Value Driven Data Science

[00:00:00] Dr Genevieve Hayes: Hello, and welcome to Value Driven Data Science, where data professionals become strategic experts. I'm Dr. Genevieve Hayes, and today I'm joined by Dr. Greg Biegel. Greg is the author of Digital Twins in Action and has been working as a researcher, software developer, architect, and product manager across academia, startups, and large enterprises for the past 26 years.
[00:00:29] He currently works as a solution architect with Amazon Web Services. In this episode, we'll explore the world of digital twins and how they can be used to enable better executive decision-making. Greg, welcome to the show.
[00:00:46] Dr Greg Biegel: Hi, Genevieve. Thank you very much, and thanks for having me. I'm really excited to be here
[00:00:50] Dr Genevieve Hayes: Whenever I go to a new restaurant, for the first few visits, I'll order a different meal every time. Sure, one of them might turn out to be a disaster, but the cost of failure is low, and it's the easiest way to determine which menu items I like the most. This same experimental approach works very well for most of the decisions we encounter every day in our lives and in our work.
[00:01:14] But what about the big decisions that land on executive desks where the cost of failure is high and there's no room to simply try again? For example, decisions around NASA's next space mission, where a wrong choice could cost millions and endanger the lives of the crew. Experiments are no longer an option, but this is where digital twins can help.
[00:01:38] Digital twins are digital representations of physical systems that are updated with data from the real world. They were originally only available to the largest and best-funded organizations, such as NASA, but recent technological advances have changed that, making them a tool that data scientists in all organizations should now consider.
[00:02:02] However, if you've never previously worked with a digital twin, it can be difficult to know where to begin, and this is a position I find myself in. I've always been fascinated by the idea of digital twins but never had the opportunity to work with one. So when I saw your book, Greg, I immediately bought it and reached out to you to appear on this show.
[00:02:24] So to begin with, beyond the definition I previously gave, which I took from your book, what exactly is a digital twin, and where are they typically used?
[00:02:35] Dr Greg Biegel: Yeah. Thank you, Genevieve, and I think it's really interesting that digital twins have seen their own share of hype and a lot of commercial propaganda since the late 2010s, I think, when Gartner first included the term in its hype cycle. And then have been driven by the Industry 4.0 movement, the fourth industrial revolution when the term has become more mainstream.
[00:02:59] So over that time, we've seen many different definitions, and I choose the definition that you provided in my book with one additional part that I think is really important, and that is that a digital twin should be built with a clearly defined objective in mind. So although the technology is available and interesting and powerful, that doesn't mean that is the only reason that we should build a digital twin.
[00:03:29] I think it's really important to start with clearly defined objective of what decisions you want to improve from building that digital twin. And I think the core components of any d-digital twin are really at the very start, a definition of those objectives with clearly defined success criteria so that you know when you've met them.
[00:03:54] Once you've defined your objectives, you need to look at how you collect data from the physical system that you'd like to model in your digital twin, how you would transport that data into the digital world, how you build a digital model of the physical world. And by digital model, I don't mean just a visualization, which is often what is billed as a digital twin but also mathematical models, physics-based models as well as visualization.
[00:04:22] And then once you have that digital model, you then need to perform analysis and simulation on that digital model and finally close the loop back into the physical system. And I think every digital twin, from one that I might have built of my house in the book all the way up to a complex industrial plant such as an oil refinery, follows that same six-step pattern.
[00:04:48] And what I often see is teams sometimes skipping that first step of defining objectives and measuring why they're building a digital twin and starting at the second step. And un-underlying these components are a huge number of technologies, including IoT, networking, storage, modeling, visualization, and of course AI, ML.
[00:05:10] And that's what makes digital twins such complex systems and difficult to approach and start with
[00:05:17] Dr Genevieve Hayes: Whenever I've heard of digital twins being used in real life, it's usually something like a digital twin of a road tunnel or a mine or a factory, some massive piece of very expensive infrastructure. Now the technological advances have made digital twins increasingly accessible to smaller organizations, have you come across any more unusual use cases for digital twins that maybe surprised you in some way?
[00:05:51] Dr Greg Biegel: That's an interesting question, and I think what I'm seeing today, what's really fascinating is people are starting to consider building digital twins of themselves or of their own systems at work. When you're working in an organization, It's really a physical system.
[00:06:08] You're accessing documentation, you're making decisions you're effecting change in the system. You might be writing emails, you might be updating document. With the advent of certainly corporate AI systems, I'm starting to see people building what they term digital twins of themselves.
[00:06:26] Now, that might be slightly different to the traditional idea of a digital twin being a physical industrial asset. But I think it is a valid definition because what they are doing is building this digital representation of their physical job, and people are having great success in really being able to build those representations of their daily tasks.
[00:06:50] So I think that's an interesting use of digital twins And there is a company called Delve.ai who are selling digital twins of employees. I haven't looked at it in too much detail but it's an interesting development I think.
[00:07:03] Dr Genevieve Hayes: That's really interesting. And I could also imagine, someone building a digital twin of themselves in their personal life. So you've got people who have smart watches that collect all that data, or Oura Rings. Presumably you could use something like that to collect your physical statistics, and create some sort of a digital twin of all those sorts of things that those devices collect
[00:07:31] Dr Greg Biegel: Yeah, that's really interesting and I think as a bit of a data nerd myself, I have a smartwatch and I'm constantly collecting heart rate data step data even temperature. I have a pressure sensor and altimeter on the watch. And so collecting all of that data and then using that to predict what my mood might be or how I might be feeling tomorrow based on the amount of activity that I've done today.
[00:07:57] And a lot of those devices have that prediction built into their software. But I'm always looking for ways that I can use that data in interesting ways
[00:08:06] Dr Genevieve Hayes: But it comes back to what you said before with the core components of the digital twins. You have to define the objectives. Because it'd be all well and good for me to have some digital twin that tells me how many minutes I'm going to sleep for if the temperature rises
[00:08:23] but who cares? Yeah what is that helping me with? If it doesn't help me to make any decisions, then it's just an expensive toy
[00:08:33] Dr Greg Biegel: Absolutely. And that's one thing in your personal life using your own free time. But in an organization, in an industry that can be represented at a much larger scale, that building something because it's cool. Technology is interesting, and as technologists, we love to use new technology.
[00:08:51] We love to see technical solutions to any problem. But the risks of building an expensive system without clearly defining what your objectives are is quite high
[00:09:03] Dr Genevieve Hayes: Yeah, in your book, you outline five capability levels for digital twins. Can you walk us through what those levels are?
[00:09:12] Dr Greg Biegel: Yeah, sure. And I like to take an example that I think most people are quite familiar with when I try to explain to, say, my family who don't know a lot about what a digital twin is. But if you start with Google Maps and you start with the base map in Google Maps, which shows a digital model of roads and buildings that is generated from, cartographic maps and satellite imagery, that's what's termed a descriptive digital twin.
[00:09:36] So it's largely static. It doesn't get updated very frequently, and it really lacks those real-time connectivity or predictive abilities. That's the first level. If you then overlay... in Google Maps, the traffic data onto that base map, which then color-codes the streets, based on density and speed gathered from many sensors and phones, that then integrates that data to visualize an evolving state of that road network, and that's what's termed informative digital twin.
[00:10:09] So it provides more information updated more regularly. Finally, st-staying with the Google Maps example, if you ask for directions, Google doesn't only provide you the shortest route by distance, but it blends traffic data with historical trends to estimate how congestion will evolve over time, and that's what's termed a predictive digital twin.
[00:10:31] So that uses historical data and current states to forecast future behavior. There's a really e-exciting development from Google. Just a couple of days ago, they published a paper de-describing what they term their planetary prediction engine, which combines data sets, Google Earth geospatial data, and AutoML to predict scenarios such as the spread of disease from just a natural language prompt.
[00:10:57] So that's a really interesting development, I think, in predictive
[00:11:00] Dr Genevieve Hayes: So it's like a digital twin for the planet
[00:11:03] Dr Greg Biegel: Very much so. Uh, predictive, digital twin, 100%. I've just started reading the paper. It was just published a couple of days ago. But from what I can tell, is using natural language to ask for a prediction of a scenario, and one of the ones that they use is the spread of the Ebola virus.
[00:11:19] This planetary prediction engine will then gather data from a number of different data sources combine it with geospatial data from Google Earth and then use AutoML to pick a, well-performing machine learning model to then predict the spread of that disease. So it's a combination of many technologies, but the interface to it is natural language.
[00:11:41] I think it'll be really interesting to watch where that goes.
[00:11:44] Dr Genevieve Hayes: That's incredible.
[00:11:45] Dr Greg Biegel: It really is.
[00:11:46] Dr Genevieve Hayes: Wow
[00:11:47] Dr Greg Biegel: And then moving up the capability levels we then get to what's called a comprehensive digital twin, and this actively simulates multiple possible futures and scenarios.
[00:11:59] And it uses a really interesting technique called data assimilation to constantly merge real world observations and data with mathematical models. So one of the ways that we see this in our everyday life without really realizing it is in weather forecasting. So there's something called the European Centre for Medium-Range Weather Forecasting, and what they do is they combine many observations from all over the world into their mathematical models of weather, and they use that to predict what the weather will be in the future.
[00:12:36] And it's sounds like a simple concept combining real world observations with mathematical models, but I think that data assimilation is really what distinguishes a comprehensive digital twin from a more traditional simulation. And then finally the fifth capability level that I talk about is the autonomous digital twin, and that's really a closing of the loop between the digital representation and the physical system based on predictions and simulation.
[00:13:06] And that may be as simple as triggering a notification on a user's device for them to go and check the pressure in a tank. Or it might be as sophisticated as automatically opening a pressure release valve when a tank's pressure is predicted to exceed safe limits. So it's a continuum of autonomy there but it's really that closing of that loop
[00:13:31] Dr Genevieve Hayes: Yeah, so with an autonomous one, you might have a digital twin for one of those traffic tunnels where it automatically turns on the fans if there's too much car exhaust in the atmosphere
[00:13:43] Dr Greg Biegel: Yeah, absolutely. And I think that's a really good example because that's taking data from the real world. So it's not just using predefined limits or thresholds, but it's dynamically changing based on the state of the real world in near real time
[00:13:57] Dr Genevieve Hayes: so what's the practical difference in terms of the kinds of decisions that can be supported by digital twins at each of these capability levels?
[00:14:07] Dr Greg Biegel: If we think about executive decision-making, often the biggest difficulty for executives is the lag between a physical event happening in the factory floor or in your asset, in your industrial plant, and then that reality reflecting in a boardroom dashboard, for example.
[00:14:28] So digital twins really elevate that executive decision-making, I think, by transforming raw operational data and we know that most organizations are drowning in data today. But transforming that raw operational data into actionable insights. So really trying to shift leadership from tactical troubleshooting based on operational data to be able to move towards proactive strategic thinking.
[00:14:57] And the way in which they do this at their core, and I think this goes up to the informative digital twin level is by providing that near real-time visibility into the telemetry from your physical system. So executives no longer need to rely on what might be stale data, and they now get a living view of their physical assets. So there's really clear value at that informative level, and I think that's where we see many digital twins sitting, at least initially.
[00:15:29] But really I think the true game changer for executive strategy is in the higher capability tier, so specifically predictive and comprehensive levels. The reason for that is that being able to run simulations gives leaders the ability to test decisions in a completely risk-free environment. So a simple example if you're an executive who wants to know the impact of pushing your manufacturing line beyond its normal capacity because you have a sudden market demand, if you have a comprehensive twin of that process, you can then use that to simulate that scenario.
[00:16:06] So you can use data assimilation to show not just if the machinery will hold up, but also how it might impact your maintenance schedule and your downstream logistics. If you produce more widgets, do you then have the transport available to get that to market? So you can run multiple simulations in a short period of time without any capital investment.
[00:16:27] And this really allows you to weigh the cost of potential downtime against the immediate revenue gain that you might get by increasing your production level.
[00:16:37] Dr Genevieve Hayes: So is essentially a digital twin just a big simulator of a physical system?
[00:16:43] Dr Greg Biegel: I think at some layer it could be considered a simulation. The way that I like to distinguish a digital twin from a more traditional simulation is that data assimilation. So the incorporation of data from the real world, from the physical system in near real time to then influence that simulation.
[00:17:02] So when I think of simulation, I typically think of something that's rather standalone. You might change parameters and run a number of simulations and see the output. My sort of mental model of how a digital twin differs from that is that continuous assimilation of data from the real world to then drive that outcome
[00:17:22] Dr Genevieve Hayes: What do the outputs of a digital twin typically look like?
[00:17:25] Dr Greg Biegel: I think that they look really like improved decisions is how I would probably term it. And that can include improved decisions from looking at a descriptive or informative digital twin. So being able to look at what the digital twin predicts the future will be and that's up to the sort of comprehensive level.
[00:17:48] And then beyond that, the outputs might be actual action at the autonomous level. But really it's about improving decision-making.
[00:17:57] Dr Genevieve Hayes: But obviously you've got a mathematical model or a machine learning model here that's going to spit out some numbers. I guess what I'm trying to get my head around is what are the numbers that get spit out by the model, and
[00:18:10] How do you use those to enable those improved decisions?
[00:18:13] Dr Greg Biegel: Yeah. And I think it also depends upon the level of usage of that digital twin. So within any organization, you have different levels within the organization, and each of those consume data and information at different resolution, so you might have an operator who's using a informative digital twin of a plant operation, and they're very interested in understanding that a piece of equipment is going to breach a temperature threshold immediately.
[00:18:44] So they would consume that alert or that prediction at quite a granular level. As you move up the hierarchy and then you get to the kind of the executive level, they would consume the output of a digital twin at a much different granularity. So they might look at the entire value chain of the organization and want to understand how changing a maintenance window in one part of your plant might then impact the delivery of a particular piece of product to a customer in two months' time after it's been loaded onto a ship and across the world.
[00:19:22] And then they might also want to bring in information about markets and exchange rates and a lot of different levels of information that single maintenance window or machine failure would have an impact all the way along that value chain.
[00:19:39] Dr Genevieve Hayes: So the alert level, it might be there is a 50% probability that in the next hour this part is going to fail. So we've got a probability that's going to turn into a alert which is either autonomously making the decision to do something to stop that part from failing or sending a message to a maintenance person to come and replace that part.
[00:20:05] With the one where what will be the consequences of this, it might be what is the average impact on profits two months into the future and the range of uncertainty around that? Would that be correct?
[00:20:25] Dr Greg Biegel: Yeah, and I do see it that way, and I think it really comes as a result of one of the key benefits that I see of a digital twin, in that it breaks down organizational silos between data stores. So what one of the key benefits of a digital twin is to add context to complex telemetry.
[00:20:45] So if we think about our operator who's looking at the plant operations, and they will be getting telemetry from many different sensors, many different systems. Now, often that data lives in isolated systems, so we find our operators having to do a lot of swivel chair integration.
[00:21:02] So they literally will be switching between different systems to understand different views. So a maintenance view, a engineering view, and a telemetry view. And together, they need to contextualize that data in their mind and make a decision based on that. So I think at that operator level the breaking down of the silos and the joining of that data together into a single pane of glass is really valuable.
[00:21:30] But then when we go up the organizational hierarchy when we get to an executive level, they're not just looking at the fact that the bearing temperature is rising. They're looking at the projected financial impact of that specific motor failing based on that motor's role in the value chain.
[00:21:50] So the availability of spare parts from their supply chain, the availability of necessary skills to repair it. And this is very much coming from an industrial background where you have complex, expensive assets that need to run continuously. And any downtime has an impact all the way down the value chain and it can run to many millions of dollars.
[00:22:14] So it's this alignment of the physical reality of your asset with your financial forecasting within the entire process that you're operating in. And that can allow you to optimize your resource planning and ultimately, build that more resilient organization.
[00:22:32] Dr Genevieve Hayes: So it sounds like we're gathering all these different data sources, and then we're combining them in order to see how these different aspects of a physical structure of some sort or a person how they will all interact with each other to effectively create some sort of causal model which allows us to determine if we make this decision or
[00:22:59] if something happens, then what will be the consequences, so that we can make a potentially costly decision with greater confidence.
[00:23:07] Dr Greg Biegel: Yeah, I think that's , a good summary. And I think what's really important in digital twins too is data contextualization. There's a lot of data in many different systems, and looking at each of those pieces of data in isolation it's difficult to understand how they relate.
[00:23:25] And now that we move into a world of generative AI, where we have everyone in the organization using generative AI tooling, I think what's really important is the contextualization of siloed data sources within a digital twin using a technique such as an ontology implemented by a knowledge graph as the implementation of that ontology.
[00:23:47] It gives the ability to contextualize data that can then be used to feed generative AI to result in much more useful output from it. So what's really exciting, I think, these days is to see executives being able to use corporate AI tools coupled with MCP servers that sit in front of APIs that then are really fronting a digital twin, and be able to interact with the data from their organization in near real time and use that in their decision-making.
[00:24:24] It's making that data so much more accessible at every level of the organization. And it's really fascinating to see executives being able to talk to their data via generative AI. But behind that is all of these components of being able to identify the data sources, transport them into the digital representation, store that data and then build that model.
[00:24:51] So that on-ontological model that relates the fact that, this sensor reading that you're looking at here relates to this machine, and that machine is creating this widget, and this widget, you have an order for ten thousand of those widgets that you need to deliver within two weeks. So all of those relationships between those pieces of data then enable you to make better decisions
[00:25:14] Dr Genevieve Hayes: What's the single most important piece of advice that you'd give to someone just starting out with digital twins?
[00:25:20] Dr Greg Biegel: Start from a decision and not a technology. So as a technologist, I love technology, but really when you're looking at building a digital twin name one decision that you make today how often you make it, and what it costs when you get it wrong, and then build the smallest thing that makes that decision better.
[00:25:39] So everybody wants to start with a three-dimensional model because it's the fun bit. But I think the 3D model, sometimes you don't even need that. So really start from a decision and not a technology
[00:25:50] Dr Genevieve Hayes: For listeners who wanna get in contact with you, Greg, what can they do?
[00:25:54] Dr Greg Biegel: Oh I think reach out to me on LinkedIn is the best way to get in touch with me. And I'm not sure how I give my details for LinkedIn,
[00:26:01] Dr Genevieve Hayes: I'll put them in the show notes.
[00:26:03] Dr Greg Biegel: But thank you.
[00:26:04] Dr Genevieve Hayes: And that's it for today's episode of Value Driven Data Science. But if you want more from Greg, next week you can catch our Value Boost episode, where you'll learn how to map a physical system to its digital representation, one of the first steps involved in building a digital twin.
[00:26:22] And if you found today's episode useful and think others could benefit, please leave us a rating and review on your favorite podcast platform. That way we'll be able to reach more data scientists just like you. Thanks for joining us today, Greg
[00:26:37] Dr Greg Biegel: Nice to be here, Genevieve. Thanks for having me.
[00:26:39] Dr Genevieve Hayes: And for those in the audience, thanks for listening. I'm Dr. Genevieve Hayes, and this has been Value-Driven Data Science.