Managing the adoption of Industrial AI – Podcast Transcript
AI has the potential to address many challenges in the industrial world, however, making that dream a reality involves more than just developing a new AI tool or addon. With an Industrial AI model in hand, finding the best way to bring that model to both tools and physical assets represents a challenge of its own.
In this episode, host Conor Peick is joined by guests Matthias Loskyll and Samir Desai to examine what it takes to bring AI to both software tools and factory systems. Industrial AI is not a one-size-fits-all tool but, rather, a collection of methods and approaches that Samir and Matthias examine from both a design and manufacturing perspective.
Listen to the full podcast here or keeping reading for a transcript of that conversation.
Conor Peick:
Welcome back to the Future-Ready Podcast. I’m your host, Conor Peick, and joining me again today are two of our AI experts from Siemens, Samir Desai and Matthias Loskyll. Last time we left off talking about Siemens’ broader AI approach and when it comes to AI, the importance of having the right tool for the right job rather than a blanket approach. Samir, could we pick things up from there?
Samir Desai:
Yeah. And you brought up a very good point, Conor, there, the right tool.
Conor Peick:
Yeah.
Samir Desai:
And that is something that we call, as part of our three-pronged strategy, one of the prongs is engine, is what we call engineering AI or in-app AI, which is the users are using the software products for years. Each one of these products is designed to do something special. And as the users are performing their task in that particular tool, in the particular context of everything that is around it, including the data, that is where we interject AI and say, “We understand what you are working on.” Do you have a 3D geometry? Or you have a simulation or you have a change? Or you have a manufacturing process plan that you’re working on? Our tools understand that and then we bring in AI to assist the user in doing the next step better, in predicting the next step that they’re to perform and predicting how the geometry is going to evolve, how the change needs to happen.
So it’s really bringing the AI to the user rather than telling the user, you go do over here and go figure out how you’re going to do AI. So that’s a big thing that we do.
Conor Peick:
So then thinking about, I think that’s an excellent look into how AI is influencing the engineering process, from moving from design through to… As we’re approaching more automated production as well, AI managing the factory floor perhaps a bit more. So Matthias, how do you think or how do you see perhaps as a better question, how do you see companies actually deploying AI on factory floors? What sorts of systems are they putting in place to automate their production a bit more?
Matthias Loskyll:
Yeah. Actually systems based on all the different areas that we touched on. So a lot of predictive AI, but also generative and agentic AI approaches. The first one’s even agentic workflows now. And very different areas of applications. So very famous and popular one in industries is obviously visual quality inspection. So using machine vision applications in industries to solve the problem of inspecting products or intermediate products if there are defects, simply speaking. So usually you have a camera and some products running over a conveyor belt, for example, and you want to automatically inspect those. That’s still a manual task in many applications today. So you have human operators standing there, all day long, inspecting the products on the conveyor belt and picking the ones that are not okay, so to say. And that’s what we can clearly automate, to a large extent today, with AI technologies.
And we do that with different products actually that we have. We have Inspector, which is an out-of-the-box quality inspection system for detection of defects, surface defects, for example, or detection of foreign objects or missing objects. So the more simple use cases works fine for metals, electronics, plastics, for example. And there are extremely complex quality inspection use cases in consumer packaged goods or food and beverage, for example.
We have thousands of pieces running over the conveyor extremely fast and you still want to inspect those. And there we have another product which is called Visual Inspection Cockpit. So this is for, we call it closed loop AI, so it inspects the product, it detects a defect, and it sends a signal to our PLCs. So the control of the machine or the plant and then this one gives a trigger to recheck products automatically. And all of this in real-time, so to say. And this is really cool, this is highly embedded into the production line.
So quality inspection, besides that predictive maintenance, obviously. So forecasting, if there is a defect of a machine, if a breakdown could happen soon. Actually avoiding that breakdown, sending a maintenance engineer in time. That’s predictive maintenance where we have Senseye is our product. And if you look at process industries like chemical production, for example, we have advanced process control. So this really goes down into the parameter optimization of your control program, tuning the parameters on the fly based on the values of the process. So this is also deeply embedded into the production process itself.
We call it, actually, physical AI if we have this close integration into the industrial environment. If we actually act also we do not only perceive and reason, but we also take action automatically or autonomously. This is what we call physical AI, maybe to add another category to predictive, generative, agentic.
There is also physical AI not only connected to autonomous robots and those aspects, but also tightly embedded into control system of a production line. And actually I would say all of this is industrial AI, how we understand it. It doesn’t matter which flavor of AI it is, all of this is under the bracket of industrial.
Conor Peick:
Awesome. And one thing that obviously we love to talk about technology and all the exciting things that we might be able to do with it. I think one of the practical realities maybe for customers is the thought of adopting a new tool. Maybe they already have other tools in place. And what they don’t want to do presumably is totally replace their entire ecosystem. So how can we work with what customers already have in place and build on top of that?
Samir Desai:
It’s a wonderful question because we are already doing this with our customers. I can do a two or three different variety of examples, but we’ll give you the value of how customers are thinking about this and how we are partnering with our customers for this. So if you take an example of our PLM system, Teamcenter, that’s the number one PLM system in the world today. We have a customer who’s been using it for over 10 years. And then we added certain elements of AI in there. And they said, “We would like to try this thing out.”
So what did we do? We provided this capability to them in the Teamcenter product, but we partnered, we ran a set of proof of concepts with them, to say, “Mr. Customer, what are your top three use cases? Let us run through those with you. With you, not hand it over to you, but with you and show to you how this is going to add value and bring more importance to your end users.” And they loved it. And at the end of it, they bought 2,500 seats of that software offering from us.
So that’s one. The other example is look at completely different industry, chip design. ARM, which is a chip design company, they have to work under Six Sigma quality, which means that when they design chips and when they go to manufacturing to manufacturers, only one in a billion chips can fail. One in a billion, Six Sigma. To achieve that, they have to run a lot of simulations. And they were running about five billion simulations over a period of months.
What we did is with one of our IC chip design verification software, Solido, we introduced a generative or a predictive model in there and they started using this predictive model. And instead of having to do five billion simulations, they’re down to 2,000.
Conor Peick:
Oh my goodness.
Samir Desai:
That’s a magnitude of 2,500 X, two point actually more than that. So it’s 2.5 million X improvement that they’ve gotten out of it. And instead of months, they are able to do that in hours. These are some very real examples of what we have done with our customers. The last one with PepsiCo. Industrial metaverse, we partnered with PepsiCo as well as Nvidia, where we are using Nvidia GPUs to do what we call a digital twin composer. And we built this together with PepsiCo. So what do you want to do? What is your scenario? What’s your factory layout? How are we going to do this visualization? And we’re with you for that. At the end of the project, they have it deployed. They’re down here today. And they have achieved up to 15% reduction in their capital expenditure. So real AI offerings with partners, keeping it real, avoiding the hype, and bringing it to them. So that’s how we work and take AI in with our customers and with our partners.
Conor Peick:
Work with the systems they already have in place.
Samir Desai:
Yes. Yeah.
Conor Peick:
In a production environment as well, you have, I’m sure, a large mix of brownfield facilities where customers have machines of all ages, I imagine. Yeah, similar question, I suppose. How do you navigate maybe integrating AI into these facilities where you might have machines that are, I don’t know, maybe older than me in some cases.
Matthias Loskyll:
Actually in a very similar manner, but with different technologies obviously. So it’s all about data connectivity in the first place. So you really need to get the data out of these systems. And as you said, it’s usually a legacy system that is their brownfield applications with mixed vendors with different generations of controls, for example, or of tribes. And we have to get the connectivity right in the first place. So that’s the homework that every customer has to do then. And we also help the customers on this journey and with some software components that we can offer to make connectivity easy, to get the data out of the systems, and then send this data to usually a centralized environment like based on our industrial edge platform, we can contextualize this data, even sending it to our Altair Craft Studio, we will expand on this later, I hope, so talking about the data fabric.
That’s really key to bringing all these data items together, getting it to a certain semantic level, contextualizing the data, and then you can put the AI applications on top. So this is also a journey for customers to transform their operations into an AI ready system to really leverage these applications. And we also recommend to go for the high value use cases first, to start with some pilots to get the breakthrough with those. But from the beginning, standardizing the infrastructure. So not reinventing the connectivity layer with every AI use case, not reinventing the AI runtime, so the execution, the inferencing of the AI models. With every use case, we saw that very often actually happening. So we highly recommend to standardize all this on a joint software-based infrastructure and then you can go and roll out these applications and put further AI use cases on top of it.
Conor Peick:
You brought up the idea of the data fabric and I think… Let’s jump into that right away because I’m very interested in the idea of how do you connect all these different, maybe, silos of data that exist within those companies? And both how it seems like to leverage AI, you need to do that connection, but perhaps AI can also help in creating that data fabric and building those connections as well. Samir, I don’t know if you want to start.
Samir Desai:
Yeah, I can start. And Matthias, please, please add on top of that, because we both are focusing on that for the key elements of engineering as well as automation. And so what we typically have seen over, again, two to three decades that our customers have been using, and this is a problem of what the customers call as either data silos or dark data. What do they mean by that?
So let’s take an example that one of our biggest industry is car manufacturers. They’re doing their design, they have their design information in our PLM system Teamcenter, but when they look at their vendor and sourcing information that’s sitting in ERP, they have their customer relationship data in CRM. And then they have a lot of homegrown systems implementations as well. Now, when they’re building this car, it’s not just the design or the sourcing or the in-house data that is by itself important. The connection of this data is more important because when they think, “Hey, I’m going to make this change in my design data,” it is going to have an impact on sourcing. It is going to have an impact on manufacturing, which is where Matthias and the team comes in terms of automation, is going to have impact on a multitude of other things.
And how do you understand the impact? Traditionally, you would have to get people from these different disciplines, the subject matter experts, they’re going to come together, they’re going to look into their systems, they’re going to try to figure out, and it’s a long cycle. Takes days, weeks, months sometimes. Wouldn’t it be cool to really provide an automation of this, that it understands the relationship of all this data, and based on that understanding, you can run AI on top of it, but you can run analytics on top of it and you can perform actions on top of it.
This is where making it smarter becomes critical. This is what customers mean by dark data. I know I have all this information, I know it’s somehow connected, but I don’t know what the connection is. So we are solving that through what we call AI fabric. And the primary fundamental offering there for us is based on our Rapidminer portfolio.
Rapidminer, Graph Studio specifically, is what we are using. It’s a great piece of technology that we have part of our Altair acquisition that we did just over a year ago. And what we are doing there is this concept, I’m going to get a little bit technical here, but is the concept of ontologies or semantic relationship of information that is in these different places.
So we are bringing Siemens because of its domain expertise that we have in all these domain areas, because of how we understand the data works and it connects, because of how we know data is secure and security is more important to our customers than anything else today. We bring all of these pieces together into this ontological view that says, “Here’s how your design information is connected to your supplier information to your manufacturing information.” And oh, now you can start asking questions. Not only ask questions, but you can build agents off of that ontological representation using either our Mendix local tool or the Rapidminer tool.
You can run analytics on top of it. And once you figured out what the connections are, now you can say, “Hey, this piece of information in one system is changing. I need to affect a change in another system.” You can orchestrate that because of the way of how we’ve established this whole ontology. It’s a very, very powerful offering. And that’s why when we look, it’s not just engineering. It’s in the manufacturing world, it’s in the shop floor world too. And Matthias, if you want to talk about how you’re looking at this thing in your domain.
Matthias Loskyll:
Yeah. So I have a very similar view. So this proper data foundation is really key to scale AI in industries. And also from an operations or running the factory perspective, it’s the same thing. You still have to integrate the data across different data silos. They’re still there very often. And it’s also, you can also give a feedback loop to the design and the planning phase of the production actually. So just imagine the example that I mentioned before, the quality inspection.
So you detect a failure or a problem of the product and now this occurs over and over again. This usually means there’s something wrong in the process of the manufacturing plant before this quality inspection step. So something is deteriorating or some process parameter is drifting. So usually this means you have to optimize something. And you have to give feedback to the design and the dissemination and the planning part of the whole production line.
So closing this loop across the whole value chain, this is where really data fabric is needed and I think it’s a great tooling that we have in the meantime here. So I was always dreaming about this doing my PhD in the field of knowledge graphs actually. And so this is the underlying technology that we use here, picking up data silos, having a joint ontology, how we call that. So it’s a joint terminology to have the different items on a semantic level. That’s the right step to really preg up data silos and this is needed then for agentic workflows that operate across the different phases of industrial production.
Conor Peick:
And so maybe part of building that foundation, it seems to me that the data fabric in some ways is providing a really solid base level for AI to operate from and for you to build more valuable applications maybe on top of. But we’re going to have to save the answer to that question for next time. Once again, I’ve been your host, Conor Peick, joined by Samir Desai and Matthias Loskyll on the Future-Ready Podcast.
Siemens Digital Industries Software helps organizations of all sizes digitally transform using software, hardware and services from the Siemens Xcelerator business platform. Siemens’ software and the comprehensive digital twin enable companies to optimize their design, engineering and manufacturing processes to turn today’s ideas into the sustainable products of the future. From chips to entire systems, from product to process, across all industries. Siemens Digital Industries Software – Accelerating transformation.
