065 - Understanding mesh sensitivity and model uncertainties with Jason Floyd

[00:00:00] Wojciech Wegrzynski: Hello everybody. And welcome to the fire science show session 65. For this week, I thought that. one tool that almost every fire engineer uses or one of the most used tools is a CFD. Computational fluid dynamics modeling and in particular Fire Dynamics, Simulator, FDS. Of tool that is like, so overwhelmingly. Using the fire industry. It's one of the staples and they did not get much love on the podcast

[00:00:26] Wojciech Wegrzynski: We didn't talk that much. About using FDS. And what does it mean to use it? How to use it properly? I felt it would be a great addition to the podcast series two to discuss this subject. So instead of broadening your horizons, as we usually do on the Fire Science Show.

[00:00:44] Wojciech Wegrzynski: This week, we're doubling down on what we fire engineers do every day. And we're diving very deep into Five Dynamics Simulator. And obviously if you're a north, one of the fire engineers who use it daily or. All very often, or maybe you don't use it at all. Uh, I [00:01:00] guess it's still, it's nice to listen and build your expectations on what you can expect from numerical modeling of fires. And this episode goes also beyond the modeling. So I guess it's going to be interesting.

[00:01:12] Wojciech Wegrzynski: It's almost everyone. My guest today is Dr. Jason Floyd. Jason is with Fire Safety Research institute, UL. At the moment. And he's been with Jensen Hughes for like 20 years. So. A huge chunk of his career done in an engineering company. Jason is on the FDS developer team. And he's very, very knowledgeable about how the software works.

[00:01:37] Wojciech Wegrzynski: And the trigger for this episode. Was actually a discussion. on issue tracker of FDS, there's a Google groups. Um, that's called the issue tracker where people submit their problems with FDS and modeling and the FDS developer team is trying to answer their questions or help

[00:01:55] Wojciech Wegrzynski: and there was an issue where someone was asking why increasing mesh resolution does not improve. [00:02:00] Their calculations. And, I felt it's like super nice point to start because one, what does it really mean to increase the resolution? And what does that do to your solution? And the second, what is a better answer? What's better accuracy. What's a better simulation.

[00:02:18] Wojciech Wegrzynski: If you look at that philosophically, it's not that obvious what simulation is better, which one is closer to reality. What the reality really is, what are the uncertainties of. The measurements in the experiments. What are the inherent, Differences between two seemingly same fire experiments and why you sometimes.

[00:02:39] Wojciech Wegrzynski: Obtain two completely different answers from them. this is something that Jason has actually answered in the issue tracker, and I've tried to pull him further in this talk. So I think it's going to be interesting to take. On. How do we model how simple choices, like the choice of mesh resolution, the choice of [00:03:00] your models, how they affect your simulations?

[00:03:03] Wojciech Wegrzynski: And what the reality of modeling fires and fire experiments really, really is. So without further ado, let's begin the intro and jump into the episode

[00:03:36] Wojciech Wegrzynski: hello, everybody, I'm today here with Dr. Jason Floyd from Fire Safety Research Institute Hey Jason. Great to have you in the podcast. Really happy to have you on the show. This is a, rapid development a few days ago. Um, Bryan Klein from Thunderhead, has posted on LinkedIn, an interesting, post from the issue tracker named find a match, decrease the accuracy, [00:04:00] where someone's reporting on how. Changing the mesh in FDS did not yield better results for their hand calculations.

[00:04:09] Wojciech Wegrzynski: And it's an interesting, issue in this, your tracker, but not very much related. So with, core of, of the S I guess, but, uh, very much related to how are we using the modeling tools that we have. And, I immediately fell in love with this, uh, with this issue, if you can fall in love with an issue, an issue tracker, but it's really interesting, like, because it pinpoints the things that we should be discussing when discussing CFD.

[00:04:36] Wojciech Wegrzynski: So, um, I would love to jump into this, uh, very deep. So tell me why fine measure, decrease the accuracy of this guy.

[00:04:46] Jason Floyd: you know, there's a, there's a number of things potentially that are going on here. , in this particular case, the user was attempting to replicate. I think it was a wood crib experiment, out of a paper. Um, right. So you [00:05:00] always have, a number of difficulties with things like that.

[00:05:03] Jason Floyd: I mean one, or if you, if you take , the same wood cribs, supposedly right in your, in you ignited multiple times, you're not going to get the same heat release rate curve. Um, you know, they're going to be similar, right? You're going to get similar growth, similar peaks, they're not going to be identical.

[00:05:19] Jason Floyd: So, Assuming you had perfect properties and a perfectly tuned FDS model, right? This is, there's no guarantee that your model is going to match that kind of an experiment. Right. I mean, hopefully you're reasonably close, but, if everything is good in your model, but you're definitely not going to replicate it.

[00:05:36] Jason Floyd: So the idea that changing the mess resolution in this particular case sort of pull his results further away from the experiment. Right. But you don't know how that single experiment relates to the mean behavior if you had repeated these tests a number of times. so, so once the one sort of issue there is, right, you don't necessarily know.

[00:05:55] Jason Floyd: Right. What's, what's a good answer, right. Just on basis of a single test.[00:06:00] and then, going into it, then there's also, right. You know, if you're just doing two grid resolutions, you've not really demonstrated whether or not you sort of. Great converge to some, you know, steady answer or if you gone finer again.

[00:06:13] Jason Floyd: Right? I mean, there, there are things that happen at different length scales and the, in the solution. And maybe the answer could have come back up a little bit, you know, if you had done a third grid resolution with this too, um, you know, you've not really established anything about grid, grid, resolution,

[00:06:27] Wojciech Wegrzynski: I think it's interesting to even move back a little bit. Why would one expect that, making, finer mesh would improve the accuracy of the simulation? it's quite interesting to start thinking about the fundamental concepts of CFD, like, I guess we all, to some extent understands the implications of the impact of mesh and the discretization of space and time on the results of CFD.

[00:06:58] Wojciech Wegrzynski: But, I wonder, like [00:07:00] I see it all the time. People would associated smaller meshes with better results, you know, and it's always finding guests as large mesh as applicable. However, on the contrary, I also have the feeling that it's not size that matters. It's To what extent your mash or your numerical domain is fit for the problem at hand, for the scale of the problem you're solving.

[00:07:28] Wojciech Wegrzynski: So I think this is an interesting direction. Of what way do you think are we associate finer mesh with better solution? And is this something we should associate or I

[00:07:39] Jason Floyd: I think a lot of times when people think about CFD third, mind sort of goes to the fluid flow solution. And, they're, you know, this sort of is this expectation that if you start with a really coarse mesh and start making it finer and finer, you should [00:08:00] get better. Resolution of the flow field for, for a set of known inputs right.

[00:08:06] Jason Floyd: To, to the model. so I think this idea comes from, that, um, you know, that if you're not fine enough to sort of resolve the important link skills for your flow, you know, as you get that resolution to those length scales you know, hopefully you should get a better answer, , but in this case, we've got this also this pyrolysis problem going on.

[00:08:27] Jason Floyd: So, one of the resolutions being used here adequate for trying to predict this would create, and we predicting pyrolysis is still very challenging, in any of , these models and just measuring the parameters and how you get, these engineering grid scales to.

[00:08:46] Jason Floyd: Really give you a good flame spread you, not resolving flames. There's a lot of ongoing work there. So, um, there, it's not necessarily clear that in all cases, reducing grid resolution is going to make things better. [00:09:00] Right? I mean, there are, um, within FDS, some correlations related to heat transfer and other things where, you know, those, don't always at this point transition.

[00:09:12] Jason Floyd: Well, from course resolution to extremely fine resolution in all cases. I mean, somebody, something we're working on, but it's not a, that's not a solved problem.

[00:09:22] Wojciech Wegrzynski: I really wonder from your experiences as a developer, how far did people go with mesh sensitivity? Like how small measures have you seen for real, , problems? Has anyone taken FDS two, like sub millimeter scales or.

[00:09:37] Jason Floyd: , I mean, some millimeter scales, , that's only been sort of single playing, like a slot burner, um, some domain like that, where you're only talking about, centimeters of, of, of domain size, on any sort of, you know, engineering type problem, we're dealing with a compartment.

[00:09:58] Jason Floyd: I'm not sure I've [00:10:00] really seen anything in a paper or any presentation, we're criticized as a drop below a little more than a couple of centimeters on, you know, on that scale.

[00:10:11] Wojciech Wegrzynski: it's, it's not the cheap thing to do because it,

[00:10:13] Jason Floyd: Yeah. It gets very expensive.

[00:10:15] Wojciech Wegrzynski: it influences not only the number of elements, but also the timestamp. So there's a big cost associated with, dropping the mesh, which is also the reason why many people would, try to get the, with, as course meshes as they can.

[00:10:29] Wojciech Wegrzynski: In, in many cases, I see, um, mesh resolution studies perform just to have a reason why I have pick the course mesh where, the choice was done even before the mesh study, because that's a, what's your, okay. With your calculation time now to think about the flow problem, like where would you need a fine mesh?

[00:10:48] Wojciech Wegrzynski: I mean, if you, even, if you consider like a system, like a pipe or something, the places where you have high gradients of, of variables like velocity or pressures or sheer stress or [00:11:00] something else is, is the boundary layer. Like just as the other walls in the middle of the pipe can go with a very large, uh, numerical elements.

[00:11:09] Wojciech Wegrzynski: And that will, be good enough. it's, there's not much happening there. You have an average flow. so this tiny meshes are often needed for, for the boundary layer problems, which let's face it. We don't have that many in, in fire. I mean, we're working with fairly low, uh, flow velocities. And, in terms of like compartment fires are burning in buildings.

[00:11:33] Wojciech Wegrzynski: you deal a lot more with it, like in men then, and free plumes and, Some small velocity flows caused by inlets outlets. from your experience like vast experience as, an engineer. did you ever felt, need to really explore this, super fine mesh for a real real project problem.

[00:11:53] Jason Floyd: Um, yeah, I think the cases that I've encountered where you sort of need to pay a lot more attention to the mesh or where your,[00:12:00] where something important you're trying to get deal with is, , flow through an opening as being, you know, an important part of what it is you're trying to do , with the model flow is being purely driven by points, not by some kind of mechanical ventilator system, in which case, , that defines the flow.

[00:12:20] Jason Floyd: But if you're actually trying to predict flow through an opening, through like a doorway or a window or some other small, orifice, if you want to sort of get some reasonable answer in terms of the pressure drop or the flow, you know, you can't just put. couple of grid cells opening, you know, you're gotta be up, sort of double-digit kind of cells.

[00:12:41] Jason Floyd: If you were on, uh, start talking about getting something good, whatever good.

[00:12:46] Wojciech Wegrzynski: Because in this case, again, you have the boundary layer problem. There is an edge which creates a disturbance on the flow,

[00:12:55] Jason Floyd: I need to capture that flow narrowing in you. [00:13:00] You've got to give the model some chance to resolve some kind of vortex on either side of that opening and then the opening and , that minimum number of grid cells you need to, to do that. you know, in cases where, you're really trying to define the flow through small openings, you know, I think, that's where that critical part of your answer, place where you might need to pay attention to, to grid size, No, there aren't necessarily a lot of problems for us where that's very important, most of the time or, you know, people using a lot of the usage of things like, you know, ventilation systems for fire, where they're mechanically driven, very few cells across the opening because whatever it needs to make up, what you're exhausted elsewhere.

[00:13:44] Wojciech Wegrzynski: So in the cases, I mean, that was the topic of Ida, I think 2014 or 15, everyone was dealing with jet fans. And, I think I remember issue tracker being flooded by jet fun, uh, topics like my FDS does not solve jet fans. I don't have good, good results. [00:14:00] And

[00:14:00] Jason Floyd: I definitely, you know, you need fairly decent. I mean, if you want to sort of really resolve that flow in the core of the jet and get something that looks correct in terms of the penetration distance of a jet. you've got to be putting like, a dozen, a couple dozen cells across your jet at least, um, you know, to get that.

[00:14:23] Jason Floyd: And that's, you know, something I've seen in, some work I've reviewed where people doing like a slot diffuser and a ceiling put the one across the slot and it's, it's like, you're no, you're just not, you're not gonna get anything meaningful, out of this.

[00:14:38] Wojciech Wegrzynski: Yeah, to the listeners some context with jet fan systems. That was a, that is a thing. Still is a thing. Uh, it's a system often using car parks and, because of how the system operates. Fairly very difficult, if not impossible to design the system without any CFD because this flows are very [00:15:00] unpredictable.

[00:15:01] Wojciech Wegrzynski: Like they are predictable once you know how to design. Yeah.

[00:15:04] Jason Floyd: Are trying to get out of the model. You know, if you're looking to understand that near-field behavior near the extra, the judge fan, I mean, yeah, you're going to need really good grip. but if, but if all you're trying to capture is sort of some reasonable estimates where the bulk entrainment of air into that jet, that may not require the same resolution if you're, if you're to

[00:15:26] wojciech_wegrzynski: uh,

[00:15:27] Jason Floyd: put, so you gotta, you gotta think about, you know, when you're, when you're the model, what is it you're using it for?

[00:15:32] Jason Floyd: And that your, your doing size may change depending on what answer you're looking to get out of the model in terms of what grit is good enough.

[00:15:40] Wojciech Wegrzynski: Yeah, I think back then there were some changes done to sub grid, scale modeling and FDS and a different, I, I know the different, uh, and smalls were rotating in and out of FDS as some variance of Les, obviously, uh, to, to help like solve this issue [00:16:00] because there was like a large group of FDS users who relied on, this software to, just do their job because that was part of their job to simulate CapEx.

[00:16:10] Wojciech Wegrzynski: And we, we were, we were among them looking a little bit from the side, but it was really interesting timing and a discussion

[00:16:19] Jason Floyd: put a lot of effort into investigating turbulence models. And then there was also just the issue of, how you treat. You know, what happens, that sort of, an edge in your, in your domain. And that's not, it's a tricky problem as to how you formulate that. And we know there's also a lot of effort into trying to make that better.

[00:16:42] Wojciech Wegrzynski: Maybe let's try also to, maybe you can try to explain the, to the listeners, like what's on the level of the software does changing the mesh size due to the solver, , because , you have a bunch of equations you solve [00:17:00] and, uh, they don't ask for the, for this cell size specific, like what does really change in the solver, from the solver perspective, when you, when you change the mesh, is it just, , just the size of the element that, uh, that's being, solved, meaning you, you just gain more resolution, like it's more grained or does physics change in a way.

[00:17:21] Jason Floyd: there really isn't any change in the physics in FDS as you change mess size, it's mostly resolution. We do have an FDS, um, these sort of different modes of operating up the nose, right, where you can be, you know, DNS, Las VLAs and that's the Las, but, but those are user selectable modes of operation and, you know, roughly the sort of intended to correspond to grid sizes, but we're not having, you know, FDS tend to choose on its own, which is appropriate to use.

[00:17:54] wojciech_wegrzynski: Um,

[00:17:55] Jason Floyd: And again, so if you do change that mode, you change some of the underlying physics.[00:18:00] , for example, right in DNS, there's no correlations for heat transfer. You just use the actual temperature grade and at the wall. And, you know, if you pick SDLs, , as your mode of operation, you know, FDS just treats everything as constant specific heats, uh, without using any temperature dependent functions, um, or, you know, things like that, that, that change and other, some of the ways that, massive affection and things like that operate in the model change, you with mode, but as long as you haven't changed the boat, changing the grid size, doesn't really your, other than there are some sort of the, like, for example, you're in some of your.

[00:18:41] Jason Floyd: Initial time steps are like tied to grid size. You know, that that changes what grid size changes, but other than, , some basic things like that, changing grid size doesn't change, the equation is being installed.

[00:18:53] Wojciech Wegrzynski: but still they're the size of the grid is then used to debts or mine. Which part of, of [00:19:00] flow is solved directly with Les on which part goes into sub grid, sub model, right?

[00:19:05] Jason Floyd: Yeah. I mean, the only place where the sub grid models would sort of change, is it, you know, if you're, if you're, if you're in a Gasol it's adjacent to a wall cell, you know, a wall model will be invoked for friction along that wall. And obviously, as you reduce grid size, which cell that happens is gets smaller, you know, yourself wallets smaller, but know the actual, but it's the same equation, , being used in that case.

[00:19:31] Wojciech Wegrzynski: I always thought that, if I go to chorus, like the bigger and bigger chunk of my and flow will be solved with this like one dimensional models instead of Navier Stokes equations. So, and I guess it's probably true when you go to very, very large, grid scales and, yeah. Um, I also, as we are still messing with, with the grid, that there's also one thing I must ask and that's the, the, D star, , criterion and, it's something [00:20:00] that people love to refer to when doing their analysis as the ultimate, condition for which the mesh was chosen.

[00:20:08] Wojciech Wegrzynski: Maybe you can give the listeners a background of how this even came to life. This, D* must be if I'm not wrong between four foreign 16, or is this the rain.

[00:20:18] Jason Floyd: That sort of came out early on in FDS. part of it was there's this effort between , NIST and then Nuclear Regulatory Commission to investigate different fire models for you. And risk assessment. And, you know, so one of the questions that came out of this, you know, the, the documents that came out of this went before the advisory committee on reactor safeguards, which is this panel empowered in the U S to oversee, um, the nuclear regulatory commission.

[00:20:46] Jason Floyd: They had some questions about, well, how can you evaluate sort of goodness of grids? And that was also coming out of other users. so a lot of the problems that were of interest, there were [00:21:00] problems where you're mostly looking at layer development, you know, is a, is a plume impinging on some pieces of equipment, or are you making a layer that's hot enough to damage, you know, cables or something like that.

[00:21:11] Jason Floyd: So this one, this, the star, it's just basically, it's a, it's a scaling. Quantity related to basically, it's with the length scale of the plume. And so the idea is, is that, well, if you have, some reasonable amount of cells across this length scale, that you should get plume entrainment okay.

[00:21:28] Jason Floyd: In your calculation. Right. So if you're pluming treatments, okay. Are you plume? Temperatures should be okay. You're you're there the rate at which your layer drops, you know, shouldn't, shouldn't be reasonable. So this was, so it came out of that and that four to 16 was really came out of that was the range of resolutions that were being used in the validation suite at the time.

[00:21:48] Jason Floyd: , you know, we hear this, this is, this is the grid size range that we had used and, Look, the validation results results are reasonable. So, you know, this w know range was, was noted in the guide as being the range that was [00:22:00] used in the validation guide. I think people took that, as gospel, right?

[00:22:03] Jason Floyd: That as long as you have this, this, these star with DX in this range, that you're always good. and that was definitely not what it was intended to be. And, we're actually in the, you know, the current guides, we remove some of that language. We still know that the D star quantity, but, , definitely try and explain that it's not a one size fits all solution, but this sort of, this, this myth still seems to have around, you know, that this is, this is what you need to do.

[00:22:29] Jason Floyd: And I think, you know, this or that, these star four to 16, if you're really just interested in a smoke layer in a, in a compartment. I think it's a reasonable rule of thumb. I mean, you still need to look at your results and, you know, make sure things seem to make, make sense for you.

[00:22:45] Jason Floyd: Um, if you do it, anything else, you know, if you're trying to predict, you know, radiation to an object really close to far, you're trying to predict flame spread or, anything else other than just really plume behavior. you know, you should be putting some effort into making sure that, you know, you're [00:23:00] using the appropriate grid for, for what you're doing.

[00:23:02] Wojciech Wegrzynski: I must say it is very convenient as a, as an argument in your publishing. And I must say I am guilty of it once I had really a rough review, like it was really annoying and really rough. And, the reviewers were constantly asking like, about the mesh sizes and mesh them, even though we did much sensitivity and the, and everything, but they were very insistent.

[00:23:26] Wojciech Wegrzynski: Is, are we sure this meshes are okay? And so on eventually I said, yeah, we know, check that the, this, it fulfills the star criteria. And so it's good. And they were okay then it's good. So I'm, I'm guilty of it as well, but. I appreciate that you you've mentioned it was made for like layer calculations, plume calculations.

[00:23:45] Wojciech Wegrzynski: There's a completely different fire physics than fully flashover compartments or analyzing a forced flow in, in, in roads, tunnel these are like, like from a different planet. [00:24:00] And now you you've mentioned that you, you should check your mesh. I wonder, what's your idea on how a good sensitivity study should look like?

[00:24:12] Jason Floyd: classic approach to, you know, a grid study is to run, you know, at least three print sizes, where you're lawfully having or something like that, you know, that, that can be difficult to do for your real problem.

[00:24:27] Jason Floyd: Oftentimes, you know, if you've got some tunnel simulation, you have some fire revolving over a period of, tens or hours or something like that. And the tunnel, you know, it might be difficult to afford that finest resolution, which you can, I think in, in most cases, Maybe important points in time and your simulation and, , run a, you know, a steady kind of scenario, and compare, the results of the grid resolutions there, to see, if things are not changing greatly between the, you know, the grid sizes you're looking to use and [00:25:00] your final answer, I mean, there's some discussion in the, in the FDS guide and there are some of these sort of quality metrics that we can output.

[00:25:07] Jason Floyd: Um, And those can be useful to look at. some idea of, it was like this mean turbulence resolution quality, you know, if, if that's, if you're hitting that your region of interest, you know, that, you know, your grids, maybe reasonable.

[00:25:20] Jason Floyd: but yeah, it's, tricky to sort of specify, you know, one size fits all approach,

[00:25:26] Wojciech Wegrzynski: sometimes you also get very confusing results over the grid study. Like your course mesh gives you a value than a finer mesh with give you a lower value than the finest mesh. Would you give, would give you a higher number than the course mesh. Then you are left with a scatter that doesn't converge in any way and it's kind of confusing.

[00:25:44] Wojciech Wegrzynski: what would you do then? Like add many more points or like move back and change something else?

[00:25:50] Wojciech Wegrzynski: happens lot of times.

[00:25:52] Jason Floyd: yeah, one of the things to look at is how much are things changing, in your results. I mean, once you [00:26:00] have to, one of the things you have to understand, right, is that, you know, FTS as a model itself has an uncertainty associated with it. And, you know, the validation guide, we estimate what those uncertainties are for various quantities of interest like bloom temperature and layer, height, and species concentrations, et cetera.

[00:26:21] Jason Floyd: Um, if you do have this case where you sort of, you know, of course grid is a higher value, a medium grid is lower and then the fine grade sort of comes back up. If those are all sort of. No essentially equivalent in terms of, you know, the uncertainty in FDS. it may just be, you know, the sort of that model uncertainty showing up, in the, in the results and you can, pick your medium resolution or something like that, but if you're, you know, if your answers are really changing significantly, then, you know, maybe you can think about, more detailed about what it is, you know, that sure.

[00:26:57] Jason Floyd: You're modeling, with FDS, you know, is [00:27:00] this, especially if it's something you know, where you're dealing with pyrolysis, you mean there, there's just the note. Interacted with in the solid phase. And the gas phase is not a very certain thing right now. So, there, I've definitely noticed, you know, strange things can happen as you, as you move bird size around more so when it's just a specified fire and

[00:27:22] Wojciech Wegrzynski: so when there would be a solid, , assault fuel, and, uh, you would use pyrolysis model to generate the gas fuel from that and burn it, you would expect there's a much more interesting things happening then when it just specify a burner vent and, and just drop the fuel in it and then burn it.

[00:27:39] Jason Floyd: yeah, I've definitely When I tried to play around with, you know, paralysis or use, , predictions, like, I see more of an impact there, with the grids. I think some of it is that yard. You get, we always wind up, where do you measure, your properties of material with some very small scale tests.

[00:27:55] Jason Floyd: people coming right now who can do some kind of like TJ DSE, try and get reactions [00:28:00] and specific heats, and then maybe do a cone test and try and, you know, predict that. And then you go to the full scale. but you know, there's definitely things don't necessarily translate well in scale that way.

[00:28:12] Jason Floyd: So it's almost always seems to be the case that there's some. Tuning needed of your very small scale experimental properties to get something reasonable at a larger scale. But then right now you sort of got properties that are grid dependent. In some sense, we don't have a really good handle on that in all cases.

[00:28:30] Jason Floyd: This is very, I mean, sort of how the radiation model and convictions, mean, those, to have some grid dependencies on them that we have not gotten rid of yet.

[00:28:40] Wojciech Wegrzynski: understood. Yeah. And, when you do, when you try to apply this, small desk per meters into, into real size model, But you didn't know the outcomes of like when you're modeling an experiment, there's always something you can refer to. You've mentioned there's difficulties in that, and maybe we'll [00:29:00] come to that in a second, but, when you try to model something that you do not have an experiment for, and CFD, your only tool, like you want to model a flame spread in a compartment using, solid fuel and the some sort of ignition methods, is there a safe way to proceed that you would recommend?

[00:29:20] Wojciech Wegrzynski: I don't know, maybe starting with a smaller model or working out the literature example it's you know, and, then moving from that, or maybe just, uh, don't touch that because, you have no idea what you're going to get. where are we in, in this kind of modeling?

[00:29:35] Jason Floyd: Yeah. if you've got properties for the same material that you're working with. I think you just got to put a lot of effort into building your calculation opera. You just can't, do a very small scale test, take those things and plop it into your final FDS simulation and run with it.

[00:29:54] Jason Floyd: you know, I think you need to, do the very small scale simulation then, maybe simulate some [00:30:00] simpler model problems, you know, with, with those properties and, look at the results. I mean, there's a lot of times you can look at, and we sort of have some idea, sort of what we kind of expect in terms of burning rates and flame spreads for, you know, a lot of items, um, Oh, I mean, if you go from the small scale to some more intermediate scale and you know, that doesn't look right.

[00:30:22] Jason Floyd: You need to go back to the start, but if, you know, if that looks right, you know, then you can move on to the, next scale. So, you know, I've, I've done simulations looking at, trying to assess, cardboard box commodity and warehouses, um, and there, you're your only fire growth, you know, when you're trying to look at when sprinklers are operating all of that, and that's really hardboard, that's driving that, , in there, there is, , test data at various scales or people have tested just the cardboard or stock of a couple of boxes or, a rack with, two by two by eight array or something like that.

[00:30:58] Jason Floyd: So, you know, there, you do have the [00:31:00] opportunity of, you can try and figure out some properties, go to the single box, a couple of boxes to this rack. And, if it's on each of those steps, you're getting something that's reasonable in terms of the heat release rate measured. I mean, there, you might have feel like you have some confidence that, you're, okay.

[00:31:16] Jason Floyd: For, your larger problem. But what you do need to, probably put a lot more outputs and really look at what the model is doing and making sure that things are okay, right. You're not getting burning rates of six kilograms per meter squared per second, something like that, which, , just isn't plausible, who's never seen other than really exotic fuels perhaps,

[00:31:39] Wojciech Wegrzynski: Yeah. It's , someone once said that bad simulation is when bad simulation is where it doesn't look real, but it's even worse when it, it looks real, but it's completely untrue. , so that's a kind of a danger we face. And now you being associated with UL, I guess there's a whole career to doing modeling plastic cups for the sprinkler as, [00:32:00] as a whole, that's a whole industry that are mine, and I appreciate your approach, , building your own skill set and building your own in a way expertise.

[00:32:10] Wojciech Wegrzynski: And, and, but it also in a way. That's a base of models you're confident with. . And if you, if you had to model a very complicated array of fuels, I see many people try to get into loophole modeling like 10 or 20 different pyrolysis behaviors in the same model, having a different one for an armchair different one for the carpets, for the wardrobe, for every single item like they trying to model them were, uh, if I was faced with such a problem to model a complex, compartment with a complex set of fuels, I think I would settle on some surrogate fueled that.

[00:32:46] Wojciech Wegrzynski: Well, literally like averaged them all because in the end, it's going to be gases that will be burning. And, if I introduced that much complexity into my model, there is no way in the world. I can handle [00:33:00] uncertainty of that.

[00:33:01] Jason Floyd: Yeah, I think be careful of sort of complexity for complexities sake. It's, especially, if you're, the default combustion model right. And FDS, right. You just specify your fuel with, without kinetics that you're just a fast reaction, that point, there's not a lot of benefit to having all these different fuels in a lot of cases, unless maybe, there's extreme differences maybe in there sitting behavior that for some reason was important to count fuel at the start that ignited some high slitting object later.

[00:33:33] Jason Floyd: I mean, that case then you'd have two fuels, right? The low and the high. but if things are kind of similar in terms of their, hazard that you're trying to evaluate it, I don't think there's a lot of benefits in making things too complex and it all just kind of gets as you, as he noted.

[00:33:49] Jason Floyd: Right. But if everything's all fast reactions and all it just averages out together anyway. So at that point, just, make it the average and save yourself the time and effort. But then, then the [00:34:00] other, complexity in, in these kinds of problems is, , No, we don't have, right now, solid objects that don't, disintegrate, right?

[00:34:08] Jason Floyd: I mean, you know, the, the solid cube might use up its fuel and get removed in the calculation. Right. But for, for something like a pollster piece of furniture, we don't have, the phone melting and dripping into a pool they need, there's some books we put in to sort of support.

[00:34:25] Jason Floyd: People trying things like that, but, that's kind of relatively new, but things like, eventually, you know, it sort of collapses into some pile of rubble that, that burns. I mean, those sort of later time, things just aren't there and you know, I'm not that's definitely that kind of stuff is a long way off.

[00:34:40] Jason Floyd: So, for predicting pyrolysis, I think you also need to be careful about those kinds of behaviors, you know, if you're, so if you're, time range of interest is where things are still relatively intact, that you don't worry about, those kinds of issues, you might be better off than, if indeed you've got, you know, your object [00:35:00] molds into a pool, the pool burns, you know, then, maybe you need to think about how you're approaching the problem that you don't want this solid object to burn as a solid view.

[00:35:09] Jason Floyd: Maybe we need to think about some other way of, modeling it.

[00:35:11] Wojciech Wegrzynski: I think we also do not really appreciate the complexity of fire. I mean, this podcast could be called complexities of fire because this is in essence what we are discussing here for more than a year. And it seems we're nowhere close to an answer of how complex a fire is because with every episode is more and more complex.

[00:35:29] Wojciech Wegrzynski: Um, and people often would expect, you know, a single answer from a single experiment, just burn the building ones and understand how does this building behave in the fire? Well, as a simple example of one of the recent experiments, we've done just a normal compartment fire with, grips as a source of, of fire inside the compartment.

[00:35:50] Wojciech Wegrzynski: And we, we had the Crip collapse and it fell out. It produced a slightly different size of fire, which triggered the. And, this [00:36:00] is like a very, like you would not model a Crip collapsing in your, CFD model. And even if you had the perfect representation of this compartment, this crib, all the fire dynamics in it, you could have modeled that and would say, okay, flashover does not the cure in this, in this compartment based on the crude behavior as it was intact.

[00:36:21] Wojciech Wegrzynski: Whereas when it collapsed and suddenly it was like three times the size in space and the air could then train in the middle part of it better. We exceeded this point and flashover happened with the same mass fuel in there, slightly different configuration, but the tiny thing that, that changes the outcome, and we often expect a single answer from a single experiment where it's just a point in your uncertainty matrix that you didn't know.

[00:36:51] Wojciech Wegrzynski: And I think that's, Something that goes back to this initial question of why find a match did not, improve the accuracy of the [00:37:00] simulation. Maybe the simulation was accurate already, as it may have been within the uncertainties of that experiment. what do you think? I mean, these uncertainties add a lot, like what kind of uncertainty you could expect in the simulation and, in an experiment.

[00:37:15] Jason Floyd: Yeah. Well, I mean, for something like that particular experiment, you know, there's a tremendous uncertainty in those material properties that you're specifying with within them, for their action. I mean, things like, right. density is something that's relatively easy to measure, right? I mean, you can measure volume, you can have a scale and you can divide one by there and get density.

[00:37:35] Jason Floyd: And those are relatively. for many materials, right? Those are measurements that you can do, clearly, fairly easily. But when you start getting into all the kinetics are things like the specific key and the, the activation energy and all that for reactions, right? I mean, those are things that we can measure, right?

[00:37:53] Jason Floyd: It's not a tape measure that you put out to measure the activation energy, right? I mean, what we do is we apply some [00:38:00] heating to some sample and measure some temperature response. And then we have a model that we apply to get, these parameters out of it. And so all of those things have uncertainties, right?

[00:38:11] Jason Floyd: I mean that heat flux or heating rate input has some uncertainty that temperature measurement has some uncertainty in that model that we use has a, has a model uncertainty. So, each one of those properties could have. 15, 20, 30% error or something associated with them, perhaps, and, and that can definitely make, a big difference in the outcome of, something like, fire, you know, predicting the, you know, the fire growth, uh, as to whether or not your, you know, your fire grew as quickly enough to get it involved, you know, the curve involved at the right rate.

[00:38:44] Jason Floyd: So I would expect that, just that simulation that he was doing probably had, , 25, 30% kind of uncertainties, Different in terms of the temperatures and the heat release rate, at least in terms of the data that the user had been posted by 15, 20% or [00:39:00] something like that.

[00:39:00] Jason Floyd: So, you know, in that case, if things are different by less than the uncertainty of your experiment, you really can't say one is better. And that's something that I don't, I don't think it's always realized when people just look at, one or a very small number of, of experiments. if you don't really sort of understand what that experiment, the one certainty is it's can be difficult to just do a couple of model runs and go, well, this one's clearly better, which, they're not necessarily better if they're both well within the uncertainty of the test.

[00:39:33] Jason Floyd: And at that point, you can't make that distinguishment. You can say they're different, but you can't say that one is.

[00:39:39] Wojciech Wegrzynski: I also think maybe many times, researchers would confuse, uncertainty related to the capabilities of their measuring systems. W w. I'll certainly do that. That is irrelevant. to the position of the measuring system, especially when, when you're measuring temperatures in flows in, in not forced flows.

[00:39:58] Jason Floyd: Yeah. And we actually [00:40:00] had, somebody here, when I, , started at the FSRI was looking at some furniture tests that at north fork and, , they weren't quite getting, , the FDS predictions to match up with, temperature predictions that were, that were measured.

[00:40:14] Jason Floyd: But in this case there was, , it was a single rake of, TCS. And you go back and you look at the test video that that was taken and you can see that, there was a little bit of lean, to the plume, right. So even though you put. Effort into centering the rake over the burner, FDS the course is going to give you that perfectly straight, fire, if you don't specify, asymmetry in your boundary conditions on the sides.

[00:40:41] Jason Floyd: Once you can show that in FDS, right? You can put up the one rake, you can put, uh, an array of rakes and go show that, well, if you're only 10 centimeters lean at, four meters above the fire, you've changed this temperature now, 25% because you're off the center line now. so that needs to be [00:41:00] considered in, in, you know, in some measurements it's more important, some areas than others, like if you're far from the fire and you're just sort of in a layer.

[00:41:05] Jason Floyd: I mean, I don't think, how accurate that position is, is not as important, unless maybe you're in a, an opening, you know,

[00:41:12] Wojciech Wegrzynski: yeah, I think it's important. Also, if you consider like a ceiling jets

[00:41:16] Track 1: Yeah,

[00:41:17] Wojciech Wegrzynski: and like in the ceiling jet five centimeters lower could be difference of a hundred degrees.

[00:41:22] Track 1: that's a big, big, big difference.

[00:41:24] Wojciech Wegrzynski: And then if you model it with, NSE of D tool, if you have like a 20 centimeter mesh, what you're getting in the end is an average temperature and the yoursel, which half of it may be a flame, half of it is not the flame anymore.

[00:41:35] Wojciech Wegrzynski: So, . I know it's not necessarily a, you are not comparing it to an average temperature measurement over the height of 20 centimeters. You're comparing to a point measurement in a certain position, not directly. So, it may be valuable to also go different, , measures than just temperature.

[00:41:52] Wojciech Wegrzynski: Uh, I'm a huge fan of comparing the, flows, the computational fluid dynamics software. So for like pressures, the flows, that's the [00:42:00] bread and butter. All of the, the reason why we have this tools. So comparing these, the parameters to, in your experiments versus, what you've modeled is a really good indication that whether the model is, is correct or not, How many times you've seen in the issue tracker someone reporting that, there are simulation crashes or, or they get an very bad, results just to figure out that they had like, uh, pressures of hundreds, of thousands of pascals, somewhere in the model because of some stupid error and they, they never, never checked the pressures in their model.

[00:42:34] Wojciech Wegrzynski: That, that, that happens all the time. I guess you, you see a lot dealing with the issue tracker.

[00:42:40] Jason Floyd: Yeah. I mean, we definitely had a lot of issues where you have a, because the way the pressure solver in FDS works, you have a fully enclosed. Volume, definitely were cases where that, let the problem, and Kevin did add in the last release this, uh, if the S now searches on its own, define what we call pressure zones, right?

[00:42:59] Jason Floyd: Sealed [00:43:00] volumes that are somehow connected to an open vent. And each pressure zone has its own treatment of the background pressure. hopefully should go a long way to eliminating many, many of those issues. Um, but yeah, but you do, you do see still people posting cases where, they've put a room or something in the middle of their domain, but they're not opened up any of the, exterior boundaries of the domain.

[00:43:23] Jason Floyd: And, you know, they're like, well, my fire comes out, you have the seal box and you've burned up all the oxygen. and again, that I think that goes to. aren't always taking the time, to learn FDS in there. It's more than just right. jumping into the problem you're trying to solve. Right. it takes, it takes a long time to become a really good FDS user and you should, start with simple problems, playing with inputs and seeing, if you're understanding, when you change this and the simulation there's, this does it, does that make sense?

[00:43:52] Jason Floyd: And, you know, you spent some time doing that before you leap into your massive, full scale. Let's get a [00:44:00] problem.

[00:44:00] Wojciech Wegrzynski: I would qualify FDS under easy to learn how to master, you know, it's very easy to set up your first simulation. And

[00:44:09] Jason Floyd: It's very easy to set up and, I think I mean, it does, there's good and bad for that. The downside is that it also means that you can get people who use it without taking the time.

[00:44:21] Jason Floyd: So.

[00:44:21] Wojciech Wegrzynski: I see much more good in it than the bad things, because the bad things relate to people's character. If you're a person that would just, take a toll, you learn five minutes ago and then use it as an Oracle, then, uh, you should be in, in fortune telling business, not in, in fire research. And, I also think that the, the fact that it's easy to learn and there's not really steep, requirements. It's a great learning tool. And if you learn and appreciate what's happening and appreciate the learning process and really put your heart into it, [00:45:00] trying to understand what's happening, you can really learn a lot. Even now I have, people in my office who would like to learn ANSYS which we're using in the office and they tell them, okay, first, first step for you, if you've never done fire CVD or anything, you should learn FDS

[00:45:16] Wojciech Wegrzynski: yes. I mean that, that's the first step, because you will understand the problem you are trying to solve with the numerical model. And then once you are good with that, then switching the tool into another tool will be a very easy, or maybe not very easy, but it will be an easier experience for you than just, stepping into very complicated tool from the start.

[00:45:39] Wojciech Wegrzynski: Without this knowledge, you get by learning as, and it was my path into engineering, you know, learning FDS first and moving into different, softwares, without any particular reason outside of that, the work environment I entered use the software. So it was not a choice that I didn't like the software anymore.

[00:45:57] Wojciech Wegrzynski: I needed a different one. It was just, a matter of [00:46:00] the, of the environment they fainted,

[00:46:01] Jason Floyd: a good sort of comment you made about FDS and then more complex tools. I think the same thing applies to just using FDS, right? I mean, there are in, many cases, problems that people using FDS for there are, can calculations or zoned model can provide you some insights, maybe, maybe not the detailed answer that you need, but they can sort of give you some insights as to where you think that answer lies.

[00:46:28] Jason Floyd: I think, you know, you see that also sometimes in discussions on the forum and an issues, you know, where people sort of want a simulation, get some strange results, you know, sort of throw their hands up rather than, maybe. Trying to simplify it down to the point where they can use a correlation or zone model and, see is that answer making sense or what's what, what seems to be off?

[00:46:49] Jason Floyd: I think, there's definitely a lot of value in, you know, before you go right into that complex tool to learn some simpler tools so that you sort of build some, [00:47:00] experienced history in your mind as to the kinds of answers you, you expect for different problems.

[00:47:05] Jason Floyd: and that way, as you, as you start getting to more complex things, you have something to rely upon in terms of experiences. Is this answer looking reasonable or do I need to rethink the.

[00:47:16] Wojciech Wegrzynski: I really appreciate this answer because it also answers a question that was posed in the middle of the podcast. What do I do when I face a very complicated problem, which I don't have experiments for? I mean, this is also a good answer that you model it with different tools and at least see the, are you within the scales?

[00:47:34] Wojciech Wegrzynski: You know, are you like 10% of, or thousands of percent of, because that's a very good indication where you should be more or less. So, this goes the experience and, and, uh, I really appreciate the mentioning, other schools in existence because we are living in this very, very weird age where sometimes it's more difficult to convince someone they don't need CFD.

[00:47:55] Wojciech Wegrzynski: Didn't convince them they need,

[00:47:57] Jason Floyd: Yeah. And I think that also, [00:48:00] challenge at times is, especially sort of in the consulting world, right. Client wants that pretty picture. , but you know, there's a, a simple tool that's going to give them as good or maybe even better. , I think there, we gotta do a better job sometimes of educating, , clients, about what these tools can and can't do.

[00:48:21] Jason Floyd: And not just always, do what they ask. You know, sometimes you have to take the time to educate them to get them to understand that no, in this case, all we really need is this hand calculation. You're getting a better answer.

[00:48:34] Wojciech Wegrzynski: Yeah. Now, because we're running out of time and there's one more thing I need to ask you. How's life at the Fire Safety Research Institute, man, that was a, a twist of career for four youth when he is at that Jensen Hughes and then moving into UL FSRI, I guess you're driving there because, uh, I love organization and, uh, when they sold it, the announcement that you were moving out, I thought, oh, what the perfect match.

[00:48:59] Wojciech Wegrzynski: You're [00:49:00] going to have a great time. So are you having a great diamond

[00:49:02] Jason Floyd: This is a really great organization to work with. you know, there's a lot of great, great, great people here too, to work with. there's a lot of really interesting experiments they've done over the years. So does this. The data that's sitting out there, you have to analyze and model and, and understand.

[00:49:20] Jason Floyd: And it's also really nice to sort of be in a position of, being at an Institute where, we're essentially self-funded so that, definitely helps, you know, we're going to eliminate a lot of that burden of sort of the grant and proposal writing. you know, and it just lets us focus on doing, doing the research and then doing, doing the work and, it gives us more time to work with our stakeholders and understand their problems and how we can help them.

[00:49:48] Jason Floyd: Um, yeah, so it's been, it's been great. I feel kind of like, getting the playground, but then

[00:49:52] Wojciech Wegrzynski: Yeah. Nice. And what was the number one? Most exciting thing you're working on now, unless it's top secret.

[00:49:59] Jason Floyd: I guess [00:50:00] one of the more interesting is I, I. Starting to do here, but work with, Dan Madrzykowski they do a lot of work with fire service, they help out, you know, at times where things have gone wrong in the fire ground, and help fire departments understand, are there changes in tactics or decision-making that might, help reduce injuries, while, fight fighting fires.

[00:50:24] Jason Floyd: And so I've been helping him with, it's sort of a mix of, It's modeling it's it's investigation, right? You're trying to piece together photos and witness statements and videos, and then try and couple that with modeling to help understand how their actions have changed the fire environment.

[00:50:41] Jason Floyd: And then trying to understand, we're, where our decision points, where you might've had an impact. So it's really kind of a fun, detective kind of problem, to do.

[00:50:51] Wojciech Wegrzynski: That's the most appreciated research. And, uh, there was a lot of that coming from UL. And I guess there, we, we will see [00:51:00] a lot more, Jason, thank you for your time in here. Mostly thank you for your efforts in, uh, in working with FDS, developing FDS within Doug about HVC solver in the, in this I certainly appreciate all the work that has been put into that part of the, of that model, because it's, it's, it's brilliant. I would love to have, a clone of that in ANSYS. Like seriously, I, I believe that, copycatting is the highest form of admiration, so we may actually go there, but I, uh, it's a brilliant idea that's. Executed even better. So, so I, I appreciate that and I appreciate you, responding to issue tracker, like daily

[00:51:45] Jason Floyd: I think pretty for us developers, to the extent that we can try and interact with the community that helps it makes what we do better. We understand the problems people are having and where, you know, the better we can make, [00:52:00] , the model. So that interaction is very valuable and users, users are great about breaking things, right?

[00:52:05] Jason Floyd: They try

[00:52:06] Jason Floyd: things. The end result, we get a better model.

[00:52:11] Jason Floyd: get an issue, an issue tracking it's like, how did this ever work?

[00:52:14] Wojciech Wegrzynski: without even thinking much, I can give you a list of like 20. Commercial software packages for which I pay a lot of money, which have worse support than the FDS that is provided, for everyone. so I, I highly appreciate that. And, on your hands props to the whole LDS teams, uh, okay.

[00:52:32] Wojciech Wegrzynski: Jason, thank you so much for coming to Fire Science Show it was a great time and, uh, see around.

[00:52:38] Wojciech Wegrzynski: And is Uh, I hope you've enjoyed that. You cannot praise the FDS team enough for maintaining the issue tracker like seriously. From outside, it looks like. A hell of a job to do. It's. The questions being asked constantly, and people are like [00:53:00] endlessly running into problems with how the FDS is used or problems with their modeling.

[00:53:04] Wojciech Wegrzynski: And these guys really tried to give them solid. Science-based fact-based evidence-based answers and I, you cannot appreciate that enough. With Jason we've went through the issues with mesh sensitivities. What does it mean that you increase my sensitivity? How to measure it. I think there's a practical takeaway on how to apply mesh resolutions. I love that we've touched the topic of D star.

[00:53:29] Wojciech Wegrzynski: And debunked, some myths around it. Why having D star between four and 16 is not enough for most of the simulations, maybe enough if you're simulating compartment and looking into But For most cases, that's definitely not sufficient. And. I think. This, this had a lot of practical takeaways like that. So I really hope you've enjoyed this episode. I've learned a bit myself And now I know that I need to get more people from the FDS [00:54:00] in the podcast. It's great to talk with them. And I hope to be able to talk with Jason again. I mean, Jason leads the. Person behind the HVC solver for FDSS. As we, as I brought up in the end of the interview, I have not talked with him about that, but.

[00:54:15] Wojciech Wegrzynski: I really liked that solver, how it can be used in analysis of building. So maybe I should bring him back to the podcast to talk about this particular issue. And for If you're still here with me. I appreciate you For listening to the show. If you have ideas on what questions to ask and to whom to ask

[00:54:35] Wojciech Wegrzynski: I'm here for you. I'm doing it for you. I'm doing this exactly for you. And if you have a question in mind and you, you have a person in mind, let me know. And I'll get it I'm the voice of my audience. And so if you tell me what you want, I'll get the delivered. So now you have bigger impact on what this podcast is. Then you think about it. Send me an email and we'll.

[00:54:58] Wojciech Wegrzynski: Get stuff. arranged Thank you [00:55:00] so much for listening and see you here next Wednesday Bye