Jerad Stack of Flowstate returns to the Pipeliners Podcast to discuss what changes when computational leak detection moves from liquids to compressible fluids such as natural gas, CO2, and ethylene. Russel Treat and Jerad talk through why changes in line inventory make these systems harder to monitor, how physics and machine learning work together in the models, and what an operator should consider before piloting leak detection on a compressible system.

Leak Detection in Compressible Fluids Show Notes, Links, and Insider Terms

• Jerad Stack is the CEO and co-founder of Flowstate. Listen to his previous appearance in Episode 451: Leak Detection is Not Just One Problem.
• Flowstate provides pipeline leak detection software that combines machine learning and physics-based methods to detect anomalies and reduce false alarms. It also sponsors this episode.
• Compressible Fluid is a fluid whose density changes significantly with pressure and temperature. Natural gas is the most common example. In a pipeline, a compressible fluid can be packed into the line instead of flowing straight through, which complicates leak detection.
• Incompressible Fluid is a liquid such as crude oil, refined products, or water. Its density changes very little under pressure, so what goes in one end of the pipeline essentially comes out the other.
• “Spongy” Fluids is Jerad’s term for fluids like CO2, ethylene, and hydrogen that behave partly like a liquid and partly like a gas. Their properties can change sharply with small changes in pressure and temperature.
• Phase Behavior describes how a fluid moves between solid, liquid, and gas states as pressure and temperature change. It is critical for operating and monitoring CO2 and ethylene pipelines.
• Line Pack (Pipeline Inventory) is the amount of product held inside the pipe at a given time. Gas operators often use line pack as storage by raising pressure. A leak detection model has to separate this normal inventory change from an actual loss.
• Inventory Correction is the calculated change in line inventory that is compared against the measured imbalance between inflow and outflow. Jerad describes it as the hardest part of compressible fluid leak detection.
• Computational Pipeline Monitoring (CPM) is software-based leak detection that uses field measurements such as flow, pressure, and temperature to infer whether a leak has occurred.
• API RP 1130 is API’s recommended practice for Computational Pipeline Monitoring for Liquids. It is the primary industry guidance for CPM systems.
• API RP 1175 is API’s recommended practice for Pipeline Leak Detection Program Management. It establishes the concept of a leak detection culture within an operating organization.
• Leak Detection Culture is an organizational commitment, from the control room to field operations, to maintaining and responding to a leak detection program. It is discussed in the episode as a key readiness factor for operators adding leak detection to compressible systems.
• PHMSA is the Pipeline and Hazardous Materials Safety Administration, the U.S. federal regulator for pipeline safety, including leak detection requirements.
• SCADA (Supervisory Control and Data Acquisition) is the control system that collects field data and gives controllers remote monitoring and control of the pipeline. It is the main data source for CPM.
• Real-Time Transient Model (RTTM) is a CPM method that simulates pipeline hydraulics in real time and compares the results with measured values to find discrepancies.
• Physics-Informed Neural Networks are machine learning models that build physical laws into their structure. Flowstate uses this combination of physics and machine learning to model compressibility across many operating regimes.
• Equation of State is a thermodynamic relationship between pressure, temperature, and density. Jerad identifies it as the part of the problem that machine learning alone cannot solve.
• API MPMS Chapter 14 is the chapter of API’s Manual of Petroleum Measurement Standards that covers natural gas fluids measurement. Russel references it when recalling CO2 flow computer calculations.
• Custody Transfer is the point where product ownership changes hands. Custody transfer measurement standards prioritize agreed-upon accuracy so that all parties arrive at the same number.
• Precision vs. Accuracy: Precision means measurements are consistent and tightly grouped. Accuracy means measurements are close to the true value. For leak detection, precision is often more important than accuracy, because the model is looking for deviations from established behavior.
• Hydraulic Transient is a temporary change in pressure and flow caused by events such as pump or compressor starts or valve movements. A good model must recognize these as normal hydraulics and not alarm on them.
• Compressor Station is a facility that raises gas pressure to move product through the pipeline. From a hydraulic perspective, it is the gas equivalent of a pump station.
• LACT (Lease Automatic Custody Transfer) Unit is automated metering equipment used to transfer crude oil from a lease into a gathering system. Russel cites LACTs as an example of how complex gathering system hydraulics can be.
• Operating Envelope is the range of pressures, temperatures, and flow rates a pipeline normally runs within. Model training data needs to cover the full envelope for good leak detection performance.
• Time Alignment and Clock Drift: Timestamps from field devices must line up for leak detection calculations to work. Older devices can drift several seconds a day, and communications latency adds further error.
• Uncertainty Modeling is Flowstate’s approach to handling lost or degraded data. Instead of shutting leak detection off, the model increases its uncertainty about line inventory for that period.
• Stripper Well is a low-volume well near the end of its productive life. Older gas gathering systems fed by stripper wells often have low pressures, limited measurement, and weak communications, which makes computational leak detection difficult.
• Marcellus Shale is a major natural gas producing formation in the Appalachian Basin. Russel uses it as an example of large, well-instrumented, high-pressure gathering systems.
• Aerial Surveys and Gas Sniffers are complementary leak detection technologies, discussed as part of a multi-layered approach for systems where computational methods are not a good fit.
• Withdrawal Test is a controlled removal of product from a pipeline to simulate a leak and verify leak detection sensitivity.
• Alarm Rationalization is the process of reviewing and justifying each alarm so that controllers receive only meaningful, actionable alarms.
• Human-Computer Interaction (HCI) is the study of how people interact with software interfaces. It is applied here to how leak alarms and supporting information are presented to controllers under stress.
• Bridger Pipeline is a crude oil pipeline operator. Jerad cites its collaboration with Flowstate, which included about a decade of data across roughly five thousand miles of pipe.
• Related Flowstate episodes: Episode 455: Leak Alarm, The Most Complex Alarm a Controller Can Get with Angie Schrader and Episode 458: Why Leak Detection is Now an Integration & Cybersecurity Challenge with Nick Dresang.

Leak Detection in Compressible Fluids Full Episode Transcript

Announcer: Welcome to the Pipeliners Podcast, Episode 460, sponsored by Flowstate, providers of pipeline leak detection software built to remove barriers and empower operators. Flowstate helps teams detect anomalies, reduce false alarms, and improve response with practical tools designed for diverse pipeline operations. Learn more at flowstatesolutions.ai.

The Pipeliners Podcast, where professionals, Bubba geeks, and industry insiders share their knowledge and experience about technology, projects, and pipeline operations. And now your host, Russel Treat.

Russel Treat: Thanks for listening to the Pipeliners Podcast. I appreciate you taking the time, and to show the appreciation, we give away a customized YETI tumbler to one listener every episode. This week, our winner is Adam Stavros with SEMCO Energy. Congratulations, Adam. Your YETI is on its way. To learn how you can win this signature prize, stick around till the end of the episode.

This week, we speak to Jerad Stack from Flowstate about leak detection in compressible fluids. Hey, Jerad, welcome back to the Pipeliners Podcast.

Jerad Stack: Thanks, Russel. This is my second time. Well, you know, I’ve been here ten times, but yeah, I’m excited to be back again.

Russel Treat: Yeah. Well, you were on recently, but listen, this is one of those conversations I’ve been looking forward to. It’s something I know a little bit about. Sometimes that’s better, sometimes it’s not. I don’t know, we’ll find out. But we brought you on to talk about compressible fluids and leak detection. So before we dive in, for the listeners that didn’t hear you last time, tell us a little bit about who you are and what you do, and then we’ll dive in.

Jerad Stack: Sure. My name is Jerad Stack. I’m the founder of a leak detection company called Flowstate. I’m a degreed mechanical engineer, which goes back longer than I’d like to admit sometimes, and I’ve kind of spent my whole career in industrial software. Seven years ago was when we founded this company. And, you know, Angie got to talk about alarms last time, so…

Russel Treat: Right, right.

Jerad Stack: So we’re not talking about that today. That’s her expertise, not mine.

Russel Treat: Yeah. Well, compressible fluids is an interesting subject. So I want to dive in and just talk about, first, what is a compressible fluid, and what makes them difficult and challenging from a leak detection perspective?

Jerad Stack: Yeah, absolutely. So, you know, compressible fluid, you and I could go back to our junior year of college textbooks and look them up, right? But in the scope of leak detection, the scope of energy pipelines, we have liquids, you know, crude oil, produced products, gas, diesel, or water, that are essentially incompressible fluids. They compress a little bit, right? But there’s a whole other class, the biggest class being natural gas, where the compressibility of the fluid in the pipeline really matters.

And it’s not just natural gas. There’s a whole bunch of other spongy liquids like ethylene, CO2, hydrogen, those kinds of things, where it really matters. The density of that fluid can change, right? So you can compress them inside the pipeline, and they compress, as opposed to an incompressible fluid that just pushes on down.

Russel Treat: Yeah. I think some people who have experience with some of these things like the ethylene and the CO2 and so forth, they’ll kind of understand the concept, right? Because to work in those fluids, you have to understand the phase behavior. So if we go back to our third year of mechanical engineering and fluid dynamics, and we start talking about phase behavior, if you remember, there’s a phase curve for every fluid where on one part of the curve it’s gas, on another part of the curve it’s liquid, and on another part of the curve it’s solid.

The easiest idea there is that as pressure changes, the temperature at which you become a solid or a liquid changes. So you have this transition zone that’s a function of pressure and temperature where things change. I’ll save your analytics right now: if you’re listening to the podcast, you don’t need to drop out. We’re not going to talk about phase change for the next hour. I’m sure people have some PTSD about it. But yeah, those of us that are educated there remember that, and that matters.

And then there’s another kind of what I would call strange-behaving fluids, where there’s this area kind of in the center. You think of a water curve, it’s pretty simple. There’s a point right around atmospheric pressure at thirty-two degrees where I go from liquid to ice. And then there’s another point where I go from liquid to vapor. It’s pretty simple. But some of these spongy things have kind of an envelope, if you will, where depending on the combination of pressure and temperature, it’s kind of a liquid and it’s kind of a gas, and it’s got very weird behavior. And for a very small change in pressure and temperature, a very big change in behavior occurs. And you’d characterize that behavior primarily as density.

Jerad Stack: Yep. And we even have that if you get outside the phase change, right? For just a gas, natural gas for example, it’ll compress a lot while staying in the same phase.

Russel Treat: Right, right, right.

Jerad Stack: Ethylene does the same, but things get really wacky with ethylene. But compared to a crude oil, all that physics starts to matter when you start talking about compressible fluid.

Russel Treat: Yeah. So historically, if you talk about leak detection, most of the leak detection done with SCADA systems and using CPM, computational pipeline monitoring, has focused on liquids, so-called incompressible liquids. What do you have to do in order to take that same approach and apply it to these compressible fluids?

Jerad Stack: Yeah. So in its easiest version, leak detection for an incompressible fluid is kind of everything that goes in comes out, right? Things get complicated with some transients and stuff that I think we’ve talked about on some other episodes. But really simple: if you have a really simple pipeline, if you push on one end, it comes out the other end, right?

With a compressible fluid, that doesn’t necessarily happen, right? You can increase the pressure, and instead of it coming out the other end, it can just compress inside the pipeline. So you really have to take into account that physics, which in our pipeline world is the inventory, right? You could have more BTUs in the pipe under the same volume. You can just squeeze it a little bit.

So that actually makes leaks harder, because it’s not necessarily true that everything that goes in comes out. Instead of having it come out one end, you can compress it in the middle. And pipeline operators use that fluid property to their advantage. As opposed to crude oil, where we have tanks on the other end and that’s where we store excess crude, a gas pipeline operator might store excess gas in the pipe by just compressing it more, right?

But that means you have to take all that into account. You have to understand, are they just compressing the fluid, or is there a leak or something else going wrong? So yeah, it’s that whole additional complexity that you have to take into account.

Russel Treat: Yeah. So if you think about natural gas, its compressibility behavior is fairly uniform, right?

Jerad Stack: Thankfully.

Russel Treat: Yeah. Well, and people that work in natural gas pipelines will be familiar with, “Well, I know that this pipeline is eighteen inches in diameter and it’s twenty miles long, and at this pressure and this composition, it will hold this many BTUs.” That’s a fairly simple calculation. When you start getting into some of these other things… I would say that gas is compressible and liquids are not. And then these other things like CO2 and such, they’re spongy, kind of compressible and kind of not. And that calculation starts becoming a whole lot more complicated.

Jerad Stack: Correct.

Russel Treat: So what comes up for me, just knowing what I know about leak detection, is all of a sudden the computer processing required to get a number out gets a lot more complicated, because you’ve got to look at a lot more things. You can’t make simplifying assumptions.

Jerad Stack: Yeah, you’re absolutely right. If you remember our conversation a month or so ago, at Flowstate we use a lot of machine learning to understand the history of a pipeline and the current version of the pipeline. It’s one of the things we do in detection. When you get to these compressible fluids, that is where that hits limits, and you have to have some physics in there to be able to understand it. So that’s just a whole other dynamic that you have to have. And it’s harder. The math is harder.

Now, thankfully, computers are way faster now too, right? Computation is not free, but it’s dang close these days, right? But building the algorithms is not free. You have to take a…

Russel Treat: Okay, I want you to unpack that a little bit, because I think I understand what you mean, but I’m not sure I do. Let me make sure I understand. So the first comment was that computation has basically become free. Very, very low cost to run a computation, which is not the way it was when you and I started.

Jerad Stack: Computation was very expensive.

Russel Treat: We cared a lot about computation.

Jerad Stack: Yeah.

Russel Treat: Right. We don’t really so much anymore. However, you said the algorithm is hard or expensive. What do you mean by that?

Jerad Stack: Well, every pipeline is still unique. The pressures and the flow rates and the temperatures that a CO2 pipeline in Texas operates with are probably way different than one in the Rocky Mountains, because you have seasonal effects and a whole bunch of other stuff that goes into it. Since you’re kind of right around that phase change, you have to get a lot of data to understand it, and you have to have a lot of physics to understand it too.

Russel Treat: Yeah. What I remember about doing CO2 flow calculations and building a flow computer, oh gosh, probably twenty years ago now, when we were doing that, we were running MPMS 14. And one of the things about MPMS 14 is you could specify the whole range of hydrocarbons that could be in the CO2, and all those affected the algorithm. But if I could do that calculation with six factors versus twenty-one, then I could run that calculation much, much faster and a lot more stable, right?

Jerad Stack: Right. Yes. So you want to get it as simple as you can while it still captures the physics you’re trying to get.

Russel Treat: Yeah. But the trick was, because we were doing that in firmware, which is kind of like doing it on older computers, you had to have a way for the algorithm to learn what data it was and wasn’t going to use, and then it could speed up. Which meant the first run took a long time, and then it started running quick.

So when you talk about the algorithm and how you implement it, the other thing we talked about is you kind of build, you talked about a physics-informed model, physics-informed neural networks.

Jerad Stack: Yeah. Right. So just think of it as we use some machine learning and physics at the same time. We need physics to be able to do that compressibility and understand the inventory of a pipeline. But we can still use that machine learning to get really good relationships from operating data and describe how the pipeline operates through tons and tons of different operating regimes.

Russel Treat: Yeah, that makes a lot of sense. So one of the things in prepping for this that we talked about is the idea of interpretation, right? How you look at the data, normal data versus what would indicate a leak. At the simplest level, it’s volume in, volume out, and is there a difference? But now we’ve got to say, well, it’s volume in, volume out, change in inventory, and is there a difference we can’t account for? So that adds a level of complexity, right?

Jerad Stack: And you just nailed that. That is exactly what we’re doing today. And you glossed over the hardest part of it, which is that change in inventory piece.

Russel Treat: Right. Yeah.

Jerad Stack: That’s kind of the trick. But you’re exactly right.

Russel Treat: Well, does that mean that because you’re doing the change in inventory thing, there are now more things that could occur hydraulically that might look like a leak?

Jerad Stack: Yep. You’re exactly right.

Russel Treat: So how do you address that reality?

Jerad Stack: So that’s where we use physics, right? That’s where you have to understand, are the pressures similar? Temperature matters too in these pipelines. Can we actually calculate what that inventory change should be, that inventory correction?

Russel Treat: Right, right.

Jerad Stack: And then compare it to the actual measured inventory correction. And if there’s a delta there, then you send it through a second set of algorithms trying to figure out, hey, is there also a leak signature? Or is there another reason to think that it might be a leak?

Russel Treat: Okay, so I know in the last episode we talked a little bit about neural networks, but what you’re really doing here is you’re taking what you’ve been doing historically, and you’re making the neural network more complicated. You’re adding more nodes in the network.

Jerad Stack: Yeah. More intelligent as opposed to more complicated is probably what I’d like to say. But yeah, they’re both.

Russel Treat: I’m not sure that’s not a distinction without a difference. So now we have to have… oh, man, I’m just thinking about what this would look like in a CO2 pipeline, right? There’s a bunch of things that start to matter more than they used to matter. The complexity is increasing exponentially.

Jerad Stack: I don’t know if it’s exponentially, but yeah, I mean, these are three times more complicated than a liquid pipeline.

Russel Treat: Okay.

Jerad Stack: Well, and you need more data, right? So temperature matters. Temperature often doesn’t matter on a liquids pipeline, or if it does, it’s kind of a second- or third-order problem. Temperature really matters on a compressible fluids pipeline.

Russel Treat: The quality of your data matters more, right?

Jerad Stack: Yeah, yeah.

Russel Treat: You need better precision on temperature than you might otherwise need.

Jerad Stack: You need better precision on temperature, and on pressure for sure.

Russel Treat: You need better time matching.

Jerad Stack: Yeah.

Russel Treat: Because a little bit of a time difference can make a whole lot of difference.

Jerad Stack: Yeah. You need better communications, right? Since your pressures and your temperatures matter, if you have a midstream pressure sensor, you need to have good communications with that as well as with the ends of your pipeline, because that all starts to matter.

We’ve had to go through not only the inventory correction piece and build all those algorithms, but there’s a whole uncertainty piece to it. Hey, if you do lose some data, these things can’t just completely turn off, right? You still have to have leak detection. So we’ve had to build in a whole uncertainty piece where, if you lose a little bit of data for a time period, your uncertainty about what that inventory is increases, as opposed to not knowing anymore.

Russel Treat: Yeah. So here’s the other question that comes up for me as I’m thinking about this. How critical is the normal operating envelope of the pipeline to being able to do a good job of compressible fluid leak detection? Like, I’ve got to run it in this temperature envelope, this pressure envelope, this flow rate envelope.

Jerad Stack: Well, yeah. Said a different way, the training data that we put into our model needs to include the whole envelope, right?

Russel Treat: Right.

Jerad Stack: So you need to put it through its paces, or most of its paces, right? You can kind of extrapolate a little bit outside those paces, but yeah, you need more good-quality data to build a good leak detection model for one of these pipelines.

Russel Treat: Yeah. Oh, man, my mind just goes spinning when I have these conversations, because I really want to dive into the mathematics of it and see what you’re doing. A lot of this is also what math you’re doing and how you’re sequencing that math, right?

So at the lowest level, you’re reading in pressures, temperatures, and flows, and you’re calculating pipe inventory and pipe inventory difference over a unit of time. That’s kind of the lowest level. Then there’s a whole other level of other things you’re doing, like, well, I’m missing this data, or I’m a little bit outside my normal operating envelope, and am I okay with that? And there’s a whole bunch of other things that…

Jerad Stack: And actually, the sequence you just went through is almost exactly what we do.

Russel Treat: Oh, okay.

Jerad Stack: So you kind of got that sequence correct.

Russel Treat: Yeah. Well, that makes sense to me, because what most systems are doing, what most real-time transient models are doing, is they’re just doing that first bit. And then there’s a lot that goes into, well, is that an alarm or not? Which is an analytical process typically done by humans.

Well, now what you guys are doing is building a mathematical, algorithmic process leveraging AI to do that additional computation to eliminate… well, not to eliminate. It’s really to identify that this is an expected hydraulic issue given what’s going on. It is normal hydraulics in the pipeline.

Jerad Stack: Yeah. And the goal is to eliminate the alarm you might have gotten during that inventory change.

Russel Treat: Right. And that’s fundamentally why we haven’t done compressible fluids, because we were just doing that first level of calculations. And because of the number of variations that can occur, and because of the change in inventory, it was very difficult to get to “Is that a leak or not?” And now, because we can do a lot more computation in a lot less time, we can do more and say, well, there was a pump start over here, and that pump start created this hydraulic transient, and we see that transient, and that’s not an alarm.

Jerad Stack: Yep. You’re exactly right.

Russel Treat: So how big is the inventory of things that could cause you to have a hydraulic transient that you would need to be able to identify and not alarm on?

Jerad Stack: So depending on how you look at it, it’s either huge or not that big. These are still pipelines, right? We’ve got pumps. We’ve got control valves. They’re not whole new mystery monsters that none of us have ever looked at. Turning pumps on and off, having control valves open and close, all of that is the same as for a liquid pipeline, right? You do have this concept of a compressor station, which we don’t really have with liquids, but that’s essentially just a pump from a hydraulic perspective.

So the list of those things is not huge. The magnitude of how those things change is big. And it’s more like what you’d see on a big gathering pipeline, right? You’ve got forty different inputs and outputs, and they could all be on or off at the same time, at different values. On the liquid side, that’s something we’ve been doing for a long time, and it’s kind of similar to what you see here. Maybe there are five different pumps at a compressor station. Sometimes you have one on, sometimes you have two on, sometimes you have three on, right? So the combination of different changes can be big, which is really hard for a human to build a model around. For a computer to build that model, that’s kind of what they’re good at.

Russel Treat: Yeah. Well, I guess the point you’re making is that the complexity related to a crude gathering system, where you’ve got a bunch of LACTs at the wellhead that are coming off and on and surging and falling off and doing all kinds of weird things hydraulically, you really can’t get to a more dynamic hydraulic situation than that. It’s not a lot different in terms of the scope of what you have to look at with these other things. It’s just that now, instead of looking at flows coming off and on, you’re looking at all the things hydraulically that could impact that internal line inventory.

Jerad Stack: Yeah, exactly.

Russel Treat: So this is all making sense to me. What’s coming up for me now is, how critical is time alignment in this? Because in my experience, as you’re trying to tune any kind of leak detection or any kind of line balance or any of that kind of stuff, timing is critical. And the clocks on all these devices we use in the field are not super accurate. They drift. So how do you deal with that algorithmically, or how do you handle that?

Jerad Stack: So the hardware guys are getting a lot better on the time thing, for what it’s worth, right? Even compared to five years ago when we started this, the PLCs are getting better. The SCADA systems are getting better. There are rooms full of people that are working on that problem, but it totally still exists. The clocks still drift occasionally on those things, and we have algorithms that are kind of watching some of that stuff. But it’s tough. Sometimes they drift in a precise manner, not an accurate manner, so the machine learning can kind of figure that out. But oftentimes, our models are only as good as the data.

Russel Treat: Yeah. Well, all this kind of stuff gets really complicated when you get down into the weeds of it. Just the simple thing of, well, I’m going to put something in the SCADA system and write SCADA time back to the clock in the local device. Even something like that, if you’re doing it once a day and catching small changes, it still doesn’t get you as tight as you might want to be.

Jerad Stack: Yeah.

Russel Treat: Because you’ve got all these other issues around comm timing and latency and everything else. It can easily pick up full seconds, much less microseconds.

Jerad Stack: Yeah. And I was going to say, we’re not really going down to the microsecond level to look at these tiny little changes. We actually have some other projects that are going down to the millisecond level. But here we’re looking at orders of magnitude in seconds, one, five, ten second kinds of data points. So that certainly makes it a lot easier. It makes it easier. But some of these old devices can drift that much, right?

Russel Treat: Oh yeah. A lot of the older ones could drift several seconds a day easily.

Jerad Stack: Yeah. But it’s rare.

Russel Treat: Yeah. Well, it’s rare, but one device in a network of one hundred can cause you a lot of grief on this kind of stuff.

Well, so where does machine learning not really help you? We’ve obviously been talking about things where it does help you. Where does it not really help you?

Jerad Stack: So that inventory correction factor is probably the biggest one. We couldn’t figure out a way to do it purely with machine learning. So you’ve got to use some kind of equation of state to be able to figure that out. And they’re pretty well… people have been doing fluids for…

Russel Treat: Oh yeah.

Jerad Stack: So they’re pretty well documented. And we try really hard to use the ones that make the most sense and are easier for people to understand, right? So if there’s an engineer at a customer that does need to figure out what’s going on under the hood, they can understand that.

Russel Treat: So this is going to be kind of out of left field. I grew up in measurement, and all of the measurement standards are written to an equation of state. But there are equations of state out there that are not custody transfer. And really, all that custody transfer means is that some group of people across the industry have agreed to this approach, and we’ve all agreed to use it, so we all get to the same number. It might not be the best number, but it gets us to the same number.

In a situation like what we’re talking about here, it might be that some non-custody equation of state gives you better performance. Are you guys looking at that sort of thing?

Jerad Stack: I would be out over my skis even answering that.

Russel Treat: Oh, okay. Well, good.

Jerad Stack: I am not a measurement guy.

Russel Treat: All right. Well, I feel better about myself, because last time I felt like I was over my skis, so I’m glad I returned the favor a little bit.

Jerad Stack: Got me on that one.

Russel Treat: Yeah. Well, no, but I understand the issue, and it’s kind of interesting. I don’t know enough about what you’re doing, but it’s an interesting question: how important is accuracy in calculating the actual volume difference? And is that more about the algorithm you’re using, or is it more about how you’re handling the data that’s being fed to the algorithm?

Jerad Stack: The answer is probably yes, both. And, you know, precision is almost more important than accuracy from our standpoint, right?

Russel Treat: We should define precision and accuracy for people that don’t know.

Jerad Stack: So as long as the measurements are consistent, right? They might be a little bit off. The custody transfer people want the accuracy, right? If I’m buying X barrels, I want to buy X barrels. But from a leak detection standpoint, if we look at your last six months of data, then looking forward, as long as I’m getting the same results I was getting, we’re able to find deviations from that.

Russel Treat: Yeah. The difference is a small grouping of arrows, but not necessarily on the bullseye, versus everything’s on the target and equally spaced around the bullseye. So that’s the difference. Precision is they’re all tightly grouped, and accuracy is they surround the bullseye.

Jerad Stack: Yeah. Interesting. You know, way back in the liquids days, something we learned the first year doing this, we were building purely machine learning models back then. And our customer would be like, “Oh yeah, we went and re-zeroed the meter.” And we’re like, “Oh, you shouldn’t have done that.”

Russel Treat: Yeah. The contracts, operations… the contracts require that, one hundred percent.

Jerad Stack: Yes. But that causes you a lot of trouble. The algorithm needs to know that, oh, you re-zeroed the meter. Start again. We had to put in a whole process and procedure for that.

Russel Treat: Yeah. Devil’s in the details, man.

Jerad Stack: Sure.

Russel Treat: So if an operator is listening to this conversation and they’re kind of intrigued about the approach, how would you go about it, or what would you say would be the idea to do a leak detection pilot on a compressible fluid?

Jerad Stack: Yeah. I think our technology for compressible is not as mature as our liquid technology is, right? So there’s some stuff that we’re still working on. A gas gathering system, for example, is super hard. In our experience, a lot of the communications aren’t great for that, and the measurements aren’t great for that, and we’re kind of having to use some other technologies to be able to tackle that. So we want to have an initial conversation of, let’s just talk about your system. And fifty percent of the time we might be like, “I’m not sure we could do that,” just to be frank.

Russel Treat: Yeah. No, you’ve been around this enough to know. When you talk about gas gathering, particularly in older infrastructure where it’s coming off stripper wells, and there’s not a lot of flow, and the pressures are low, and there’s not a lot of high-caliber measurement or high-caliber communications, that starts getting really difficult. If you’re talking about a gathering system in the Marcellus that’s eighteen-inch pipe running at sixteen hundred PSI, the measurement’s going to be tight because of the amount of gas they’re moving, and that’s probably a different conversation.

Jerad Stack: You’re exactly right. If you’re talking about one of those old ones, a computational, data-based system is probably not the right answer for them, right? It’s probably more of an aerial survey or something like that, or putting some sniffers at your stations. There’s a bunch of other technology out there, right?

Russel Treat: Sure.

Jerad Stack: They’re trying to tackle this. This is a multi-layered approach. We’re all trying to solve the same problem.

Russel Treat: Yeah.

Jerad Stack: But yeah, so we do an initial assessment. Do you have some pretty good measurement? Do you have some pretty good data? Do you have some pretty good communications on the system? And make sure that we have not just flow rates, but we need some pressures, we need temperatures, and we need a good history of their operational data, which most people now have. It didn’t used to be the case that everybody had all their data in a historian or whatever, right? But most people do now. So we can get access to that.

And we try to do exactly that. Let’s go look at some of your historical data. We put it through the first few passes of our algorithm, and then let them know, here’s kind of the sensitivities or false alarm rates that you would probably see from that.

For those spongy liquids, most of those are managed and operated by liquids people, right? So they’re kind of used to using CPM. We speak their language. Pure gas operators or pure gas control rooms, maybe they’re not used to using CPM, and the regulations don’t necessarily require them to. That’s probably a whole other conversation.

Russel Treat: Yeah.

Jerad Stack: But there’s probably some education there around whether they’re ready to be able to have leak detection as well.

Russel Treat: Yeah. Well, I think you make a really important point here, Jerad. Implementing leak detection on a line that you don’t have it on because it’s compressible, but you have it on other things, so you have capability around the care and feeding of a leak detection program, is very different than “I don’t have any resources, I don’t have any skills, I don’t have any experience.” That’s a whole different thing. The field operations, when I call and say we’ve got a leak detection alarm, they’re going to know what we’re talking about. There’s a whole, I think PHMSA or maybe 1130 calls it, a culture of leak detection.

Jerad Stack: Mhm. There are organizations like that, and that’s what we’re seeing. We’re seeing a lot more uplift on this product with the compressible liquids than we are on the gas side, because they’re just way more mature from a computational leak detection standpoint.

Russel Treat: Yeah. And they’re probably closer to getting the regulations landed on them too, because of some incidents that have occurred with those kinds of fluids that have had pretty significant consequences.

Jerad Stack: But that said, even the pure gas operators, and yeah, they’re probably years away from regulations, they still want to do the right thing.

Russel Treat: Yeah, no, they do. But the point being that they have to figure out the economic justification, and they’ve got to operationalize whatever technology. And operationalizing a leak detection technology requires a lot of work around tightening up your communications and improving your instrumentation.

Jerad Stack: Yeah. And all the stuff that you’re an expert in and not me, the alarm rationalization piece.

Russel Treat: Oh yeah.

Jerad Stack: There’s just a lot that goes along with having that.

Russel Treat: And yeah, that’s the big gap. They’re still figuring that out.

Jerad Stack: Yeah. The big pipe gas guys have kind of got the alarm rationalization stuff. But you’re absolutely right.

Russel Treat: Well, look, we’ve now done four conversations with Flowstate about leak detection. We kind of started just talking about what it is and how it works and what you guys do and how it’s different. We talked about alarming and how you deal with leak alarms and the quantity and quality of leak alarms. We talked about security. And now we’re talking about fluids we haven’t historically done this kind of leak detection on. For somebody who’s listening to all this, what would you say as a final comment? How would you put a bow on the conversations?

Jerad Stack: You know, I think we are privileged at Flowstate in a couple of ways. We get to tackle this problem in 2026, right? There is so much cool technology out there that we get to leverage to go tackle these problems. I think you met with Nick, and some of the cybersecurity struggles that our industry has, they’d be really, really hard if we had to use 2005 technologies to tackle those, right? And we get the privilege of being able to do that. So we get to be the modern answer to all of that. And that’s where we get to tackle things like compressible fluids, because we get to be the modern answer to that.

And we also get to do this in an environment where the whole industry is kind of on the same team. That’s one thing I really enjoy about the leak detection side. Nobody’s like, “My leak detection is a competitive advantage to me,” right? So they’re all really happy to help. Like our relationship with the guys over at Bridger Pipeline: a ton of data, a decade of data, five thousand miles of pipe, and they’re like, “We just don’t want to have any leaks. Can you help us make this better?” And we get the same when we get into other customers on the compressible side, where we’re able to go test this out, do withdrawal tests and stuff like that. So I think it’s a super exciting time to be in leak detection. And maybe that’s really nerdy of me to say.

Russel Treat: Maybe, but you know, what you’re saying, Jerad, is so very true. The fact that data has become virtually free, and the fact that computation is virtually free, means we have the ability to do these more complicated, or probably a better way to say it is richer, algorithms. Richer in terms of we can put these things together in ways that we just couldn’t have done even ten years ago, just because of what computers are capable of. And it’s a game changer, and it should help us improve safety performance. And that’s really ultimately what we’re all trying to do.

Jerad Stack: And there’s been decades of research in human-computer interaction and stuff like that, and I know you talked to Angie about that a lot. But it’s not just a red light anymore, right? Boom. There’s a lot of research that goes into how we put the right information in front of controllers at the right time, in this really stressful moment where they’re like, “I have a leak alarm.” How do you give them something that they can look at and analyze?

Russel Treat: Yeah.

Jerad Stack: And we’ve been able to leverage a lot of that in our work to get user interfaces that are modern and understandable, and to get really good information to the right people at the right time.

Russel Treat: Well, look, this has been a great conversation. I’m so very appreciative that you dedicated the time of some of your best and brightest on your team to have these conversations. I’m certainly getting some feedback from a few people in the industry I’ve talked to that they’ve really found these conversations valuable and helpful. So thank you for this.

Jerad Stack: Yeah, thank you, Russel. It’s been really fun for us, for the whole team here. It’s been great.

Russel Treat: Great. Well, we will definitely have you guys back. So you need to invent something new and come tell us about it.

Jerad Stack: Deal. I’ll take you up on that.

Russel Treat: I hope you enjoyed this week’s episode of the Pipeliners Podcast and our conversation with Jerad. Just a reminder, you should register to win our customized Pipeliners Podcast YETI tumbler. Simply visit PipelinePodcastNetwork.com/Win and enter yourself in the drawing. If you have ideas, questions, or topics you’d be interested in, or if you’d like to be a guest, please let me know on the Contact Us page at PipelinePodcastNetwork.com or reach out to me on LinkedIn. Thanks for listening. I’ll talk to you next week.