Rockwell Automation, Inc. (ROK) Earnings Call Transcript & Summary

Hello, and thank you for joining today's webinar. Before we get started, we have a few housekeeping items. The audio for this event will be streaming through your computer speakers to make sure your volume has turned up and speakers are turned on. We will be showing a video in today's presentation. You will be able to address the volume of the video on the bottom left-hand side of the video panel. Our webinar platform performs best on Chrome and Firefox browsers. On the lower left-hand side of the presentation, you will see a Q&A box. We encourage you to enter any questions you have throughout the presentation here, and we will answer them at the end of the presentation. If you're having any trouble connecting to the webinar, please take a moment to refresh your browser and disconnect from your VPN. If you are still having trouble, please clear your cache. We have instructions in the handout section of the webinar platform. All of the panels on the webinar platform are adjustable. To resize, simply click the corner to adjust or hit the maximize screen at the top right-hand corner of each panel. Today's event will be recorded and will be available immediately after it is completed. You can access the recording utilizing the same link that you use to access the live event. After the webinar, we will also be sending you an e-mail with the resources from today's event, including the slides, handouts and event recording. Additional information regarding today's topic can be found in the Handouts panel on the webinar platform. With that, I would like to introduce today's speaker, Michael Tay.

Awesome. Thanks, Shannon. The topic for today is to Operate Smarter with Prescriptive Analytics and talk about what that means and how that can help you in the process operations. I'm going to talk a little bit about driving value or a good bit about driving value, elevating operators to a higher level, leveraging analytics and then describe -- go a little deeper into what we mean by prescriptive and what is prescriptive, have some product demo to let you see how that works in the product and then go back to the kind of core principles and then summary and time for Q&A. We had a food and beverage manufacturer who was challenged with their drying operation. They had variability and trying to meet regulations on a product that was reviewed. The leverage of the application and solution achieved a 42% reduction in quality variability with a dryer that tends to be moisture; a 0.38% increase in average moisture content, and that is pushing up closer to the specification limit, but that's 0.38% more products shipped with the same raw materials; a 6% to 16% increase in throughput out of the same machine with the same equipment, so increased profitability. French fry manufacturers have the same challenges of any food and beverage. They take a raw material that's more varied than most and try to get a consistent fry that can be sold from various restaurants. In that, they were able to increase throughput on the line by 15% or up to 15%; reduced energy costs up to 12%; reduced quality variability up to 80% (sic) [ 60% ]; and there's a broad range of quality parameters, even though they don't completely fry the product at the manufacturing site; increased yields up to 1%, so more on spec product, less off-spec downgraded product; and operators able to focus on more value-added tasks where their human insight has more value, like cleaning and quality checking. A chemical manufacturer making styrene monomer for a polymer manufacturer had challenging quality problems and always trying to be driven as efficiently as possible. The styrene/ethylbenzene distillation is a very long, slow dynamic process like a lot of high-purity distillations. They were able to reduce product variability up to 90% -- almost up to 90%; reduced production capability, so throughput by 2%; and decreased energy use by 5%. After the distillation project, they went to the reactor and saw similar value and gains in their yields and profitability. In egg processing, a corn-to-ethanol producer was trying to maximize throughput on a 50-million-gallon-per-year design factory and wanted to see if they could get up to 60 million gallons per year of production. They were able to increase their production by 7.9%, exceeding targets; improved fermentation yield by 5.8%, that means they make fuel ethanol from the same amount of corn 5.8% more; and then realized an ROI within 6 months, so high payback, high-value proposition on a variety of process operations. Similarly, in minerals and mining, this is a flotation circuit in Latin America. They were struggling with degrading quality of ore coming out of the mine, so they've gotten the highest-value ore out of the mine and quality was degrading, variability was high. They were able to, with model predictive control, drive that complex flotation circuit from 46% yields to 73% quality, that's higher pure ore coming out of the processing; a 30% reduction in chemical dosing that is less additives, less cost to produce; and ore flow continuously maximized, so they did increase throughput as well. In oil and gas, an oil producer was working with secondary oil recovery and wanted to maximize their oil production. And oil pricing and margins are quite high right now so increasing throughput has huge value for them. Because of the secondary oil, the constraint or the bottleneck of the processing facility was the water injection or the water management from recovery system to the down-well pumping. They were able to increase water injection by 2.4% or 36,000 barrels of water per day, and they got a linear increase. So every bit of water that they pumped in, they got more crude oil pumped out. And they were able to increase crude oil production by 548 barrels of oil per day, which is a significant value proposition, and at the same time, decreasing energy consumption by 3% by better management and better leveraging those pumps. And in a pulp and paper process, they were struggling with fiber balance and trying to cut the cost of raw materials going into a sheet of paper in their fine paper operation. They were able to increase profitability between $1.5 and $4 per ton on different grades, so there are different opportunities for increasing profitability or the fiber cost per ton; reduced off-spec production by 75%; and then reduced variability on their strength and color quality so that they had a more consistent product going to their customers. Each one of these across a broad range of applications shows this idea that you can drive quality with a prescriptive analytic, achieve good value propositions and get more out of the manufacturing site. So I wanted you to stop a little bit and think about those types of opportunities and identify what's your highest value opportunities. Where do you think, on your operations, you have the hugest opportunity, the largest value or at least what you're struggling or driving to do today? Give us some answers on what you're thinking, enter your thoughts as far as throughput, quality, energy, variability or other things that we've missed here. We're starting to get some answers back. It would be helpful if you gave us some ideas or at least thought about it so that you could see what your peers are thinking of and what you're thinking of as far as value proposition. And everybody is going to see a little bit of value on different things or a balance of these, but it is good to sort of give it some thought, think about it, maybe even state externally what you think is that value proposition. Because it does make it more concrete in your mind if they talk about psychological studies they have done. When you put a statement out there, it makes it more realized, and you become even strongly -- more strongly motivated to achieve those benefits. All right. And we have a range of answers. I think we're sort of flattening out. So let's take a look at that and see -- I'm not sure if you can see. But what I see is about the most popular was reduced energy, about 31%, but also fairly high percentages on throughput, on quality. A little bit lower on just performance variability, and it's valid. That's not such a concrete value proposition. And several people now adding in other as the range. Sorry, I'm learning the gyrations a little bit, but now you can see the same data I did. As far as what's driving your peers and what's driving you as far as energy and other value propositions, and it's a pretty broad range of opportunities. You probably saw in the case studies that many of these applications and most prescriptive analytics are leveraging optimization and trying to get a balance of value across conflicting or apparently conflicting challenges. It's that ability to not just maximize throughput or reduce energy, but do it while you're still making good quality product that's sellable and has a stable quality. So when we think about turning data into action and all of analytics, you may have seen this chart before with descriptive and diagnostic as the basic analytics towards saying what's going on? What is causing issues too predictive and prescriptive? And the blue triangle at the top that gets smaller and smaller is the level of human intelligence that we think or sort of impression of the amount of human intelligence that takes the analytic to action and ability. And so prescriptive, our topic for today, is at that far end, where it's limited -- or less human intelligence between the analytic and the action. And what we talk about with model predictive control in many of these applications is a closed-loop prescriptive analytic, where there is no human in the loop. The human is supervising the loop, but the action happens without any human intervention. So it's getting human guidance on a constantly driving solution. When we look at the FactoryTalk Analytics Pavilion8, we think about it as optimizing the plant. And when we say no human in the loop, we're elevating the operators, and they're no longer thinking about temperatures and pressures and flows, which is implied by these PID loops and the PLC and the DCS. So there's single variant, what the operator is adjusting today. But now the operator is adjusting model-predictive control and targets on throughput and constraints and how high can we push the plant and energy efficiency and quality. So we're elevating the operator with an analytic that's driving it with technologies like model predictive control, with hybrid modeling and with a robust, reliable closed-loop solution. So something that can be left alone to get a useful result without constant supervision, which the PID loops, the temperature pressure and flow, need a little bit more attention to keep that plant in line. The other technology that's shown here to decide that soft sensor technology is a predictive analytic, but it's working hand-in-hand with the prescriptive model predictive control, doing an inferential or an estimate of the lab quality. And a lot of those case studies that we talked about, the flotation circuit and the french fry processing, are looking at lab qualities that are used to decide whether a product can ship or not, but we're controlling it directly with an inferential that's working in real time. And so this is sort of that idea of a prescriptive solution. The technologies under there, model predictive control is a matrix of models. And each of these models between the manipulated variables and the objective variables, the controlled variables, are dynamic. They're saying, when I move MV1, how long does it take? And what does it look like over time to change CV1 or CV2? They're multivariate. They're dealing with multivariate systems because we're coordinating this matrix of models and there are interactions in several ways. So when we talk about conflicting objectives, we're dealing with balancing multiple controlled variables, but we're also dealing with multiple adjustment variables, manipulated variables that are moved to control that. They're dynamic predicting, and we'll show that in the demo, but each of those have a predictive horizon. When I move this input, what will happen in the future? When this disturbance variable moves, what will happen in the future? And that dynamic predicting is really important for controlling slow, long systems and long dynamic systems like that ethylbenzene/styrene monomer splitter that we improve variability on. They enforce constraints. So when we maximize throughput and up to equipment limits or quality limits, we can put limits on those and actively enforce them across the predictive horizon. And that lets us push closer to those limits. And they reject disturbances and the disturbance variable lets us predict what's going to happen before it happens and then take corrective action in advance. So this is sort of the core of model predictive control. The Soft Sensor is -- we look to -- we start with the quality control lab and how you control and release products today. And there's this time impact in the middle because operators have to get a sample, the sample has to be analyzed, and the results have to be fed back to the operator. And when we look at a soft sensor, it's sooner, it's available in real time for supporting direct quality control. It identifies the -- what changes will result, and it can reduce lab frequency. So there's good value in the Soft Sensor. The good value in model predictive control is that constant driving equality. But the Soft Sensor or the virtual analyzer, building a model from lab data and historical data and putting that online is to improve that quality variability and support closed-loop control on it when you don't have closed-loop measurement. Now there's an application of the same Soft Sensor technology that we call a Predictive Emissions Monitoring System. So this is a Soft Sensor of emissions, but it has a wrapper around it to meet U.S. EPA requirements so that, that Soft Sensor it's not used to release product, it's used to tell that you're in compliance with emissions reporting requirements. And so we use the same technology. We're measuring fuel flows and quality and airflow, and we're predicting the emission outputs and reporting it in a certifiable way that meets requirements. So it has a wrapper for quality control on predictive emissions. It was first patented and introduced in 1993. It's been in more than 250 stacks. And we mentioned California because it's some of the toughest U.S. environmental compliance regions. And it innervates with all control systems in DCS. So all of these technologies will work against any control system, Rockwell and others. We have a list of kind of existing applications for -- that we have used successful compliance and a list of fuels. Part of what's here that's important on this slide is that every fuel doesn't work with a software CEM. Some solid fuels are [ omitted ] or not included here. We don't have coal on this list, and it's not appropriate for coal. But many types of combustion units and fuel types fit into that software CEM. And the reason for it, you get environmental compliance, it's extremely robust, and it tends to be a lower cost than the hard -- that kind of plastic hardware analyzer. Another reason for a software CEM, because we've been talking about causal modeling, it also includes causal modeling and a strategy for having models that are causally correct and can be optimizable, which some people are using it both for emissions reporting, but also for constrained optimization, including emission sources. So I've been talking, so far, about prescriptive analytics, and this conversation is about prescriptive. But I want to give you a little bit of time to think about other analytic opportunities for you. And what is your next step in deploying analytics? So where are you in your maturation process? Some people think about prescriptive analytics as the last strategy for analytics, and they're looking at visualization first or they're looking at other. We have some very cool autonomous AI that takes care of itself. And so Rockwell has some capabilities in that. I do want you to consider that you can start your analytic journey with any of these as your first step. So it is a good thing to think about where your opportunity is, where the value proposition is. Some people do believe -- start descriptive then look at diagnostic and predictive. My personal philosophy, and you should have your own, and this is sort of the point of this question is to give it some thought, is that you're going to look for high value and start where there's good value, and then that's sustainable in your plant that's going to be one that your plant is ready to live with, but then also focus on opportunities to drive an analytics journey based on where you see the big value. And a range of users, I encourage you to look at Rockwell and visualization opportunities, which it is that ability to see into data, mostly descriptive, but somewhat diagnostic. Autonomous ML tends to be more predictive, but there's a range of technologies that Rockwell has. It's simpler and needs less effort to do. And what we've been talking about with prescriptive is something that has no human in the loop. So there's a little bit of extra engineering and costs associated with it to drive those good value propositions. So it is -- this range of selections doesn't seem inappropriate, and you're going to look for where your opportunities are. So what do we mean by prescriptive? And part of that, I think what much of the marketplace forgets is that you need a causal answer. You need a model that's not just predictive but is correct in the derivative, what drives optimization? It's one that maps to reality. And so there's a couple of strategies. Rockwell has some good experience with causal modeling. Judea Pearl is a UCLA expert on causal modeling. And his quote that I like a lot is that every causal inference task must rely on judgmental, extra-data assumption or experimental -- experiments, right? That's an awesome quote. It has a lot -- it's probably worth thinking about. And I'm starting right away with this idea that if you do experiments, you can develop a causal model. You can do controlled step responses, and I'm showing the step-test automation and identification of the step-test automation from Pavilion8 in the right here. And so that's the strategy. A design of experiment is an efficient way to do -- develop causal models, and that is also used for the environmental solution. Another way to develop a causal model is to use a physics-based model, one based on known physics with expressions of known physical relationships and then train the coefficients. Still, there are coefficients that can be optimized to map as best with the plant data. So a physics-based model, as shown here, is another strategy for causal modeling. And the last strategy has a lot of more bullets on it. So we're getting a little bit more effort, a little bit more technology requirements to do that. But the hybrid composite models that I referenced early is a type of modeling that leverages or combines physics space with empirical modeling in ways to leverage all the information you have. That flexible composite modeling in Pavilion8 is shown here and a couple of ideas as far as combinations of models that are supported. But in addition, you're thinking about -- particularly with the empirical modeling, independent parsimonious input. You don't want too many input features, you want them all to be independent. You want to deal with correlations and input data. We're pretty strong in this process world of time influences. So when an input takes a longer time than another input, then managing that in the modeling environment can be very helpful. Most AI or ML once balance data and improving the balance of the data can be important, particularly if it's poorly balanced, and a lot of plant operations data is poorly balanced. When we say structurally dealing with confounders, there is a lot of things in process data that confuse a modeling engine, gets you the wrong derivatives. Some of it's closed-loop control action. Some of it has to do with correlated inputs, but also that going back to the mean, that idea that data tends to look at extremes and predictive analytic technology, we'll look at extremes. So regression to the mean is another confounder in data that you sometimes want to deal with. And we have a technology called gain-constrained training, that's very helpful if you have other sources of information that you can use. The core kind of message to this causal modeling and why is that important for prescriptive is that correlation is not causation, right? Something can be highly correlated, but it may have nothing to do with correlations. It may -- because of confounders and data. And so we're usually using extra data strategies to get a prescriptive model that can be used in a prescriptive analytic. So I'm going to show some of the technologies and some of the tool sets that we have in a quick -- maybe a short video here. Let's see. [Presentation]

So what you'll see is we're just looking at a predictive control display. And it has a predictive horizon. So that vertical line in the center is now -- and I'm changing the range of the data so that it focuses a little bit more tightly on the control action and the moves. But we have a predictive horizon to the right. We have a short-term history to the left. I'm also showing how the user would interact with it to change limits. And so I just lowered a constraint as an operator might adjust the model predictive controller. So if you've lowered the limits on this dryer, you're going to see the controller then react to try to move that down and support the constraint. You can also change setpoints. And I'm just showing the range of parameters. So I've just increased the moisture on this dryer. So the target on moisture has gone up a little bit. The pressure control limit is there. While the controller is thinking, I switched to a history mode that we have that looks at how the operator looked at it in the past. And I'm just stepping back in time and watching the model predictive control actions in the past and just showing a ways to interact and troubleshoot. What you see is that a disturbance happened, but the control stayed quite stable and the control is acting the way that you would want it to. So the history mode can be quite useful towards identifying disturbances here. We see another disturbance, another stable control response where the controller reacted. And here you see the model predictive controller back in the real world has changed the target, is enforcing the constraint, and so it's moving up to that target and stably controlling the constraint. And now I'm going to turn on a maximization. So if I have a target on one of the manipulated variables, and I turn that on, and I'd say, "Hey, I'd like to get more throughput out of this dryer," I can do that type of optimization. And so it's just a small change in the [ feed ] rate target, so I'm trying to push more into the dryer, but you also see that I've got this constraint. And the predictive horizon looks like the model predictive control is going to exceed the limit a little bit more than I'm happy with. So I can go back into that and change the coefficient or put more weighting on it. Now I'm looking at sort of what an engineer might do, where I want to enforce that limit more strongly. And I can wait a little bit, and you'll see that the model predictive controller will not exceed the limit so strongly, it can also reduce the weighting on that target to make it less important. And so the model predictive controller that's optimizing this dryer is something that you can interact and the engineer or the operator can interact with. So now it's enforcing the constraint. It's pushing up the throughput. But I also start to see some update where the moisture target isn't quite where I want to put it. So now let me just increase the weight on the moisture target a little bit more like I did that constraint, and that will let the moisture target be more strongly enforced. So again, I'd like more throughput from the dryer. But I want to make sure that the moisture target kind of pushes back towards the limit in a way that I'm comfortable with. And so you're looking at kind of a dynamic optimization. So this is another technology where we're building that virtual analyzer or we're building a steady-state model. This is a paper machine data set. And I'm showing just some data cleanup when I talked about data balance and -- or just measurement problems in the data set. And so what you're seeing is sort of an easy way to cut data that we think -- here is the machine shut down where we think it's not as interesting. Actually, this is a polymer process operations that we're looking here. But I'm eliminating outliers, making sure that all the data is stable, so exploratory data analysis and then here data wrangling to eliminate these outliers that are not of interest. And just trying to show that it's fairly easy to cut or filter or clean up data the way that you want to. This one had data weighed down to a negative something. I just set it to just about 0. And so I've now cleaned data in a way that all of my measurements look reasonable. And I can then create a virtual online analyzer. This is showing correlations against melt index. So this is looking at a ranked index to show what's strongly correlated or less strongly correlated with one of my output variables, melt index on this polymer. So then if I create that model, now I'm just selecting inputs and outputs. I have 3 things that are going to be outputs for my model predictive controller or my virtual online analyzer. And I can train that model, get a good R-squared. This is a little bit faster than a real data set, but it is showing how the product might work. And I can investigate any of these models towards model accuracy, model inputs and sensitivities. And this is what we're thinking of as an explainable modeling engine. So we've built a predictive analytics on 3 polymer processes. And we're looking at the solids concentration in the melt index and looking at the prediction accuracy across the data set with the residual plotted below and then the overlap of the predicted and the measurement data at the top. And then we can look at the input versus output models. So if we look at -- I think this is melt flow, and I'm looking at the nonlinear relationships of each of the inputs to the outputs. And I can also look at, on all 3 of these outputs, the sensitivity ranking. So the relative ranking of the input variables on each -- so here we look at my 3 output variables: production rate, melt flow and solids, and the most -- from the most significant to the at least significant. What you also see is that all those inputs were fairly similarly ranked, which makes this a complex multivariable control problem. To build this into a nonlinear model predictive controller that's important in polymers, we can export that as a predictive model, and then we're going to show building a controller from another data set and bringing that nonlinear model into the controller. So first thing to do is to import this model into a SolutionBuilder so that it's available for developing a nonlinear model predictive controller. I just go out and grab the model that I just exported and imported into a SolutionBuilder. So now I've created a predictive analytic. We started with the visualization console. I've created a predictive analytic. And now I'm going to use that predictive analytic to build nonlinear gain -- nonlinear model predictive controller. We start this with the data set. This is a dynamic data set ready for dynamic identification and identifying the manipulated variables, the control variables, any disturbance variables in that model predictive controller. The data set has information on time that's used in the model predictive controller. And in the identification window, we set that up for identification. And we've got steps on each of the variables. So this is experimentation to develop that causal model. I did make a mistake here on selection, and I noticed that with the video. What you're trying to do is select each step for identification, and we're now selecting each step where each manipulated variable and disturbance variable shows movement. In the back end of this data set, every variable is sort of moving, overlapping, and we're showing to activate for multivariable identification. And then we can identify those dynamic models, run through that identification. See the step responses, they're all shown as high-quality models. Predictive versus actual are looking quite good and mapping well, not for the one that I mistakenly selected, but we'll ignore that for now. Hopefully, you'll grant me that. But it shows then how well the modeling was on the -- each of the stages. And this is with the linear identification and the linear modeling. So there's going to be some offset, particularly on the gain because this is a nonlinear control process. So to fix that, we import that nonlinear model that we identified in the other tool, map each of the variables to the input and output to make sure that the variable names are aligned between the model predictive controller and the nonlinear model. Then update all those models, and this is creation of a nonlinear model predictive controller. Now then we would go back and reidentify and fix that identification step. But that was what I wanted to show as far as the product and usability and how we build these types of predictive analytics. So I gave you a quick run-through of the key products that we're using in this prescriptive analytic technology for both the VOA and for the model predictive control. I did not do the software CEM, but that's also shown in SolutionBuilder. To buy this product, the licensing is set up with a base functionality that's focused on the first technology that we showed, which is the visualization. It includes some runtime applications, then sort of a base functionality of one virtual online analyzer and one software CEM license. The SolutionBuilder license for developing the applications are necessary when you're developing, but may not be necessary if your used cases is operating and another partner or someone else is developing your applications. SolutionBuilder is an optional add-on. And then there are runtime licenses based on the needs of your application, the number of manipulated variables for our model predictive controllers. So it's scaled to buy the manipulated variables. And then the number of model predictions for virtual online analyzers is the scalability for that runtime license. And then the software CEM is the number of emission sources or emission points that you are predicting on, and it's generally a number of virtual analyzers. You saw in the previous slides, it could be NOx, it could be CO, it could be O2, it could be other things, SOX if you had sulfur measurement in the feed. And so the CEM is licensed with an assumption that there are multiple outputs in general based on emission sources. The technologies that I showed, SolutionBuilder is for building that model predictive controller. You can create calculations and applications with -- that run behind the console. But you can also develop a model predictive control from data, from a step-test automation that's included or from a previous model predictive controller. You can compare models. It has a pretty strong version management, including snapshot and snapshot comparison and merge capabilities. And the technologies that we've shown in SolutionBuilder are optimized for developing, and we spent a lot of time focused on this building of the model predictive control. Data Explorer, the steady-state modeling engine used for both virtual online analyzers and also nonlinear model predictive control models is an explainable modeling engine with exploratory data analysis, data wrangling built-in and then analysis plots on both accuracy, relative input importance and the response vector, input/output response of the model. So that's the Data Explorer technology that comes with SolutionBuilder. So when you buy a SolutionBuilder, this technology comes with it, although it's integrated. The Console Server is a visualization focused on model predictive control, identifying the variable type, the predictive horizon. It has that playback on time that you can -- that we've shown, shows KPIs as well. So we have a calculation engine behind it, but the interactivity with model predictive control is built in. And there's a lot more pictures here that show sort of the wind-up status, so when the PID loop can't go any higher or lower, that will also be mapped to the controller and indicated. So this is the MPC Console Server. And then the persistence not only shows that back calculation of the model predictive control, but how the value is running and how the stability of the model predictive control performance has been over time. So that's a part of the console server value proposition. Kind of just to close, Murray Goulburn is one of our reference customers, another dryer application. They are, which I like a lot, is that, "The solution performs as well as your best operator, 24 hours a day, 7 days a week, resulting in the highest-quality product consistently." model predictive control is sort of now a new tool that the operator is leveraging to drive performance, but it's designed to do what the operator used to do all the time consistently, constantly driving to quality, to value proposition. And so when we see these improvements, these reductions in variability, it's because we're controlling quality directly, it's because we're driving efficiencies and yields directly and doing it in an automated way without a human in the loop. So it drives throughput, quality, energy efficiency. And that was the last slide, and I think we have time for some questions and answers.

That we do, yes. First one, you gave a lot of really good examples in the beginning of some use cases. And so somebody wanted to know, are there any samples of water and wastewater industry use cases?

There are. There are. I should have added it. I tried to hit every industry, and that is one that I didn't hit. But there are uses of advanced control, model predictive control in water and wastewater. That includes some pumping optimization or pump balancing is one of our success stories that you can find. Also, there are some investigation on chemical dosing and some investigations on aeration beds that are supported by both autonomous -- some of the autonomous technologies, but also by model predictive control and Pavilion8.

Perfect. Can it help with inferring Motor Health by GPM and current?

It can be used as a -- so if you look at the causal modeling in the Soft Sensor, you could build a prescriptive analytic -- a predictive analytic and then look for deviations against it and then look for sensitivities about what the most likely thing is. There are other technologies within the autonomous AI capability that we think may be more cost effective. But we -- one of the things that you think about, and I would probably -- with that problem and specifically steer you to the autonomous AI categories, the Logix AI and the soon-to-be Guardian AI capabilities. Pavilion8, we tend to look more at equipment health by avoiding problems. So that constraint management means we can run equipment longer without problems because we're running it and enforcing constraints constantly. So it may be more interesting as and more cost appropriate for that constraint avoidance and equipment kind of sheltering and look at the autonomous AI is likely a more cost-effective answer for Motor Health.

Great. What is a hybrid modeling?

So hybrid modeling is a combination of physical, so thermodynamics, chemical physics, but also empirical modeling. And there are weaknesses or there are strengths of each of those, right? The physical modeling is known to be causal. The empirical modeling is not, it takes more effort to be causal. But you don't always have all the physical models, and it takes a very long time to develop a new physical model. Also, physical models can be computationally intensive and slower than empirical modeling. And finally, they may just be missing part of the things that are part of your data. And so that combination can get you a more accurate representation that extends the physical knowledge. So hybrid is a combination of physical and empirical.

Great. Do you have an opinion on if this should be run on edge, on-prem or in the cloud? Or why should I pick a place?

That's a great question. And there's just -- there's a lot of answers. But because we are changing setpoints in the control system, we tend to run in the edge. The construction could be either in the edge or the cloud, and that's flexible. But you have to then transport the application to the edge to operationalize because most customers are not going to be comfortable with an extremely distant application driving setpoints that protect equipment and operators, and if they fail, may not protect equipment and operators. So that's sort of why the edge is the right place and the place for operationalization today.

Okay. There's another one. Does Pavilion8 work only with ControlLogix or other DCS?

Any DCS. Pavilion8 works with any DCS. We even have deployed applications that are in sights with multiple units, simultaneously controls that we're sending setpoints to different DCS at the same time. But it's DCS-agnostic, and we have many, many of those applications running against other DCS.

Perfect. Let's do this one. Lots of machine learning say that say they optimized. Can't any machine learning do optimization? I think I asked that correctly.

That's a great question. Yes. So I believe no. The reason is because you have to use a causal modeling strategy. So all of that definition and all the time we spent on explaining causal modeling is because I think it's important. If you build an AI that's predictive, and it's driven by correlation, then you know it's not necessarily causal. That's what AI -- that's what predictive analytics are. You can put an optimizer on it. But if it's not correct in the derivative and what you really care about is from where I am now to where I want to go, how far can I get, the derivative is essential, right? The derivative is core. And predictive analytics do not focus on getting the right derivatives. So their optimization tends to be flawed, tends to be wrong. And causal modeling strategy or -- do they process extra data assumptions or sort of inherent in being able to do it? So a lot of the marketplace ignore that and just say we can optimize because we put a solver on our predictive analytic.

Great. It's like we have time for one more. What's the highest value that's been delivered with MPC?

There have been really, really high values. And some of the use cases that I showed were well over $100 million per year. You have to think of the application and kind of guess, but the actual value proposition is going to be opportunity dependent. So how much opportunity you have is going to drive how much value you can achieve, right? You can't achieve 100%. But if you can achieve a good amount of the opportunity, then there's a good value proposition. And then scale of the unit, right? If I made -- if I could improve my margin $5 a barrel and I make 100 barrels a day, then that means one thing. But if I make 500 barrels a day, that means another thing. So the scale of the equipment and then the opportunity, how much opportunity is there from perfect to where you are.

Wonderful. Well, that -- if we did not get to your questions today or if you have any other questions, please make sure that you fill out our survey at the end of this webinar, and someone will be back in touch with you. Michael, I'd like to thank you for your time and information today. It was a wonderful presentation. And we look forward to seeing you all on our next webinar. Have a great rest of your day. Thanks so much.