What it takes to build a modular factory – Podcast Transcript
The idea of a modular factory brings to mind something space age and futuristic, with the ability to rapidly adapt and scale production systems through a network of connected shops providing different services yet, to achieve that, there are many real challenges to overcome. From quality to the human factor, creating a modular factory that can meet the demands of modern product design is no easy task.
In this episode, host Conor Peick is joined by Alastair Orchard to discuss what it takes to ensure a modular factory concept is up to the task of creating a high quality product as well as the importance of combining both advanced digitalization concepts with human in the loop elements to achieve an ideal result.
Check out the full episode here or keep reading for a transcript of that conversation.
Conor Peick:
Welcome to the Future Ready Podcast, I’m your host, Conor Peick joined by Alastair Orchard, VP of Digital Enterprise at Siemens. Picking up from where we left off last time, when you have this, I mean, this network of these, you know, modular factories that are all producing this mix of products, How do you make sure that the quality stays at the required level and make sure that every unit is sort of in spec and producing at the desired quality and not mixing colors incorrectly or whatever?
Alastair Orchard:
Good question. So I’d say the first thing is to make sure that obviously that cell is designed correctly. This is why we use all of our digital design techniques to design and optimizing the digital world so that we’re sure that whichever recipe we throw at that cell, it’s going to operate correctly. That recipe then becomes a critical element. The recipe contains not only target set points which guide the movement of the machines, but these are also used by the machines themselves to self-correct and to test the results of the products that go on. I would say that what throws a spanner in the works is actually personalization.
Conor Peick:
Okay.
Alastair Orchard:
The cool thing, again, I think we mentioned this before, but the cool thing about adaptive manufacturing is that every order is a kind of new product introduction. And these adaptive cells don’t really care therefore if a formulation is a personalized one or a standard one. They couldn’t really care less. So that opens the door for personalization. But what you have to be aware of is that extreme personalizations can break the product in some way. And so what we do is we put guidelines or constraints within the product design. We put personalization points within the product that are constrained. So maybe an example, let’s take the Gatorade. Maybe a customer, a standardized Gatorade can be 0 sugar or with sugar. Maybe we want to give some personalization there. So we’ll say, we can allow the consumer to choose between 0 grams and 10 grams of sugar. Any more than that, maybe we change the viscosity and the mixing breaks. Any more than that, and it’s no longer Gatorade, it doesn’t have the philosophy of Gatorade. So we give this range and we allow the customization within the range. Maybe on average it’ll be 5 because that’s what consumers kind of like. And so also from a cost perspective, we’re okay.
Conor Peick:
Yeah, interesting. And just that idea of personalization in these production cells is very, very fascinating. But we talked about a little bit about the software embedded in each of these cells. And of course, one thing that obviously we talk about a lot at Siemens is the digital twin and now AI as well.
Alastair Orchard:
A lot. We talk a lot.
Conor Peick:
A lot. Certainly, yeah. So even in big production facilities that are obviously not modular and adaptive and all this, there’s a lot of integration happening with digital twins, AI, these sorts of technologies. But within this modular concept, what capabilities does the digital twin and AI unlock or provide to help with predicting maybe demand or optimizing the robotics within the production cell and yeah, just helping the the manufacturing cell to evolve in real time in response to the market.
Alastair Orchard:
No, it’s highly relevant. Actually, you’re absolutely right. Digital twins, AI, it’s not specific to adaptive, but I think it’s more acutely felt if you have an adaptive approach because the potential for variability is built into an adaptive machine or an adaptive cell or an adaptive network. And so you really need digital twins and AI to keep things on track. Actually, your question was an interesting one because it isn’t just the machine. Again, typically we think about digital twins or machine within Siemens, but let’s take the PepsiCo use case again. I said we had 16 or 17 adaptive factories organized across Canada, the US and Mexico. why 16 and why did we put them where we did? And this was the result of a simulation exercise using our supply chain suite.
Conor Peick:
Okay, interesting.
Alastair Orchard:
We actually used this also to calculate the economics of the network. So was it worth bothering? Was it better to create this network or was it better to leave this demand on the table and not fulfilled? And the simulations actually calculated all the dynamics and the economics and worked out that it was definitely worth going after that demand with an adaptive network and the digital twin used simulations in order to identify exactly the size and location of these facilities. So that was kind of cool. And then within the factory itself, I think I already mentioned the machines. These are all designed using pretty classic digital twin approach. So we want to make sure that we design agility, quality, reliability, performance into the cell itself. So that’s kind of classic digital twin. But then more in the case of adaptive than in traditional manufacturing, that digital twin becomes a really important asset for online optimization. And so we take it, we either embed it directly into the cell, we put it on the industrial edge device, or so we can do that if the simulations are fast simulations, they can run faster than real time, something like a 1D simulation. Otherwise, if it’s a CFD, so a highly computationally heavy simulation, then we’ll run those up front, usually in parallel on the cloud and train an AI. And then the AI can take instantaneous decisions on the shop floor. Anyway, we put this, whether it’s the digital twin itself or the AI on the edge, we connect it up to real-time data. And then it can run, yeah, It can recalculate set points. It can predict maintenance. It’ll kind of keep the cell going. Well, yeah. So it’s critical.
Conor Peick:
Really cool. You brought up the idea of maintenance and things within the cell. What kind of support staff is required for one of these things? Is it ideal that you can just sort of drop it and leave it pretty much alone or maybe a small team?
Alastair Orchard:
Yeah. It’s another good question because what’s the point of having a fully autonomous cell if it then takes 100 people to kind of look after it? So ideally it’s extremely autonomous. But some things are still kind of, some things are still built for people today. So I think in the near term, there’ll still be operators. I mean, And this is where AI kind of comes in. And I think the idea of operators supported by co-pilots is something that everyone understands now. Several months ago, this was not the case. People are still kind of struggling to understand this. Of course, everyone knows what ChatGPT is, but at the time I remember going to try and explain this to a very large customer and It was quite informative. If you take, you can try this now, actually it works very well. It gives you a good idea. If you take ChatGPT or Gemini or Claude or whatever, give it a PDF of like your toaster oven or something and say you are a copilot, it’ll instantly support you in troubleshooting the toaster oven, it’ll give you kind of recipes, it’ll help you be a better operator. So that’s exactly how a copilot works. But the things obviously evolve from now. So you go from a single operator running a single machine to a single operator orchestrating across multiple copilot equipped machines. Then the really interesting kind of paradigm flip occurs when the human becomes the copilot and the machines themselves become the autonomous agents. And then you put those agents, you embody them, you put them in a humanoid robot and now you really don’t need the human anymore to pull that lever or whatever. Yeah. So that’s the, I would say that’s the That’s what will happen, but I don’t think it’ll happen overnight. So agents that kind of prove themselves reliable in answering a certain question or taking a certain decision, they’ll be allowed to be autonomous in that context. And I think for a while we have this kind of heterogeneous landscape of agents in some areas being highly autonomous, in other areas just supporting humans. But we design these specifics. I would say that’s more of a discussion for a brownfield classic factory, but we design these autonomous cells to be pretty autonomous. Sure.
Conor Peick:
Yeah, which makes sense because just like you said, I mean, what’s the point of doing it and then you.
Alastair Orchard:
And then it’s in a container, so you’re going to shut the operator inside a container. It’s pretty nasty.
Conor Peick:
Not a very nice working environment, that’s for sure. Yeah. No, it’s really interesting what you said about as these things evolve and eventually, especially in the case of this highly autonomous modular factory of the human kind of becoming the copilot. And so I can imagine perhaps in that scenario it would be a human somewhere in a remote location. can be monitoring the operation of one of these cells. And then if something maybe goes wrong or he sees an update is required, maybe you can then interact with the agents operating the cell and say, hey, we noticed this problem. Can you address it essentially?
Alastair Orchard:
Absolutely. So that’s maybe a step beyond what I was thinking. It’ll be an operator not only operating multiple pieces of equipment within a factory, but multiple factories. Absolutely. Yeah. Maybe with some Boston Dynamics spot the dogs running around with cameras on. Yeah.
Conor Peick:
Very cool. So in the scenario of a big disruption, perhaps within a supply chain, you know, your supply suddenly has to shift. Given that’s assuming a pretty highly intense, maybe time critical sort of environment that you’re trying to make decisions in. How would you maybe use an AI agent to help you either let the agent make a decision or inform you to make a correct decision as quickly as possible.
Alastair Orchard:
Well, again, I think let’s use the, as we built it, let’s use the Pepsi example. So if Actually, I presented this on stage and maybe we can link that stage presentation somewhere in the notes. But yeah, I’ll describe what I presented, what we built downstairs. So when the, I’ll take the example of the network. So when the football match occurs, this is essentially a demand signal and it’s picked up by a Pepsi supply chain agent. which maps the color of the jerseys to the color of the Gatorade and then works out kind of what’s going on. There’s an agent to agent discussion that then occurs between the supply chain agent and the agent managing the warehousing and stocks. So that second agent says, look, our standard demand predicts we’re going to have this much on that date. And the first agent says, okay, then we have a shortfall. And it then speaks with the agent from our distributed manufacturing platform. and says, look, can you fulfill this demand adaptively? And this is where the discussion around data is really key. So we don’t want the agent just to make stuff up. So in this specific case, the agent, we use Rapid Miner, we use a knowledge graph to a kind of ontology to allow the agent to understand what is a factory, what are capabilities, what is the recipe, and we basically give that agent the ability to map those instructions and materials in the recipe to the capabilities, to the location, and to the inventory levels of all the factories in the network, and to then map, intelligently map this request onto the network. It’s obviously going to try and get as much done as possible in the production node as close proximity is really key here. We’re trying to reduce lead times and transportation and so on, but it may not be able to fulfill all demand there. So it’ll then distribute across the other areas. In our specific platform, we also have agents running for each production node. These are producing agents. And so there’s a negotiation between the the network agent and each producer agent and a smart contract is put in place. Platforms work when they’re frictionless and that means you can’t have humans in the loop negotiating things. So the agents negotiate, they encode their decision on the blockchain and a smart contract and then production is scheduled and immediately kind of triggered. So I mean, in most cases, you can just think of these agents either as just automating and speeding up the process that humans typically do. In other cases, as we’ve identified our platform, we’ve actually taken code, either that we actually coded or that was captured in low code workflows in Mendix very often. and we’ve removed that and replaced it with an agent. And at that point, it’s obviously much less code. The code base is smaller, but it’s actually much better. With a single prompt, you can replace lots and lots of hard-coded if statements and so on and so forth. But that’s essentially how it works. The whole network is managed by this interaction between agents.
Conor Peick:
Wow, I mean, that’s really, yeah, that’s a really interesting The idea of a contract between agents is really interesting to me. And presumably there’s also some amount of traceability within that interaction where if you had to, you could potentially go back through and maybe sort of run an audit of the contract that was made and then what got executed and make sure that it’s follow through was correct.
Alastair Orchard:
No, that’s exactly right. It’s what I mentioned the blockchain. I know the blockchain is kind of gotten a bad rap. More recently it was this hyped technology and then everyone’s like, but it is immutable and it’s a great place to put smart contracts. And so you can always go back and you can’t change them. And those smart contracts are code so you can then really execute the smart contract, which is nice. We also stream data out of the whole supply chain, including production process, but also logistics and supply movements back into the blockchain to create a genealogy, an end-to-end traceability of everything that happens. And then we can compare what we contracted to do with what actually happened, which materials we’re supposed to use with ones that were actually used. and we use another AI agent to verify and validate the production. And again, this isn’t normally required if, that gateway was made in my factory, but now it’s maybe made in someone else’s factory. And I need to make sure that, you know, that agent is something that I’m confident that it’ll pick up any problems. But good point.
Conor Peick:
That was interesting, but I think this is a good place to end our episode today. Once again, I’ve been your host Conor Peick joined by Alastair Orchard on the Future Ready podcast. Tune in again next time as we go beyond the factory to look at the future of distributed manufacturing.
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.
