David Ariens

Welcome, welcome. We're so glad you made it into our breakout session.

Just want to do a quick show of hands: who of you works in or with manufacturing or infrastructure companies? Okay, yes. And who doesn't, who has nothing to do with physical things? Still really, really, really welcome.

This is indeed a session about digitalization in manufacturing. Before we start: obviously everybody here knows DevOps. It's the reason this conference exists, the reason why we are here. For those who are a little bit less familiar with operations, this is what operations looks like for us.

Hard to change by GenAI.

Let me introduce myself. I'm David Ariens. Both of us are from Belgium. I started my career 15 years ago in the chemical industry. I'm trained as a mathematical engineer; that helps nowadays. A couple of years ago I switched over from the end-user perspective to the consultant/vendor perspective, so I wanted to experience both sides. Currently I'm typically working with food and beverage customers and infrastructure customers in Europe.

Willem van Lammeren

Okay, I'm the other one. I'm Willem. I started 20 years ago as a chemical engineer, and for the past 15 years I've been working in manufacturing digitalization in some way or another, either in IT or on the shop floor. Today I work at a large chemical company, Syensqo, in IT, where we work and develop solutions for operations.

But David, today we're not here as those people. We are here as the writers, the authors of The IT/OT Insider. We're not trying to sell you something, so we don't have a product to sell you. We're not pitching any services. We just wanted to write and think about one problem that we're facing in our jobs. That problem is: why is it so hard to digitalize manufacturing? I know the people who work in manufacturing here, I don't know what your experiences are, but it's hard. There's a lot of complexity when you go on the shop floor.

So we listen. We talk to a lot of people, people with much more experience, people much smarter than us, and we try to learn from that. We'll share those with you like we will do today.

David Ariens

Absolutely. What we should start with is just to define that term OT. You've seen it in the title of this session. We mentioned OT, or the IT/OT gap.

Now you all know IT, and IT in manufacturing is the IT you know; there's nothing really special about it. But the thing that makes these processes run automatically is OT. That stands for operational technology. It's an umbrella term. It's a combination of hardware and software.

Together it makes sure that the crisps, the chips, are perfectly crispy, perfectly nice to eat because the speed of the belt is running smoothly. It makes sure that the welds on your car are at the right position and of good quality so the door doesn't fall off when you're driving on the street. It makes sure that there is the exact amount of medicine dosed in each and every vaccine or pill, or whatever that you have, so you have a safe product to use. In the case of the water of this wastewater treatment plant, it makes sure that the water we send back into nature is safe, is clean.

We all do these things with OT, with operational technology. It's basically what runs the world. It's what makes the chairs you're sitting on and the phones you're using. So it's super important.

Willem van Lammeren

We've been working in manufacturing for around 15 years, working on digitalization. To explain the problem that we're observing, I'm going to use the phone as an example.

In the past 15 years, the question is: what part of your life doesn't go through your phone? Your flight, your drive, your taxi, your love life, everything. Go on the shop floor and you will wonder what happened. It feels instead of having a modern situation, you end up with some sort of upgraded BlackBerry. That's what's happened in the past 15 years, unfortunately.

Does it work? I don't know, but apparently not. I'll go to the next slide for you.

This, I think you all know: DevOps, where we're trying to bridge dev and ops. It's really hard to get them together and work together, talk the same language. So if that was hard, bridging IT and OT is even harder.

We're talking about people reporting to different C-levels: the CIO, the COO. They're using different technologies, really different technologies. They're talking different languages. Finding that common ground to get IT and OT to work together is really hard in practice. The problem is that usually the ones left in the cold are the people that need those solutions the most: people in operations. I'm not even mentioning the other people, like in engineering.

We were looking for a picture with local flavor. I think this picture of the Grand Canyon is a perfect representation of the huge gap we're facing.

David and I, we've been working, talking, thinking about what exactly is making this hard. If we start to understand the underlying reason here, we can start to look for solutions. We've chosen three. Three is a great number. David, let's find out what comes first.

David Ariens

Rule one when you work in manufacturing: innovation is important, but it comes after "don't explode." I know that Facebook coined the term "move fast and break things." I hope it's clear to you that's maybe not the best philosophy in production in a chemical plant or a pharmaceutical plant. When things go wrong, we don't really have a Ctrl-Z or a rollback. When wrong doses are made or something explodes, you're stuck with the physical reality.

I think that's really important. It changes the way you approach innovation. It means that you will build in guardrails. You have procedures to follow. You need to do testing before. That's important from the moment you're entering that physical reality: your way of innovating is going to be different than when you're completely virtual.

That almost automatically brings us to the second rule. The second rule is that change is expensive. What we're not trying to say here is that change is easy in IT systems; it definitely is not. But if you take another example close by, if you take the Hoover Dam as an example, built 90 years ago, if you want to make a change, it's going to take you more than just one pull request. I wouldn't want to refactor the Hoover Dam. It would be a bit hard.

Willem van Lammeren

The third rule is a consequence of the second rule. In our projects, very often we know that scaling is linear, and there's a good reason for that. If you look at a plant, it's usually an expensive endeavor. It takes years to plan, years to build. You cannot deploy just with a click.

Very often plants are made according to their economic environment, according to the technology of the time, according to a lot of factors. That means that each plant is going to be unique. Each plant is going to be different, even if they're making exactly the same products. Each plant has a lot of complexity.

What it means when we're trying to deploy digital solutions is that at the moment we're entering that level of complexity, and we need to bring that into our solution, we're going to slow down really hard. Instead of growing exponentially, we need to slow down, and at best we scale linearly.

David Ariens

There we go. We hope that we gave you already a couple of pointers why this IT/OT gap exists. There are more things to it, but it's already a starting point.

That brings us to the question: why should we care about it? Why should you care about that gap?

On this graph we're showing you the difference in labor productivity output in the manufacturing sector, which you can see in red, and the overall labor productivity in the US. These numbers come from the US Bureau of Labor Statistics. If you just look to the last 15 years, that's actually the era where Industry 4.0, the digitalization initiatives in the industry, were introduced. You can see in the red graph that it stagnates or goes down.

That means the productivity in general, the average productivity, drops. Obviously, it's an average graph, so there will be pluses and minuses. But on average, the productivity actually seems to be dropping. That's a big, big, big trouble, especially now today if we were talking about reshoring and all these things. This graph shows that we seem to have a big problem.

On the other hand, we are visiting manufacturing plants, I wouldn't say daily, but at least weekly. When we enter a facility, we always see untapped potential. It's scattered around. You enter a facility and go: oh, we can improve this and this and this area. Not talking about one or two percent. We sometimes talk about 10, 15, 20 percent productivity increase easily.

That obviously brings us to the question: can AI help us here? It's an AI conference, and we do an AI talk.

Is AI then just another incremental step, another button on the old BlackBerry, or can it truly transform the industry? And if so, how?

We have a lot of concepts. I invite you also to take a look through our blog. One of these concepts is the way we approach industrial digitalization. I'm going to explain this because it shows you how manufacturing works. This is not our concept. Using a physical twin and a digital twin is a concept which is tried and tested, obviously, but we're going to use that to explain how we work.

On the one hand, on the right side, you obviously have the physical production assets. See the videos we showed in the beginning: where stuff is made. On the other hand, we have something we call a digital twin.

In many marketing slides, that's often shown as virtual goggles or super-complex integrated systems holding all data. The reality is that in 99.9 percent of plants, you don't have just one digital twin. What we actually have is a collection of systems sitting on the left side, a collection of systems all holding a small piece of the truth.

Some hold sensor data. Some hold production data, for example manufacturing orders. Some hold quality data, some simulation data. Even the ERP system is, in many companies, an essential part of the digital twin because it might hold supply chain data, for example.

You have operators and engineers interacting with both of them. They would use a certain system, come to a certain conclusion, and then apply that conclusion to the physical twin: turn the valve, for example.

Luckily, it's not all manual work. Luckily, we have an automation layer as well. Automation was invented in the seventies, I would say, or at least modern automation. This is really where the term OT comes from, the operational technology thing. What we do in the automation layer is the stuff which is deterministic. The stuff we can easily describe, we do that over there. For example, we automate the speed of the belt to bake or cool the chips.

That's where we are in general. Then the question is: where does AI fit in today?

This is a very interesting one, and I think it also shows you a bit of the problem we're facing. If we look today to AI applications, and there are a lot of them, and not just in the last years, there are some of them already around for decades because there is more to AI than GenAI.

Today if you look to these applications, which are sometimes really clever, really tailor-made for a specific use case, they are bolt-ons. That means they do something: for example, they make it easy to search sensor data. Or if it's a bolt-on to the automation system, it might be some code completion thing to help engineers build a new control logic, a new screen, or something like that.

Then we have to ask ourselves the question: is that the thing we need to take that giant step, that quantum leap toward more productivity? The obvious answer would be to make those bolt-ons a bit bigger, but that's clearly not going to cut it.

Willem van Lammeren

It might be a good idea to take a look where we want to go. For the moment, we are in the assistant stage. Using the analogy of a car, you would see that as your windshield wiper going automatically when it rains, or your GPS. It gives you information, but in the end, it's you, the driver, that has the responsibility to take the action based on the information you're provided. It's super useful, and like David said, we have great applications there, but they're quite isolated.

If we would look really far into the future, what would be the perfect, the most ideal situation for manufacturing from an automation perspective? The autonomous plant: everything runs automated, lights out. You don't need even a single person to press a button.

Maybe you've seen a couple of movies about advanced plants where they do this. It's honestly not that special. It exists for a while. Just think about a solar farm. You don't need somebody walking around the solar farm most of the time. Windmills, for example. Even a chemical process like air separation has already been around for decades. You can have local small plants creating liquid nitrogen near a consumer without having somebody physically operating that.

The problem is, once the complexity starts to increase of the process that you're executing, the business case completely disappears. The cost of automation becomes way too high. We're going to need something in between full automatic self-drive and automatic windshield wipers when it rains.

That's the in-between step we're going to call the virtual operator or the virtual engineer. We're going to look for solutions that, within their limited scope, can already take some actions. They can replace, they can be a counterpart for our engineers. It's not easy. It's going to be quite complex to get there because we have a couple of big problems we need to solve before we get there. It's not just slapping a chatbot on top and hoping things will work.

We've seen two big problems. The first big problem that is going to have to be solved, and we don't have the solution at The IT/OT Insider, is we need to have a more holistic view of all those separate digital twins into one general digital reality.

For the moment this work is done manually. We connect the systems. We create data models. The issue with that way of working is, as we're adding more data, we're going to add much more connections. It's going to become an exponential problem. It's not sustainable. For the moment, we're also barely managing to connect all those data and maintain them. That would be the first problem that AI should need to solve.

The second is that virtual operator. I have lots of experience, and you too, in process industry, in chemical industry in general. It takes an operator around 10 years to really know a plant, sometimes more. We think it's going to be a complex problem to solve with AI. Sure, the physics is well-known, well-described. There's math, no problem. But there's a lot of complexity hidden behind that.

We even know cases of identical plants built on the same place, next to each other, at the same time, exactly the same layout, and still behaving differently. We're going to have to find systems that can learn quickly, because we cannot wait 10 years, how this specific plant is working. Once we get those two connected together, I think that's where we're going to see really interesting applications emerging, AI that can make that step change.

David Ariens

We want to end our presentation by giving you three examples. Three we selected out of a gazillion, because there are so many.

The first one is this one. I have a dog too. If you buy dog food, you want to have the perfect blend between enough nutritious materials and water, because water is really, really cheap to buy if you're a dog food producer. That means the producer will always try to find that balance in the dog food: to dose just enough water to make the food still not too soggy, and the dog and the owner really, really happy.

We got that example from our friends at Tvarit. They are an AI manufacturing company. To get there, what they actually did is they already started solving a bit of these two problems. They combined a couple of the digital twins: the one where you have the sensor data and where you have quality data. They combined that together with a limited model of the physical reality to understand: if we're dosing more or less water, what is actually the physical result of that?

Then they brought that advice to the operator first in what we call open loop, which means the operator can decide what to do with the advice. Secondly, in closed loop, if we trust the results good enough after a while, we can make that closed loop. Closed loop means that the AI will automatically apply the recommendation to the process.

It's a very simple one. It's using classic AI. This is no GenAI agent, whatever. But this is where, today, the real business cases in the industry can be found. Here we are talking about not half a percent improvement. Here we are talking about percentages: five percent, 10 percent, 15 percent easily.

Willem van Lammeren

The second one is very close to us. It's very close to our hearts, because you heard we're not native speakers. We come from Belgium. The right term for fries is not French fries, by the way, it's Belgian fries. You guys know us for waffles. We know ourselves for fries.

For good fries, David, you need a couple of things. You need good potatoes. We have lots of them, so that's easy. You need to bake them properly twice, very important. Third of all, they need to have the right length, just so you have your balance between mayo and fry just right. No ketchup.

The guys at PolySense helped a fry manufacturer with quality control to find out what's the distribution of short fries versus long fries. People in manufacturing might say that's already a solved problem. We have vision systems for a long time, and you would be right: that exists.

The problem they were facing was increasing the accuracy fast enough. That meant they had to generate enough training data really fast. The guys at PolySense were really smart, and they created a virtual fry generator. You could say they simulated the production process. They simulated it on quite a basic level and used that as training data.

Instead of relying on the physical process to train and then do your vision system, they just could deploy factories like you could deploy an application. I found it a really interesting case where they're augmenting their digital twin and then using that physical information and bringing that back to the edge. It's a really fun case.

We brought a small video. This is the actual process. This is the actual video feed. What you'll see in a second is the synthetic data. This is synthetic data. It's actually a game engine. They are simulating a random fry generator, which is actually like a dream come true for me. Zero calories.

By doing this, they were actually able to make very high-accuracy predictions. This is interesting because here again, what we are doing is really stepping into that physical domain.

David Ariens

The final one is a bleeding-edge case where this company, Applied Computing, comes from the energy industry. The problem is that if you have a chatbot, a normal ChatGPT or whatever chatbot, and you're going to ask it questions about a physical plant, it will respond based on the knowledge it got from reading documents.

What they did is they built a digital model, a sensor model, which contains all the different interesting parts of the digital twin, and combined that with a second layer, which is a fully physical model, and then put an LLM on top of that.

What you actually get is, when the operator asks a question to the LLM, it will trickle that question down. It will check it against the physical reality, the digitized physical reality, and all the sensor data it actually has real-time access to. This gives us results, as you can see here, which are actually checked against what is possible.

This is really important because here we are not making guesses. We are taking guesses, and we're checking them against the actual physical model.

Willem van Lammeren

With that, we are through our time. If you go back to our question, can AI be that quantum leap that we're looking for? I think yes, but we need to rethink our approach. We need to go away from slapping a chatbot onto your application or trying to just improve your own application, and try to look for those new fields, those two problems that we mentioned: creating that virtual, real digital twin, and trying to make work of that virtual operator.

That brings us to the help we are looking for. If you are in manufacturing IT, help us define that virtual operator. We really want to set a bar somehow. If you are developing AI applications for manufacturing, also show us where you are. We really want to understand where you are putting your focus. Obviously, if you struggle to align IT/OT initiatives, that's what we do best.

With that, if you want to stay up to date on manufacturing digitalization, go to our blog. We post weekly articles, all free to read, at itotinsider.com. If you want to know more, we also have our academy at itot.academy. With that, thank you very much.