Why is trust so critical in the integration of AI into existing chemical engineering technology? In this episode, we explore the role of AI and machine learning in chemical engineering with Dr. Sebastian Werner. Dr. Werner has a doctorate degree in chemical engineering and currently works as an AI Evangelist at Dataiku, a data science and machine learning platform that designs, deploys, and manages AI applications.
Dr. Werner explains how AI-enabled predictive maintenance, regression, and time series forecasting improves efficiency of chemical engineering operations. He and Deep touch on the challenges of bridging the divide between operational and informational technology, and why this has slowed the integration of AI into manufacturing. Finally, Dr. Werner speculates about the future role of AI in manufacturing, as well as how the global push for sustainability will influence the development of new technology.
Learn more about Dr. Werner here: https://www.linkedin.com/in/sebastianwerner/?originalSubdomain=de
Deep: Hi there. I'm Deep Dhillon. Welcome to your AI injection, the podcast where we discuss state of the art techniques and artificial intelligence with a focus on how these capabilities are used to transform organizations, making them more efficient, impactful, and successful.
Welcome back to Your AI Injection. This week, we'll be discussing the use of AI and chemical engineering and industrial processes with Dr. Sebastian Werner. Sebastian has a doctorate degree in chemical engineering and currently works as an AI evangelist at Dataiku, a data science and machine learning platform that designs deploys and manages AI applications.
Thanks again for coming on. Maybe. Get us started about how did you get into chemical engineering and how did you get into AI and machine learning? Let's start there.
Dr. Werner: That, that's a super interesting question. So I started, um, thinking about chemical engineering, cuz I couldn't really decide what to do after high school. And I was interested in like, let's call it things that are between the world. Like not really, I couldn't decide on mechanical engineering, not really on chemistry. I was always interested in like modeling and that's also like modeling real processes for, let's say some benefit. That was something that like really got me. And like one part of there is sustainability because like efficiency is one thing that really drives chemical engineers. So that I would say is the motivator brought me into chemical engineering.
Deep: How does a high school kid have a vision of that?
Dr. Werner: So that goes back to like, let's say to the old times to nineties. So I saw this. So back then a guy in Switzerland called GRE cell came up with this GRE cell cell, like the, the, like photoable Thai cell to like artificial photosynthesis and all that stuff. And it was like super fascinated by that. Okay. Wouldn't it be cool to like, um, build in the end, this idea of let's say painting your Walgreen, spraying some water on it and like down drip, some let's say sugar, or like in the best case alcohol, wouldn't it be cool to do that instead of like, I don't know, like making in big factories, that was the idea. And that was kind of like bringing me,
Deep: You got rooms, you got paint, you wanna drip sugar down, like just cuz the effect would look cool. Like what's the sugar's role. Are you trying to, so let's say photosynthesis, we have CO2 in the air.
Dr. Werner: Yeah. And like we can use sunlight photosynthesis in the first case. Just imagine we do artificial photosynthesis. So we manage like put photosynthesis as kind of, let's say just a paint on the wall. And then just add water to it, similar to a tree and use that in order to produce sugars for us. And then add a second step in that process, fermentation to take that sugar right away and make alcohol.
Deep: So there we go. Now we get to the high school kid
So somehow that turned into like sort of desire to kind of like, maybe understand how to model these more physical. Chemical reaction processes. And that took you into that world. Is that yeah, like, yeah, absolutely. And to, to save more precisely in high school chemistry, we went to a trade fair called AMA in Frankfurt, and I saw these giant vessels, these giant, uh, factories, cuz we have no scale about it. Like how big such a factory is. And I was like, cool. Then, like, there must be a way to learn how to do that. And I like chemistry, I liked physics. And so I liked engineering. And so I was like, Hey,
Deep: I mean, that makes a lot of sense. I always felt terrible. Um, when my kids were younger, like take your daughter to Workday or something. And you know, like the most exciting thing for a seven year old is like a bowl of liquorish or something. Just a bunch of computers in a loft somewhere with a bunch of geeks staring at them. And I always felt jealous cuz when I grew up, my dad had this huge research lab with all kinds of contraptions and I there's just something about seeing these larger systems. That's way more interesting than seeing them in a model in your, in your computer or something.
Dr. Werner: and I mean like as a kid you ask. Yourself, how do these stripes get into my toothpaste? And then you ask the second question. Yeah, there you go. How on earth do you manufacture toothpaste? Like how it's made and the thing is all behind. That is always chemical engineering or process engineering. Yeah. Cause it's about scale, raw material. Into something you can use as like human at scale, because like these factories are ginormous and you can't sell toothpaste for 50 bucks a tube. So you have to figure out how, yeah.
Deep: So from chemical engineering, somehow you wound up in machine learning and AI, what was that? Uh, what, what led you there?
Dr. Werner: Yeah. And university times, um, like when I started chemical engineer, one important part is to understand the chemistry in physics, behind the processes. And I got somewhere attracted by the thing that, uh, calls of most people, meaning thermodynamics mm-hmm, like understanding for like the, like, if you have no idea, what's that about? It's about like, Hey, what happens if you mix mix certain things, let's say take the practical example of water and Ethan. You mix that together and you heat it up. So distillation, that's a very practical thing, but like the theory behind it really fascinated me. And that got me into the point of, let's say modeling. So called flow sheet modeling. When you try to end to end model a chemical process and that's in the end, you could say how I ended up at machine learning, cuz there's a lot of these same techniques behind it that we nowadays use for machine learning. And later on for. Yeah.
Deep: So in the traditional chemical engineering, pre AI machine learning world, walk us through what things were like then, and then maybe what is in, you know, an undergrad learning Now that's kind of incorporating these statistical processes that help modern model some of the uncertainty that. You know that maybe the physics and chemical modeling doesn't get too directly.
Dr. Werner: That's an excellent question. It's to a certain extent, it's a little bit like chicken and egg. When you want to model whatever process you have, you can either start with a data. Let's say that's for machinery. AI is in, you have observations or vice versa. You can start off with a model first and try to adapt that to a certain. It's likely chicken and egg problems. Think about physics. There's a guy, uh, like, uh, Mr. Newton, he sits, uh, under a tree. He sits in, uh, sees an apple falling. He does an observation. He says, huh, there must be a relationship behind it then coming. With a fundamental relationship. The other way you could say, Hey, Mr. Higgs came up with the observation, Hey, Higgs bow on that must be there. So we built a super complex, uh, LHC Collider and made the observations afterwards in, in order to validate a model and fit the data to it.
Deep: Yeah. Makes makes a ton of sense of super high level view in this. Yeah. Yeah, no. I mean like, you know, like modern Newton sitting under tree now you've got a video camera. Maybe you do some semantic segmentation, you track the apple as it falls, you measure the acceleration. Maybe you weigh it. Maybe you measure the force, you put it all together. And then you wind up with that, you know, like your actual point.
Dr. Werner: And that's the super rigorous points that you said, like all these details. It was like the first pot, like chicken and egg. When you start. So, and so. Going going into the modeling of a process. The same is true. Like when you model, let's say a chemical process, you can start up building it up from the fundamentals, these little bits and pieces, starting from like Newton's gravity law to all these other pieces, you can put them on top of each other and build so-called first principle model. That's what you learn in university flow sheet modeling. You learn all these relationships, physical chemical ones, like continuity laws and whatsoever. That's the first thing you can do. And you built that for, let's say a factory. The one thing you find out. Hm, it doesn't really fit. So what you need to do is you need to then take observations from that factory, if it already exists and you need to fit the data to the model. So you adjust some parameters and that's already, you could argue machine learning, cuz you do statistical fitting of the yeah.
Deep: Walk us through like what are, where are some of the points in. Chemical plant where reality butts up against the equations that you've got that make everything. Cause you know, those of us who've studied this sort of thing. You understand that the equations kind of have a whole bunch of parameters and there's a lot of approximations built in, but in reality, like what are we talking about? Like, you know, pressure, variations, temperature, variations that are just sort of shifting throughout the day that maybe don't get reflected.
Dr. Werner: It's almost at all levels. So certain things we understand very, very. Let's say thermal conductivity. We understand fairly well. If we have pure substances on both sides, meaning there's pure water on one side and there's like pure water on the other side. And there's just, let's say two inches of metal in between you say, and one hand it's like warm water on one side and cold water on the other side. And you say, okay, explain to me and build a model for, let's say how. Energy flow from one side to the other. And what temperature will I have on the other side after two hours? That's let's say fairly well. Understood. Yeah. That's complicated. As soon as let's say effects that are, let's say slightly chaotic come into play.
Deep: Let's say boiling Kool-Aid mixed with granola.
Dr. Werner: Maybe like is you have like first thing is, let's say we have boiling water. Yeah. Are you really sure that you understand what happens when the bubbles form on the surface? That's not completely analytically understood. Mm-hmm in the same. If you now, instead of just pure water, you take a mix of, let's say water with a little bit of, I don't know, salt or yeah, a little bit of ethanol. It already gets more complicated. So that's the starting point for where it's get, uh, gets a little bit more fuzzy. So now let's go to the factory. Assume you have a heat exchange up at the lab scale mm-hmm and you want to cool down a mixture. Of let's say whatever, whatever material you're dealing with, the heat exchanger will not be the same when you buy it. Then three months later or two years later, because there will be something called fouling mean there will be some material built up and for sure the thermal conductivity will change over time. Do you really understand what's going on? Not so sure. And here is exactly where these, these points come into play at all these levels. Any like certain things come in that we don't understand at.
Deep: Makes sense. So walk us through, like maybe we should take a scenario, pick something that you maybe, that you either worked on or that you can think of. That's an interesting chemical transformation or process that you're trying to engineer and walk us through. Like where is AI being applied in a particular scenario?
Dr. Werner: Let, let let's, before we start going into AI, I think staying with that, like that example with the process is a good one because okay. Like E especially that heat exchange, we just talked. Now just imagine you have these observations from the seed exchanger, where you have on one side, um, a material you want to cool down and you have cooling water coming from reservoir. What you have is the temperature of the, to be cool on both sides of both sides. And so on, in a simple way, you could just say, Hey, I have these observations on all sides and I know how much cooling water I'm putting.
Deep: And you probably have some inputs, like you've got the thermostat that you can turn up or down. Yep. Maybe you have pressure. I don't know.
Dr. Werner: Um, so, and now let's just say, you want to make sure with the CD exchange, the whatever fluid you want to cool down has a certain temperature on the outlet and you see, Hey, over time I have to open my valve for the cooling material more and more and more and more to reach this tempera. So you don't have a model about this, uh, about the fouling that is happening there. However, what you can do at scale with artificial intelligence is for every single one of your heat exchanges, take exactly these data pairs and let AI for you.
Deep: Figure out that's a predictive maintain how much to turn the knob one way or the other.
Dr. Werner: Yeah. When you need, not only that, how much to turn up, cuz this control software for that. But when to call a crew, a cleaning crew. To clean the seat exchange and get, uh, all the following stuff out. Cause if you're in a situation where like you cannot open your valve anymore, then you have a problem in your process.
Deep: I missed something. So you've got this, these big vats of the thing that you're trying to cool. And then the thing that's doing the cooling and what's the fouling that's happening. Like some something goes inside the,
Dr. Werner: oh, oh, sorry. I have to go back a step. So the cooling water is never clean. It always has some stuff floating in it. And whenever there is something floating in it, bacteria. Like, whatever scale, like whatever scale, like whatever you can come up with, it will happen to build up in the heat exchanger, heat exchanger. Yeah. It'll start to grow. If it's bacteria, for example, it starts to grow and so on. You can try stopping it. And also there's some rust that might build up and so on, no matter what this effect fouling is in there, uh, you will not get rid of it. The thing is taking an heat exchanger out is costly. So you don't wanna do it too. , but if it doesn't do its job anymore, it can be a problem for the process that you're operating it in with. And so predictive maintenance of something like this can save a ton of energy and to do this analytically with a classical model. Is not always easy. Cuz there are so many unknowns. You don't know how fast will the stuff grow on there? How fast will it rust all of these things. You don't know. However, any of the observations that's exactly where I can help you to figure out, Hey, by the way at about this time. Likely will reach a critical level. You probably should, uh, schedule. Okay. And that gives you a chance to get your resources lined up for that day. And with that example, we pretty much explained the entire use case of predictive maintenance, which is one of the key use cases of artificial intelligence, not just in chemicals, but also in entire manufacturing. Just observe what you see. You have AI figure out a certain trend by means of a regression model specify a critical limit. And use that in order to alert you when a problem comes up. Yeah. And the key is that it's something that's not straightforwardly modelable, um, or at least with a high level of certainty. Um, like in the case of the backtrack, who knows what's floating around in the air at that particular plant or, um, you know, the state of the metals and, and just imagine you directly use water from, from the river, right. There is a lot of stuff floating in that, and that's in many cases, cause it's costly, cooling water. You take water from the river and exactly this is what happens. There's all kinds of stuff in there.
Deep: You're listening to your AI injection, brought to you by xyonix.com. That's xyonix.com. Check out our website for more continent, or if you need help injecting AI into your organization.
So, what did folks do before, you know, pre pre machine learning? Were they using more basic statistical methods to do this? Or.
Dr. Werner: I would say yes, but in many cases, it's also, do you really have the time to have a, an expert engineer sit down to build 500 models for your 500 TD exchanges in your plant? So what you would do is you probably have a schedule or you just simply have an alarm in this case that just says, Hey, if the output temperature here becomes critical, this is when we just, uh, when we just do. And of course, this is the most simplest case where you could do AI at scale in the chemical industry or process industry, but it's a starting point to like, What is the gap that we can overcome? Whenever we have phenomena, we don't really understand that. Well, use the observations of the systems that we have in order to bring our understanding of the system to the next level.
Deep: So let taking this, um, example that you've got here. I mean, I'm just. Guessing that you must have some modeling software that lets you run through some scenarios like this, where you can play between the physical world and the machine learning world. So in the physical world, you can set up your, you know, your exchange, your scenario, you can set up your materials on both sides. Maybe you can even like introduce some bacteria or some, some stuff. And then. Something like mess around with it. And then you can try out your observational models. Like how do you assess efficacy on something like this that has such a long potential?
Dr. Werner: So that, that's an excellent question. The first point I would say or argue is that I believe pretty much everything we see in the process engineering space you could solve by traditional first principle modeling methods in your software that is available out there. They have been around for a long time, and here's the thing given enough time and manpower. Meaning resources of experts to model that you can do it. However, there's always the question of like, how much do you have to invest versus like how much is what's the payback for this standpoint?
Deep: Yeah. I mean, I guess I was getting more at like the question. Not so much like replace potentially the machine learning solution with the physical solution, but how do you test the machine learning solution when you've got this big giant vat and you've got a theory of a model, um, you put it into place, like, do you wait months and months to see if you actually called predictive maintenance? Like you predicted your maintenance at the right time, or do you take it all and maybe sort of try or maybe do both. Maybe you trial it out against different physical scenario.
Dr. Werner: All of the, all of the above is done. So I've seen people doing just, um, like running it and seeing if it, it works. I've seen people benchmarking a machine learning model against the first principle model and vice versa. I've been, uh, I've seen first principle models being benchmark against reality and the same reality. Concurrently also being trained, uh, tested against a machine learning model. All of that is possible. And the thing that you, you mentioned that is actually the, the agist of that is the problem is the process we are dealing with are slow. When there are slow processes, it's in some cases very hard to also see the impact of certain decisions or certain things that are happening because you have so many influencing factors.
Deep: Now I'm kind of curious, like when you build such a model for a particular process, like what are the maybe open available resources look like for, um, chemical engineering, AI people.
Dr. Werner: so super interesting question. I would, uh, I would for one second, take a step back and to say, Hey, what, what challenges are we dealing with in most cases? Yeah, that's probably the most, most challenges we are dealing with are regression problems. Problems where we want to forecast a certain property or dealing with serious of numbers. And most importantly, we are dealing with time serious, meaning effects like, or measurements or observations based over time where the context of what previously happened really matters. Mm-hmm . And now, like going back to the AI and machine learning world, when you look at literature, I don't have the exact number, but like my impression is the vast majority. Focusing on classification.
Deep: Yeah, yeah, yeah, for sure.
Dr. Werner: On the other, very few of them are actually regression, uh, regression techniques and out of the regression techniques, even fewer of those focus on time, serious problems. There's a couple of things that happening there, but it's like, I would say not as far as like some of the other fields have gone in the past couple of years.
Deep: Yeah. I mean, just for quickly, I just wanna clarify for everyone. So like, when we talk about a regression problem, we're talking about taking, you know, a machine and trying to predict a particular value. So that might be for example, like, you know, Zillow or Redfin predicting the actual value that your house is worth as opposed to classification where we're just trying to predict a category like. Things spam or not spam. So, um, yeah. Sorry, go ahead.
Dr. Werner: And of course you, and, and, and what's definitely true is you can take almost any regression problem, andt it. Into a classification problem.
Deep: Yes. And it's definitely true.
Dr. Werner: However, as soon as you add this like, time context, meaning. Whatever you're observing is very much dependent on what you previously observed. In that case, it becomes very, very difficult to do that transfer. And that's one of these challenges and we are dealing with real systems. Meaning when you observe a certain value, it's not necessarily clear from the observations, how you guts to the state that the system is in that produce these certain value. Whereas like for forecasting, like looking in the future, this part really, really matters. And that's where. The challenges with machine learning and AI in that space come into play to your knowledge.
Deep: Is there any type transfer learning, like, you know, like we think of Alex net for imagery and for NLP, you know, we've got all of the embeddings, um, that are trained on like really common, large corporate. Is there some kind of analogy in your kind of process world that. I've seen that would be appropriate.
Dr. Werner: I've seen a couple of papers on transfer, learning on time series. That's the first part I've seen a couple of people that have worked with time series and used, let's say established techniques from statistical learning, like Alima and so on. Mm-hmm and enhanced them with effects from neural networks. What all of these approaches are missing so far is being able to let's say incorporate like detailed feature engineering and so on. Cuz like there's been a lot of like, um, like development going on, but it's like at a very more early or like infancy state compared to what we see in other domains. Cause the complication is, it's not just that we have, let's say with a classification, we have a certain value and you want to, we want to classify it. It's always this history, this. Time the time domain matters.
Deep: Yeah. I mean, I'm just gonna, I'm almost imagining like, so you've got this time series that generate, let's take your, your heat exchanger template, right? So you've, you can, you can model that out. You can have this kind of standard heat exchanger template. My guess is there's like a standard set of variables, like temperature and pressure and et cetera. The question I would ask is like, if I saw like a million or thousands of different heat exchanger problems, and I'm looking at that maybe multivariate signal. Are there properties, uh, within that signal that are common to this problem class in general, if so, it feels to me like you could bootstrap a network that learns the general gist of heat exchanger stuff, and now go take it and apply it to your particular problem.
Dr. Werner: Absolutely. Um, totally true. The challenge is not there. The challenge is to get to the state that you get the data into such a model. Get it out of, let's say from. Sensor that is out in the field mm-hmm and plug it into a framework, uh, that it can actually, uh, reach there. I know there's some research in the direction of like usage of neural networks and so on for exactly these problems. But as I said, it's in an early state for the easy cases like that, it's completely straightforward for the more complex ones or the more complex so-called unit operations, meaning heat exchange as a unit operation, like 40 more complex unit operations. It's not so simple anymore.
Deep: I didn't quite follow your sensor argument. So you're saying like, is there something about the signal in the sensor that's different than what we can model or something
Dr. Werner: let's take a step back in order for your machine learning to, uh, model to. You need data and observations, right? Yes. So that, let let's now step that aside for a second. So now let's go to the reality factory world. You now have a sensor that is sitting on a heat exchange. Mm-hmm the measurements that you take from D have to come to an infrastructure where you can run your model that we just set. And so the sensors are in so-called operational technology. They're embedded in so-called process control systems. And so getting them out from there. That is where like let's I would call it that's the ABUS that is in between, or let's say the, the chasm that is in between connecting these two worlds because the chemical industry traditionally doesn't do much in the cloud. They don't do much.
Deep: Oh, so you're talking about just the, the, just the, the semantics or the, the apps of getting these signal.
Dr. Werner: Absolutely. And, and then. To go to an industry where the operational technology runs on a process control system from one of the major vendors that in some cases doesn't have an infrastructure that records all these measurements over time, which is the basis for all machine learning. Step one, step two. If you have such a, uh, such a recording system, a so-called historian, then the question is how do you get the data from this historian? Number one, one time as a dump to a data scientist that would build such a model or in an operational setup, how would you continuously feed this data into a machine learning infrastructure in a world where all of these, uh, let's say the data scientists and all the computation infrastructure is hosted by it. Information technology, not operational technology. And then the next thing is the OT things. In many cases, Not running on, um, open protocols, they're using proprietary interfaces. And so that's where the challenge becomes into, let's say data accessibility and getting the actual data.
Deep: This makes sense to me, right? Because. You know, I mean, I come largely from the software world in the software world. Most companies are totally instrumented at this point because your sensors are basically other pieces of software. Who do you hire software engineers. They know how to like, take their own data and shove it somewhere. Uh, and, and therefore you wind up with, you know, all of the data kind of being centralized, typically in warehouses or, uh, you know, or lakes or something. And then you can start to like manipulate it, but in. Industrial world. Uh, you have, it's more complex cuz you have legacy physical things like distributed geographically. Oftentimes each of those things hasn't necessarily been wifi enabled with like data gathering that automatically lets you centralize it and get it up into the cloud.
Dr. Werner: And, and I would even say, take, take banks. Banks were the earliest ones that did digitalization like 50, 60 years ago with their mainframes. And so. They still those ones that don't have physical, that, that don't have like physical assets out there. They have the issue integrating these legacy assets into data lakes and so on. Like, and now imagine you have a factory you invested in some cases, hundreds of millions or even billions of, uh, of us dollars into building such a factory, you put all your process control systems. Now the thing is you will not take a lot of money in your hand just to upgrade them. So accessibility of these super heterogeneous systems, meaning. They are all different. They're all Cove to get them into one platform is a huge challenge.
Deep: So are you seeing a kind of huge movement as a part of? I don't know, we hear it about it as the internet of things, but are like sensor manufacturers, for example, making wifi enabled or Bluetooth access, centralizing componentry to like, get this stuff. Digitalized digitized.
Dr. Werner: That view in many cases, I think too easy because in an environment where you handle explosives, you'll not put in a Bluetooth or wifi sensor, right. Uhhuh, Uhhuh. So this is where the problem basically starts in this case. And also. Yes, they are offering systems to like put that to the next, uh, to the next century. But it's, it starts with the fact that if you replace a process control system, you, in many cases has a, have a five to six digit numbers to update them. Cause they are process critical systems. And to let's say, piggyback a new sensor in will not solve your issue. It's not like the 50 Euro, uh, sensor can plug in somewhere because it has to be harden has all the certifications for safety and so on, which makes the entry barrier for this like industrial internet of things, a lot higher. So what's happening, uh, on the other side is that, uh, vendors are going ahead and building opportunities. So you can use the existing legacy systems and enable them to let's say, stream out their sensor data by means of MQTT. Let's say that's an that's another example. So you retrofit existing systems to be able to send out their data, but that's also a huge effort.
Deep: Is that like, somebody makes a little thing with a camera that points at a dial and figures out what the dial says.
Dr. Werner: Oh God, no, no, no. Just, just imagine you have a process control system, meaning let's say companies like Siemens or Honeywell make those, and you have 50,000 sensors and control circuits in a. You will not replace that right away. However, you already have the sensor readings, you have all the data there. The issue is you need to somehow interface with that process control system and get the data out in a secure way. And that's basically step one. There's an interface called OPC open process control that exists. And so on, but again, it has to fit in the industrial certification standard. So there are a lot of software companies now that make middleware to retrofit old industrial assets in order to enable. Getting the data out.
Deep: So why did that data wind up centralized in that plant in the first place is that cuz there was, there were human operators or something that needed it.
Dr. Werner: So again, take a step back. How do you operate? Let's say a refinery or a chemical factory and it's not that there as millions of people were running around and turning knobs. No, they sit in a control room and that was digitized, uh, or digitalized in the let's say seventies and 80. So you have central control rooms with huge big screens. And, uh, what you do is basically you sit on your terminal and you type in a certain number and a control circuit out in the field, uh, make sure that the valve is open door closed. So the data is technically already there and you see the current sensor readings on your screen already.
Deep: So yeah, uh, just to kind of make sure I got that right. So the vision I had originally is like, you've got all these sensors. Go straight to the sensors, get the data. You're like, no, don't need to do that because we already had a reason to centralize them because we had these humans who were operating these plants. So whatever was just humans. It's not just humans they're like, or other systems out there.
Dr. Werner: Yeah, yeah, yeah. Yeah. So, and that's the other thing that the, like there's already, uh, an entire field of experts, like process control experts, like in electrical engineering that figured out rigorous models in order. Automate certain of these parts already. And, uh, that means the sensor readings are already there. However you have to get them out. And this entire first part is like the domain of operational technology.
Deep: So would you say maybe, I don't know, the 80% of the industry is in that stage trying to get those centralized operator data up into the cloud where they can manipulate it, or do you feel like that already happened largely.
Dr. Werner: Ooh. That's a very hard question to say that. I think that very much depends on the region, on the language industry, uh, like specific industry you are in. I feel like, uh, in some cases, um, the, the thing is why would you put it in the cloud? Or like, why would you put it somewhere that needs to be a pain? And if you are with the operating method that you have right now, you're pretty happy. And you have like huge margins right now. Why would you go ahead? Like, if there's no pain, There's not much happening in some other fields. Um, people have, let's say basically made this push in the first case to say, Hey, it's not just nice that we have the data in one place for the operators. Why don't we store that data to be able to do analysis data on that is the motivation to do, to make a so-called historian where you save all your sensor readings, like every minute or so on, so you can graph it out and do some analysis. This is where like, I would feel. Most major companies are in the next state is from that, that you are able to use that data in order to fit your models. This is where I would say a good number of the larger companies are in again. And then they're using the traditional first principle of flow sheet modeling tools in order to use that data.
Deep: That's yeah, that makes sense. So you've got, you've got this team of. You know, I don't know, maybe industrial engineers or other engineers that are, are, are, or data scientists that are trying to find efficiencies in the process. And the first primary directive is to drive the plant. But within that, there's like all kinds of data being shot off that, you know, we, where we can learn quite a bit. So that ends up, um, being the forcing function for the centralization of the data. Like, or in, in, in general, do you want to have people run around in the factory turning knobs?
Dr. Werner: I don't think that's very efficient, right? Yeah. So the first case was, let's say a central control room was just like you need, I don't know, five people to operate a $500 million asset, something like this. Instead of like having 50 or so that are running around. So it's a cost factor in terms of not to mention potential errors. I would imagine with humans running things, and now we're coming to the point where it's so complicated also to use artificial intelligence machine learning in there. Cuz now suddenly if you have an algorithm in there, you need to be as good and as adaptive and as secure and like a safe, uh, in operating. Compared to human, which means the level of certification you need to get for a certain algorithm to be approved, to be run automatically. That's pretty high because let's be honest, a machine learning algorithm or AI, they don't care if they make a mistake. However, the people living around, uh, your factory very much care. Uh, if the algorithm makes a mistake.
Deep: Perhaps you're not sure whether AI can really transform your business. Maybe you don't know what it means to inject AI into your business. Maybe you need some help actually building models. Check us out at xyonix.com. That's xyonix.com. Maybe we can help.
So is the, does the evolution look something like what we see with self-driving cars? Like self-driving cars you've got, or if you go back a little bit, it starts to be suggestive. So you've got the operator still, but you're trying to like, identify the guardrails, like, Hey, you're about to go off a bridge or, you know, and you try to be suggestive. They start using that suggestions and we're not yet taking their hands off the wheel.
Dr. Werner: So you could call that that's an open loop system. Yeah. Where you basically give a recommendation. Someone looks at their recommendation and says, yep. Makes sense. I'll act on this. Yeah. Whereas like a closed loop system would mean the AI is really in control or machine learning is really in control that you see in very limited cases, because coming back to our heat exchanger for a simple heat exchange, you could probably do that. However, if you take an entire factory. The level of, let's say certainty, you need to, uh, to do that. And the level of checks you need to do becomes so big that the investment in that is probably higher than just having an open loop system and say, Hey, by the way, I think this is what you should do and have someone validate it.
Deep: Yeah, it makes sense. I mean, it's also like, even if you feel like you're. Machine learning system is, uh, you know, well vetted, well tested that will only be relative to whatever environment it's seen before. So if a Fukushima like thing happens or something, then all of a sudden it's like, ah, and you know, as humans, we tend to be more trusting of a human in that scenario than a, than a, than a machine. And whether it's right or not. That's a different question. Yeah. Is that, so it sounds like the, the bulk of deployments. Of kind of AI in these operational context is in these. You know what you're calling, I think open loop systems where you're basically being suggestive to an operator when that happens, are we also discovering a lot of peripheral use cases or things off to the side, a little bit of that operator where the machines are picking up in like. Not just getting us to the point where the humans were operating before, but maybe into a more enhanced place.
Dr. Werner: Yeah. Augmentation, I think mm-hmm, the first part, uh, to get there is basically gaining trust, whatever. Like when, when we change the way we operate day to day and there's like suddenly a computer that is, let's say. Could replace you maybe in a farthest in the future. And of course the first thing is you're not really the person who will trust that system right away. There's that kind of like skepticism behind it. And of course that's dependent on where you are. Some countries have that more, some others have less of that. So that's the first step, like with taking easy use cases where you help an operator have their job, um, like to make their job easier. That's a good starting point because if they see they can do their job better. Get a better performance. That's the step. And then you can suddenly introduce more and more parts where, uh, where you can really help them and bring them up to the next level. But if you go to the let's. I tell you everything you're doing is wrong because this machine now, uh, is doing everything better than you as an operator. This will surely not fly because then they will not accept the machine. They won't, they will actively do everything to not work with it. Yeah. And to alert said, this is a, this is like a user experience thing, right? Like, you know, depending on how you present things and, uh, you know, and also changement even more so a change management thing, how you place it. And if you say. This machine is there to help you do your job easier, make it better and so on. It's probably a better placement. It's not just user experience of how you present it. It's just like how you place it in the cultural landscape.
Deep: So talk to us a little bit. About the software landscape that the systems that this operator is looking at, like who builds these kinds of systems. Can we think of these as like SAS systems where there's evolution happening behind the scenes or is it custom built per plant
Dr. Werner: prepare to be, uh, zoomed back to the 1980s. So every, like almost all these Sy. First of all, for many reasons, uh, you would not have the automation of such a plan in the cloud, just first of all, for reliability and safety reasons. That's the first part, cuz you need to like have very strict guarantees in terms of interlocks. Like when something goes wrong that you can put the players, uh, the plant in a safe state. I don't think that's a good idea to do that offsite. That's the first step. The second part is these process control systems are built by specialized vendors like large companies, Siemens. And so on that when you built these, uh, factories, you put them into place, you validate them, you run them. That's the entire operational technology part. It's completely custom made the software that the customer, uh, that the operator is seeing. Is in most cases, a graphical user interface that is very specialized, uh, 2d purpose. It visualizes how the process looks like and shows you where all the sensors are. And, uh, that is of course, a huge disconnect from basically what the process engineers sees, who sits with their machine, uh, in their office. In that case. And even there, I'm not really aware if any of the flow sheet modeling tools are already available as a software, as a service, most of them are pieces of software. They're installing a machine. In some cases you would add a backend on like a cluster, or also like when you think of similarly computational fluid, dynamic softwares and stuff like this, where you develop on your machine, and then you offload your computation into a cluster. but also there,
Deep: I'm not aware that this on a large scale is done, um, in the cloud, at least for the chemical engineering space and, and those capabilities, like you mentioned, Siemens building. Is it more like Siemens building a toolkit that some engineers that work for a specific manufacturing plant would modify and configured to their purposes?
Dr. Werner: The operator will very likely not modify anything of that. Once is put into place. It's basically preconfigured. You use it as. An appliance. I would almost say it's good place. And if like you want to do changes in there, like you wanna change a sensor. So on a maintenance crew on operational technology, we'll change the sensor. We'll reconfigure the P D loops. It's a project to do something like this. Cause everything has to be like made up to code, has to co uh, be compliant with Asia and safety regulations. And so on. You cannot just simply ship an. Everything. So it's actually like, like se like you buy an expensive package from Siemens and then they have some consulting folks that walk your plant and work with your people to configure it, something like that. It could look like that. In some cases you have internal experts that would, uh, perform, uh, updates on the, on the let's say parameters and so on, but updating a process control system, I would say you do that in a maintenance interval every couple of years, because it's a major. Major thing that is ongoing. And just imagine if you have a factory that produces 5,000 pounds of a material per day, and you sell this material for $500 a ton, that means 2.5 million every single day. Meaning if your plant due to a software or does not run for one single day, big problem losing two, 2.5 million. And so sorry, I should have said that probably earlier. This is why the aversion to like go into anything that is not let's. I wouldn't call it Bulletproof, but like really working is very, very high because just the risk of let's say that your machine locks up and doesn't work anymore is a huge risk.
Deep: So in that world, gaining incremental efficiency improvements via AI becomes a thing for the select few or something. The strong stomachs,
Dr. Werner: it's a tough cookie.
Deep: Let's see fast forward for us. Like. 10 years with everything you're seeing in the evolution, in the, you know, process, industrial chemicals, manufacturing space, and everything you're seeing with the evolutions in machine learning and AI, like what does that intersection look like? In, in 10 years?
Dr. Werner: I would say first thing is I think the similar to all other industries, I think in the next 10 years, we'll get, uh, to say our shit together when it comes to the integrations of legacy. That's the first thing. I'm pretty sure we'll get that dealt with is like huge. Let's say access issue one way or the other we'll have to deal with it. The second part is that I believe with the, I would call it very painful learnings that are right now done in the field of autonomous driving. We'll probably get a way better idea of how to deal with AI and safety critical. The third part is that I believe the, uh, the research that we're currently doing in terms of time series modeling will be so commonly applied to these issues that all these points that we currently cannot address, we can address at scale with standstills easily. You, you said earlier about like having reusable units and stand up, uh, pre-trained models and like, yeah. Yeah. You basically know what, what kind of type of model you need to take out of the, uh, out of the cabinet in order to solve the time from like the right now, there is to the best of my knowledge, no type of model structure that you could just pull out and say, Hey, I have a time series problem. I want to forecast a time series. And by the way, I understand some parts of that physical system behind. Because by the way, all the observations that I have go, I believe here with the research that is going on in 10 years, we'll have quite, uh, quite a bit of software, quite a bit of understanding and model structures on how to do that.
Deep: Do you see a lot of activity and maybe the startup world addressing some of these problems who's innovating towards this 10 year future that you're seeing.
Dr. Werner: So I see some activities in universities that are working on, uh, from the chemical engineering world. Mm-hmm, , there's a couple of groups working on this, uh, in the field of time series modeling. I see some interest, uh, from the financial sector, cause like all the data we're dealing with there, uh, is in many cases, time series data.
Deep: Yeah. I mean, there's an awful lot of people that are probably the best time series forecasters are all, you know, predicting the stock market. like on some level.
Dr. Werner: And, um, startup wise, the thing is the entry barrier that we talked about in the beginning is, is pretty high in this point, like to, to get your foot into, uh, an industry that has a tradition of being very, let's say. Strong in modeling. It is very strong in automation and has a, a culture of continuous improvement. I mean, there's all this six Sigma and so on applied for many times. Plus there are the legacy systems and there's strong, like disconnect between operating technology, information technology, these parts make it very difficult. I believe for startups to enter into that space because will like, for something like this, will you really. The funding that you need in order to hire this many people compared to like out autonomous driving and so on. I'm not so sure. We said earlier, a lot of the things are already solved in a, to a good state. However, I feel like there is one thing that could really change that. And that is if a strong push towards sustainability comes into play and suddenly there's a huge innovation pressure that you like really have to solve things that you previously couldn't. In that case, I could see that something is happening.
Deep: Well, let, let's talk a little bit about the sustainability tsunami or wave that's coming towards pretty much the entire economic world, right? Like, you know, you've got everything from governmental entities, you know, mandating. You know, carbon accounting, if you will, to, uh, hedge funds, you know, wanting to report on their, um, companies and being able to understand their companies and their progress towards sustainability, a how, what's the magnitude of that from where you sit and B how does that affect, you know, this, this world that we've been discussing,
Dr. Werner: I ask a hypothetical question. Do you think people will still be tooths? Coming back to, I'm pretty sure he has. And exactly. That's the thing like, um, even though like, if you look at the stock market, the traditional industries, haven't been the, let's say hottest child, uh, in terms of investments, you will still meet a lot of these things. And even if we remove, let's say all transportation, uh, fuels, meaning refinery, and so on, the question is still, will we still need paint? Will we still need the liquid crystals for our LCD displays? Will we still need. Let's say all the chemicals that we're using in our day to day products that we don't even think about. Yes. I would say very likely. yes. And then the question is, uh, then, then the question is how do we produce them? So the question is like, how can we find new routes to produce a lot of the things and the performance materials we're seeing in day to day life in a different. And that will likely mean that we have to think the entire value chain or the supply chain from raw materials into the goods we're using. And that creates, I believe a huge pressure in terms of like, how can we do that? Because I'm not sure if we have answers if we have the people there and that's similar to many other things, that's a tsunami wave coming in. Yeah. If you really say, we don't want to use anymore oil anymore. For like all the products that currently are produced by the chemical industry, I would say that's a huge challenge.
Deep: Yeah. I mean, it, it there's, there's many steps to it. I mean, I imagine like some of the earlier steps is just like, where does the power come from from a plant that produces toothpaste or what all these materials?
Dr. Werner: Absolutely. That seems, you know, like a, a good place to that's the tangible, that's the tangible, tangible first. Yeah. But you're kind of now talking. Well, okay. Let's assume you've done that. That's sort of easy within our world. How do you start to source materials? Um, you know, in an, in a carbon optimal manner, reduce carbon optimal manner. How do you get into your actual processes and, and, and let's spin a future. Let's say, think about the circular economy that basically all the waste that you're producing. Let's say you have a Ziploc bag. You have used a Ziploc bag. How can you reuse it? How can you also then if you cannot reuse it, reuse the material to make something else, all of these parts, not completely clear, cuz probably use in a thermal way. Meaning burning it to produce power is probably not the smartest to do with this material.
Deep: I was gonna wrap up, but I , I still have more questions. Couple more minutes. If you don't mind, we get a couple more minutes. Cause like this environmental stuff, like I've had like a, so we can do a follow up on that one. Like talk like, uh, have a couple, like have a couple of thoughts on, let's say the sustainability big picture for, yeah. I would love to do that. Like to do that. We didn't talk anything about Dataiku, do you want,
Dr. Werner: what I can say is right now is let's say when you have these whole points and let's say you have your data all in a, in a data infrastructure, mm-hmm you want to do exactly these analysis and you especially want to operate then such machine. So let's say from the heat exchanger we talked about previously. Yes. That's exactly what, uh, where platforms like data IQ come into play. Like having the whole life cycle from model development, signing off quality control into deployment, and then monitoring both the data and themodel in operation. That's exactly where you, you, you are looking into, uh, platforms like this when it's about. Governing them at scale, making sure that you have the risk in mind and that you sit abstractly on the infrastructure, but we will surely not be the ones that, uh, that deal with the process control system and interfacing. And there, we would rely on partners and we rely on partners actively.
Deep: Gotcha. Thanks a ton for coming on. This was, uh, this was super fun. I feel like I learned a, a lot about a lot of stuff I know. Very little about. So that was cool.
Dr. Werner: No worries. That's why that was why there. I also learned a couple of things and, uh, also that's, that's the important part. When you, when you do something like that and have a conversation, I'm also not the experts in all of these fields, so yeah, no, that was fun. Learning.
Deep: That is all for this episode of your AI injection as always. Thank you so much for tuning in. If you enjoy this episode on AI in chemical engineering and industrial processes, please feel free to tell your friends about us. Give us a review and check out our past episodes at podcast.xyonix.com. That's podcast.xyonix.com. And please also drop us a note. If you've got a specific topic you'd like us to cover. That's all for this episode, I'm Deep Dhillon, your host saying check back soon for your next AI injection. In the meantime, if you need help injecting AI into your business, reach out to us at xyonix.com. That's xyonix.com. Whether it's text, audio, video, or other business data, we help all kinds of organizations like yours automatically find and operationalize transformative insights.