that scale. Because obviously, you know, you mentioned going from lithography to LDI, right? That's one step in the, I guess the size spectrum. And then there's this kinda gap between, you know, LDI scale, and then, you know, what they do in semiconductors, which is, you know, like EUV at this point. That gap in between to get that packaging and get that substrate even smaller, that's an area, for me, that's still kinda no man's land. Like, I'm still learning in that area and it's something that I know I'm gonna have to get better at understanding, especially when it comes to, like, DFM around it. Because, you know, like you said, packaging demand is coming back over here in the west and I've started to see a little bit of it myself and the conversation always seems to go there. So it's something that we all have to learn a little bit about. I could see a time, in the not too distant future, where more companies have an opportunity to take control of what they do with packaging, but they're gonna need designers that have more of that knowledge and especially more knowledge of the manufacturing process for it.

Geoffrey Leeds:

Yeah, I think that's a great answer. This has now become I'm interviewing you. Yeah.

Geoffrey Leeds:

I was just-

Zach Peterson:

That happens sometimes.

Geoffrey Leeds:

Yeah, it's always good to have a conversation. I was at the IMAPS Packaging Conference last week in San Diego, 'cause it's just a stone's throw from where I live. And that was a great conference to go to and I highly encourage any circuit board fabrication company to go to these conferences and just be a fly on the wall, because those problems and processing and from designing, all get addressed in these rooms. And you start hearing some really interesting perspectives from, you know, pre ASIC designers, right? Or Pre-Silicon designers, sorry, that are, you know, and then the post-silicon, AKA, what we have to deal with all at the same table. And then we've got the thermal person in the room, and then we've got the injection molded people on the other side who are actually putting the plastic over this and doing the overmolding. And then the assemblers who are technically or traditionally the OSATS, that are now having to buy some really sophisticated pieces of equipment to try to reconstitute some of these, you know, chiplets that are coming out. And let's not even get into the fact that chiplets really still isn't defined yet, even though we're seeing a lot of architectures coming out from these tier one design companies on using them overseas. So it's the. . . It's still the Wild West and there is a bit of, or I should say a big push around standardizations right now, which is really cool to see the NIST metrology R and D standards conference two weeks ago. I did that one online, which was awesome to be in. But, you know, in my online session was the president of engineering for Qualcomm. And I was like, "Oh, well, this is interesting, to get his perspective on this. " But their design loop is one year, they're pumping out a new processor, right?

They're putting out new variants on what they're doing, and that drives a lot of innovation that isn't occurring here at the U. S. So you're seeing a lot of companies that are trying to say, "Okay, how do I jump into this pool? " At the same time, those big companies are saying, "How do we get more designers that are here in the states, 'cause 97% of all these package designs are done here. " Like you said, Zach, experience with going into a packaging house or into a substrate facility and looking at, "Oh, before we even put alternative oxide on our signal layers, it's going through several other etches. " What are those etches doing? They're creating surface topology. Oh, that changes if I have ED Copper or RA Copper, huh, like, it's. . . We're at the point where you do have to have a certain amount of discipline, and I shouldn't say discipline, understanding of physics in general outside of it's aligned on a piece of paper that it's fun to be in manufacturing again. It's great to see this kind of drive towards what do we need to do to get there. And, yeah, big burden to overcome if you're a manufacturer and it does require a large amount of data collection. 'Cause overseas, they're struggling with how to do these 20 and 20 packages, which is crazy, but that's the density we're hitting to build a true heterogeneously integrated package. It's pretty wild to see.

Zach Peterson:

Well, getting back to the data point just for a moment, 'cause we're running a little low on time. But one thing I'm wondering on the data point is that I think as more companies realize the importance of being able to get access to that data and process it in a relatively automated way and being able to visualize it quickly and make decisions. I wonder the level of integration there will be between these, like, data operations and the actual equipment on the floor that doesn't require a third party solution like a Mitsubishi or a Siemens or whatever else. So you don't have to go buy that Arduino and hook up the sensors to it and then start worrying about security and then the tech stack that's behind it. So I'm wondering what that's gonna look like and if that's really gonna help a lot of companies maybe upgrade themselves to a point where they do start to see that ROI faster and then they can get into these more advanced areas of electronics, manufacturing, and possibly even packaging.

Geoffrey Leeds: It's not an easy road, it's gonna take a decent amount of investment, but just like upgrading your equipment, it will yield fruit if, and only if, you approach it with the right mindset. You gotta walk into it two eyes open. You have to say the data-driven decision making we've been doing in the past. We're saying, "We're data-driven firms." That's not quite the way it's done for good reason, and be open to hearing new solutions in different ways or how it's done somewhere else, rather than the, "This is the only way I've done it for the last 20 years to survive. It's the only way it can be done." It's gonna take a new perspective. It's definitely doable, but you just have to approach it with the right way and you have to have the right team to help figure that out. And if you do decide to go down that road, please give me a call.

Zach Peterson:

Great, great point. Well, I hope folks will check out the Shownotes and they'll be able to get in contact with you and learn more about your company. To everyone that's out there listening, we've been talking with Geoffrey Leeds, founder of Leeds Engineering and Data Science. If you're watching on YouTube, make sure to hit that subscribe button, hit that like button. You'll be able to keep up with all of our tutorials and podcast episodes as they come out. Make sure to check out the Shownotes to learn more. And last but not least, don't stop learning, stay on track, and we'll see you next time. Thanks everybody.