00:00:00:00 - 00:00:40:22 Unknown Our next speaker will be the title of the presentation will be on 4D Electrical Resistivity Tomography Monitoring of Vadose Zone Soil Flushing at the Hanford 100-K Area Reactor Facility: Machine-Learning Based Assessment. And the speaker is Dr. Tim Johnson. Dr. Tim Johnson is a computational geophysicist at PNNL, is internationally recognized for his technical contribution to subsurface geophysical imaging and process monitoring. 00:00:40:24 - 00:01:23:07 Unknown His primary research is focused on characterizing and monitoring subsurface properties and processes, using autonomous geophysical measurements with an emphasis on electrical methods. Dr. Johnson is the original developer of the award winning E4D software, parallel electrical resistivity tomography and spectral induced polarization inversion code designed specifically for 4D subsurface imaging. Dr. Tim Johnson Thank you. 00:01:23:09 - 00:02:22:14 Unknown I'm going to share my screen here. Okay. Hopefully you can hear me and see my slides. So I'm going to talk about the relay application actually at Hanford, where they use soil flushing to remediate some hexavalent chromium contamination in the bay. So I've listed all my coauthors here. This is, of course, an effort among many people from different companies, including PNNL and CPC Co, the cleanup company, and how much of this was also funded by by the Richland operations office of DOE. 00:02:22:16 - 00:02:55:23 Unknown So I thought I'd start to give just a basic background on what electrical resistivity tomography is. and when I do this, I like to start with the medical imaging analogy because people are familiar with that. And there actually is a medical imaging analysis analogy called Electrical Impedance Tomography and Impedance Tomography. As you can see here on my pointer here, here on the left is are basically electrodes, metal electrodes that are attached to the subject to make electrical contact. 00:02:55:23 - 00:03:26:15 Unknown They're connected to wires and to take measurements. You basically inject currents between any two electrode pairs and then you measure the corresponding voltages that develop across many of the other electrode pairs. You take many measurements like that of different offsets and you turn that off and you process the data. And then fairly computational intensive inversion algorithm, that's the tomography part and you get an image. 00:03:26:15 - 00:03:48:07 Unknown So here you see an image of you can see the heart, you might see parts of the valves there, you can see the lungs and a little bit of the spine down there in the bottom. And so useful technology. But what you might gather from this result, as you might expect from normal other types of medical imaging. 00:03:48:09 - 00:04:11:12 Unknown And the reason for that is the technology is very simple, but it's something that we can use in the subsurface because it is simple. We're just using electrodes and metal wires to connect them. So I had an example at the Hanford 300 area of electrical resistivity tomography, where we have a look very closely spaced in this case along the surface. 00:04:11:12 - 00:04:48:17 Unknown Each one of those orange dots that you see is an electrode. And just like the medical imaging case, we do some data processing and we get a tomographic image of the electrical conductivity of the ground. So here you see some stratigraphy underneath, underneath one of these lines represented as the electrical conductivity of the subsurface. So later, just as a preview and we'll talk about how we're replacing this data processing step with machine learning and what is output is not an image in terms of electrical conductivity, but it's the hydro geologic properties of the subsurface. 00:04:48:19 - 00:05:17:16 Unknown And we're doing that with time lapse imaging data. Time lapse image data is a little bit different than what we're showing here because we take images like this over time and process them and then we subtract out the baseline image. And what is revealed is what only what is changing over time. And that makes the imaging very sensitive to what's changing and is really, to me, the greatest power or use of electrical resistivity imaging in terms of monitoring. 00:05:17:18 - 00:05:38:17 Unknown So I thought I'd show you some just some data collection hardware. So I showed you these surface arrays. These are basically the electrodes. They're just stainless steel, eight inch stainless steel pipes, spikes that go a few inches and eight inches into the ground or so, six inches just enough to make electrical contact with the ground. You can also put the electrodes down. 00:05:38:17 - 00:06:00:09 Unknown Boreholes, as I've shown here on the left bottom, and they go in the analysts. And so the borehole can be used for other purposes. They're kind of out of sight, out of mind. Down here on the bottom is what a commercial instrument looks like. This is for a very large array 302 electrical system. But operating these systems is completely autonomous. 00:06:00:09 - 00:06:25:21 Unknown So once they're set up, they'll just continuously collect data. They're multichannel. That means that you can collect many voltage measurements for a given current injection, which enables rapid data collection resulting in a faster time resolution. If you're doing time lapse imaging, a lot of these times they're remote. So we put these systems out in a remote location. They just need power and we can access them through an Internet connection. 00:06:25:23 - 00:07:01:02 Unknown The biggest disadvantage of VRT is that it doesn't work with metal casing. You have to have a PVC or a non conductive casing and you can use it in the spring intervals if you don't have electrodes on the outside. So why what why is what we're imaging electrical conductivity here and why is that a useful property? Well, we don't really care that much about electrical conductivity, but we know that it's governed by several things that we do care about, and that's the porosity and the moisture content and the soil flushing example. 00:07:01:02 - 00:07:24:19 Unknown I'll show you in a moment. We're basically imaging changes in moisture content by their effect on the electrical conductivity. It's it's sensitive to surface area of the or the texture of the soil groundwater composition. And basically what that means is the fluid conductivity of the groundwater. That's another really important one that we often use to image how things are changing over time. 00:07:24:21 - 00:07:54:09 Unknown If you're injecting things, for example, and another one that's not used as often as temperature, but the conductivity, the ground is sensitive to the temperature of the ground. So we're focusing on the site of the 100-K area in the Hanford site. So if you're unfamiliar with the Hanford site, I put some statistics here that I won't read through all of these, but basically we have subsurface contamination. 00:07:54:09 - 00:08:24:06 Unknown Some of it falls not perfect, purposeful from various makes some things and discharges into the subsurface. That happened when the site was in operation generating plutonium for the US weapons arsenal. So we're going to focus up here in this area that's called 100 K, which is way up in the northern part of the Hanford site, where there were two reactors. 00:08:24:08 - 00:08:52:00 Unknown Here are a few historical images of the Hanford site. You see in 1942. This is this was before in as part of the Manhattan Project and before the site was installed in in 1964, you can see the Key West in the case reactor here. And the area that we're interested in are these chromium storage tanks and associated subsurface piping that leaked at some point during the operate operations and contaminated the vital zone with hexavalent chromium. 00:08:52:02 - 00:09:25:18 Unknown And there's a consequence chromium plume. You know, it's threatening the river there. So this is what it looked like in 2017 after much remediation operations. And in the end, the area again is that we'll look at where the tanks were is here. So the first approach to remediating the site was was to excavate. So they basically dug a pit and they took the contaminated soils away. 00:09:25:20 - 00:10:03:11 Unknown They would test the soils as they dug and chased them downward. At some point it became infeasible to maintain the site slopes that they needed and they could no longer dig. So they had to. So they backfilled and plan B for the second phase of the remediation was to do soil flushing in situ. So a flushing. So the way soil flushing works, if you just apply clean water at the surface and because hexavalent chromium is mobile, as the water infilled stream down to the water table, it takes that chromium with it and transported to the to the groundwater where it's removed through the pump entry system. 00:10:03:11 - 00:10:35:05 Unknown And you see several wells out here, some of which are used to pump the water out and to capture all of the flush water that moves down through the system. Chromium also. Well, you can imagine it's very it's difficult. I mean, in order for this process to work, you have to work out where the contamination is. And because of heterogeneity that exists everywhere, it's difficult to know where the water is migrating or, for example, how many poor volumes of water are moving through a given part of the subsurface. 00:10:35:07 - 00:11:02:20 Unknown So we installed it and any r t array. I can go to the next slide here, which so we installed an air Terry over the over the flushing boundary. And each one of these black dots is where an electrode was placed. But the idea that we would monitor how the flush water was moving down to the subsurface over time as a performance metric to make sure this entire area was getting treated with water. 00:11:02:22 - 00:11:29:07 Unknown And so we had eight lines, about 5.4 meters between lines 256 electrodes in total. And we took a survey every 2 hours or for about three months in 2022 and 2023 when they were doing the soil flushing operations. So here's what the electrodes look like. The electrodes are you can't see them. They're spikes down to the bottom of these boxes and then they're just connected to these cables and then the box is closed. 00:11:29:09 - 00:11:53:11 Unknown That's just to make sure nobody touches it while it's operating. Because it does. They can the systems can put fairly high voltages on the electrodes to get enough current in the ground. So during operations that the flushing was applied in three and actually three different zones, there was an eastern zone and a central zone or a western zone and a central zone. 00:11:53:11 - 00:12:25:19 Unknown In an eastern zone, the central zone was continuously flushed and then the east and then the flushing switched weekly between the east and Western zones and they applied the water. They started about in 2024, 2023, 340 liters a minute. And then they ramped up to 454 liters per minute over time. So one of the important parts of this is the geology or the and how it influences the freshwater, as you'll see. 00:12:25:19 - 00:12:48:23 Unknown So we have this pit boundary. And again, the pit was back filled with cleaning materials. So they're potentially have much different models on transport from a lot of the geologic properties than the than the Hanford formation in which it's embedded. The Hanford formation is sand and gravel, and the Ringgold formation underneath the sands itself is generally lower. 00:12:48:23 - 00:13:20:08 Unknown Permeability in the water table is way down here at about 120 meters in elevation. So we had we were limited on how wide we could make the electrode array because there was a foundation in a way. So we can only see to just beneath the Hanford interface in these images. We could we unfortunately weren't able to see through to the water table in this case. 00:13:20:10 - 00:13:48:18 Unknown So here's what the soil flushing looked like. So you see that basically you see the application lines of the soccer hoses placed on the surface. And again, they had, you know, 494 liters per minute moving through these lines at the high point. So these electrodes deployed out here and here in the array. This is when the war started. 00:13:48:20 - 00:14:17:21 Unknown We went out there several months after they started and looked again, and it looked like a green oasis out there was one of the only green places, Hanford site at the time. But there were there was many weeds that had grown and were continuing to be watered by the fresh water. So it was interesting. So what I'm going to show you here is a this is a a time lapse image of the flushing operation, the 2023. 00:14:17:21 - 00:14:46:21 Unknown So what you're going to see here, this is the this is the flushing boundary here and the electrodes. And what you're going to see is the changes in conductivity over time as they start applying flush water in 2023. In this case, we're doing 12 surveys a day again. And the way this works, as we take a baseline image before they do the soil flushing and we take many of those to establish the noise conditions so that we filter the data properly after it switches on. 00:14:46:21 - 00:15:14:09 Unknown And then so flushing starts and you're collecting 12 surveys a day and the data are automatically are autonomously transferred to our computing resources at panel. And they're there, they're processed and inverted and archived and also posted to an interactive website so that the operators and those interested stakeholders can watch this water flushing happening over time and kind of see where the water was going. 00:15:14:09 - 00:15:49:13 Unknown So this this is kind of a summary of what we saw in 2023. Again, these blue ISO surfaces are caused by increases in conductivity because there's water there. And I'll just let this rotate around a little bit until we get to a view here that we should stop at which I started it over. Didn't mean to do that and see if I can stop it one more time here if it starts over. 00:15:49:15 - 00:16:20:06 Unknown So what you might see here is that the flush water is being significantly influenced by the boundary between between the pit and the Hanford formation. So it's running down. Flush water applied here is running down kind of to the bottom of the pit where it's pulling up and then moving downward. I should use my pointer for that. That's right along this boundary here, water flowing down and aggregating here and moving down. 00:16:20:08 - 00:16:53:13 Unknown You can also see under these so application lines, there's not a lot of water being infiltrated into the into the subsurface here, which continue to rotate. And there are some anomalies or at least we thought these were anomalies until we went out there and ground was wet here, like here outside of these these areas that are outside of the outside of the soil flushing boundary, those are leaks basically. 00:16:53:14 - 00:17:31:14 Unknown They're actually there. Then just a few more. Go ahead and let this finish. Rotating around After a while, there are flushing. It kind of reaches a steady state where, you know, you're just getting more pore volumes of fresh water through the soil here. And that was over one month of soil flushing the first month in 2023. So we also had this in the website. 00:17:31:14 - 00:18:02:17 Unknown This is what it looked like. This was kind of a prototype that we put together for the stakeholders to use and clean subcontractors. I'm just showing here. It involves images similar to the last one, but this website had the capability to rotate images around and slice them in different locations and change the transparencies and things like that so they could kind of investigate where this water was going over time. 00:18:02:19 - 00:18:33:02 Unknown So this is just basically a summary of what we the ERP, the standard ERP based and now. Well, so what I'm showing here in the in the bottom, these are basically at the highest the changes in conductivity, at the highest at the highest flow rates are experienced in 2022 and 2023, which gives you an idea of where the freshwater water was going at the time. 00:18:33:02 - 00:19:05:01 Unknown So it looks different point of 2022 because the application rates were much, much smaller than they were in 2023. So 111 interesting thing to note in terms of whether this was useful or not to the contractors is if you look at this cross-section, G, which is the one down on bottom here, you'll see that there's not near as much freshwater reaching the deep subsurface, particularly below the pit, because below the pit is where the contamination is. 00:19:05:03 - 00:19:30:20 Unknown And so the pit boundary is shown right here. And in both cases, you're not seeing as much flush water moving down through there to that section as you're seeing in the northern or northern cross-sections. And it seems to be impeded by some low permeability materials here in the in the pit backfill. Well, at some point they recognize this. 00:19:30:20 - 00:20:00:16 Unknown They stopped flushing and they moved they moved this water flushing array further to the south to try to get more water to this area. And when they did that, they had a huge pulse of chromium come through of hexavalent chromium, come through the pump entry system. So the air imaging in this case did provide some very useful performance diagnostic information that they used to to get more chromium out of the subsurface. 00:20:00:16 - 00:20:34:15 Unknown In this case. Okay. So this is basically the state of the art in T right now is time, time lapse imaging. And we get asked all the time, what can you use these images to to to estimate things like how much poor water went through, you know, each part of the subsurface or you know, more quantitative measures of things that could be useful to better understand, you know, what's happening as they're flushing water here. 00:20:34:17 - 00:21:03:12 Unknown So that is where the machine learning is coming in. So this this is kind of a busy flow chart. I'll walk you through it here. So basically what we're doing is we're training a neural network to process the timelapse data and not to give us a change in conductivity, but to try to estimate the unsaturated hydraulic properties of the soil that are governing how water moves through the subsurface. 00:21:03:12 - 00:21:39:19 Unknown So the distribution of permeability and porosity and the unsaturated flow parameters they're called that connected parameters that govern how the water is moving down. If you can do that and you can do that accurately enough, you could simulate how much water was moving through each part of the subsurface. You basically have a very a well calibrated and I know the buzzword is a digital twin of the system that you could use to assess the performance if you had confidence that your model inputs, i.e. the porosity and permeability and all those other parameters were accurate. 00:21:39:21 - 00:22:04:23 Unknown And so we did this testing to see whether we can there was enough information in the data to estimate these parameters. So the way it works is that we have randomly generated scenarios and that's that those are these are these P1 through PN, they're basically different hydrogeological scenarios where you pick scenarios and they're true to the subsurface. 00:22:04:23 - 00:22:31:12 Unknown Those are input to a high performance simulation code called Pfortran and Pfotrain is not only to simulate not only the flushing process, the flow and transport, but it's also able to simulate the neural data that that would that we would measure during that process. And so it's able to is basically the teacher that's able to generate realizations of the ERT data for a given set of parameters. 00:22:31:14 - 00:23:06:17 Unknown So we use that time lapse data to train a deep neural network, actually split it into two parts. The first part is used to train a deep neural network where the inputs of the neural network are the ERT data and the outputs are the sort of hydraulic properties. Once the network is trained, we have a data set that the network has never seen before, and we passed that through the through the neural network and predict the sort of properties that those ERT data came from. 00:23:06:19 - 00:23:35:17 Unknown And that gives you a performance assessment of how the network is performing. Basically. How well does it predict what soil hydraulic properties that that that were used to simulate the data? And I'll show you some examples of that in a moment. But this is important because it gives you a measure of uncertainty in terms of how well the network is able to predict the soil properties. 00:23:35:19 - 00:24:10:05 Unknown So once the network is trained, you apply it to the actual the real field ERT data that goes into the train neural network. If you predict the hydraulic parameters. And I should note here also that in addition to the hydraulic parameters, you have to predict also all of the petra physical parameters. So the petro physical parameters relate how if you remember the slide I showed you earlier, how porosity influences electrical conductivity and how saturation or water content influences electrical conductivity or or the poor fluid conductivity. 00:24:10:05 - 00:24:37:14 Unknown So all of that there are there are Petra physical relationships that use all of those parameters to convert them to more conductivity. Those have to be estimate also. So once those estimates are made to the neural network, we plug those into people and then we simulate the true sort by flushing behavior, including the uncertainty, because we have those estimates from a here. 00:24:37:16 - 00:25:01:11 Unknown So the way that this is done is we use the pilot point approach where we the computational mesh is discrete sized and then you have pilot points, which are these red points distributed throughout the domain. And so what the network is tasked with is estimating the hydraulic properties at each of these pilot points, and then they are correlated to the right. 00:25:01:13 - 00:25:30:04 Unknown So to fill out the entire volume. So the so we do that, we and then we generate the corresponding air data from each one of those realizations. This case we have 500 simulations, it's about an hour for simulation and 3200 CPU hours, which is, which is, seems like might seem like a lot, but that's actually not very many CPU hours at all. 00:25:30:04 - 00:26:01:22 Unknown And in a world of high performance computing, so very moderate amount of computational effort here and we do some tricks called adding noise to all of these simulations called lightning to to generate more realizations, up to 10,000 of them. And then there's this data compression step and then we do the training. Interesting that the training, the network that we're using, that we're training is actually a very vanilla standard, deep neural network with seven layers. 00:26:01:24 - 00:26:33:15 Unknown And it only takes 10 minutes to train ten or 15 minutes to train on a on a decent graphics card. So just to give you an idea of the computational effort and how it's distributed in this application. Okay. Next slide. So what I'm showing here and these are this task for the tasks were funded this year, so a lot of these results are hot off the press. 00:26:33:15 - 00:27:02:22 Unknown But so what I'm showing you here is a few examples of how the neural network performed in terms of predicting the permeable state of the subsurface and the porosity for a few different cases. So this is our this is our true or our test dataset. This is a distribution at the neural network had never seen before. And when we pass it, so we generated the ERP data from the soil pushing process to given this distribution of permeability. 00:27:02:24 - 00:27:28:14 Unknown We ran it through the neural network in the neural network predicted that the permeability would look like this. And you see that it's a little bit smoothed. And that's you might expect that. But it does a really good job of doing what the permeability distribution is. Just from the time lapse, the ERT data. And if you look over here in this example for the porosity, you see the same type of thing. 00:27:28:14 - 00:27:50:07 Unknown It does a really good job of predicting what the ferocity is. Does the same with the other of the petro physical parameters that are distributed. And we also have to estimate the conductivity of the freshwater that's being applied at the surface and the baseline for water conductivity and things like that. All of those are estimated by the by the neural network. 00:27:50:09 - 00:28:17:19 Unknown So basically what this is telling us, there is a lot of information in the time lapse of the data, and we kind of know that because we can do the inversion and get the bulk activity. But there are a lot of information regarding the hydraulic properties which are really governing how those data change over time. And it's the machine learning that's allowing and allowing us to extract that information and then apply it to a simulator. 00:28:18:00 - 00:28:47:02 Unknown This is much, much more useful data analysis than the air imaging is because it produces a model, a simulator, a digital twin that honors the ERP data that you can use to to do all kinds of computations regarding the performance of the remedy in this case. Okay. So here are a few more. This is just a couple of the porosity and permeability distribution. 00:28:47:04 - 00:29:16:06 Unknown In different cases, you get the same kind of results. So if we plug that, if we plug that those process, so if we plug the field the data into the neural network and predict the hydrologic properties of the true site, this is what this is what the simulation looks like. What we see is the pit is in terms of the permeability in the pit is much more permeable than the Hanford formation. 00:29:16:06 - 00:29:52:12 Unknown And we know that's true. That's very evident in the ERP results. And also something you would expect based on the the materials that they put in there. The Hanford formation is actually say the formation is more permeable than the Ringgold. We know that is not true. So in this case, that's that's very suspect and it's likely that may be happening because the Hanford formation or the Ringgold formation is very deep and the data are very sensitive to it. 00:29:52:14 - 00:30:23:02 Unknown If you remember that our line, our online links were limited. So we don't have a lot of resolution there. So that's what we have. I mean, we're still modifying this. We're now using more training data to better represent the true uncertainty in terms of what the distributions might be, ideas of property distributions. But the results are very encouraging so far. 00:30:23:04 - 00:31:01:18 Unknown So just so if we use those permeability processing distribution that I just showed you and simulate the this is in this case the change in saturation over time during the actual flushing, this is what they looked like. So this, this would kind of be one result that you could compare, you know, in terms of the use of this to the extent images just and what you see is water piling up on the interface between in the pit and the Hanford formation and then moving downward to this is one week in two weeks, just like we saw in the data. 00:31:01:20 - 00:31:30:06 Unknown It piles up on Hanford formation then, which downward from there becomes very saturated in Ringgold because the permeability and porosity are are the permeability or the porosity is less in that case. So anyway, it's a work in progress, but this is kind of the idea and where we're heading with this this year and this machine learning is making a huge impact on what we can learn or the information that we can extract from the time of ERT data. 00:31:30:08 - 00:32:24:04 Unknown So that was my last slide with that. If there are time, I'll take questions. But yeah, thank Tim, good presentation. All right. Somebody, thank you for the talk. Would there be any room for improvement in your inverter by using a, say a physics based neural network instead? Yes, in theory. But when you have when you have very complicated heterogeneity and things like that, physics based neural networks are they're very challenging to implement. 00:32:24:06 - 00:32:45:04 Unknown And I'm not saying we couldn't do that or that's not worth anything. It's just much, much, much more of an effort than what we're doing here. And that's fine. I don't know why we haven't gone down the path at this point, but I would imagine a pen would be better or could do better if you could if you could implement it. 00:32:45:06 - 00:33:22:24 Unknown Yeah. Any more questions or no online questions? No more questions. Yeah, I have just a question regarding you are talking about resistivity data analysis and it is impacted, of course, by the moisture content and the water content. What is the impact of the Columbia River seasonal flooding? You know, and the models that you are doing, you try to make interpretation and analyzing the data. 00:33:23:01 - 00:33:50:00 Unknown Have you looked at the seasonal changing and variation in the living on water that is coming to the formation of. Yeah. So this is the primary impact of the variation in the Columbia River stage would be that the water table goes up and down. And as I noted earlier, we can't we can't see to the water table. It's too deep for this array. 00:33:50:06 - 00:34:20:14 Unknown So the water table movements in this case don't affect the data at all. And so that the, you know, the river stage variations, they don't impact the analysis at all because it's influences are just too deep to see. So. Okay. Thank you. Any more questions? There's a there's one on the line. No where. One more. Yeah, there's there's one or one person has two questions. 00:34:20:15 - 00:34:49:06 Unknown The first one is what are the assumptions during training data generation? Do answer that one first. Yeah. Yeah. So the assumptions, the hydro, all of the, all of the hydraulic properties are and all of the hydro drone properties are estimated at or produced at every pilot point. We assume that in this case that the arbitrary parameters are, are homogeneous in each of the layers. 00:34:49:10 - 00:35:18:19 Unknown So they don't change what the pilot points. And we assume the fluid conductivity doesn't change the freshwater fluid conductivity and the baseline for water fluid conductivity. And we're also making assumptions because we agree we're making something assumptions about the about the spatial statistics of the burial grounds. And so the test data have the same zero grams that the data that the gender or that the training data have. 00:35:18:19 - 00:36:01:00 Unknown And so one of the things we need to do next is go back and say, okay, if the statistics aren't the same, the spatial statistics, so what happens? What happens to the predictions? And certainly the the you know, certainly the predictions won't be as good, but I, I don't know how much that will influence things. So we do need to go back and in our testing data, add more and variability that represents the true uncertainty, you know, to see how well we can predict off off cases where, you know, the spatial systems are much different than the training data software, for example. 00:36:01:02 - 00:36:30:13 Unknown Thank you, Tim. Any more questions? You'll have more. There are more, but all right. Okay. So this is kind of a quick one, but how many amps of current are you applying to the ground through the electrodes for the study? Yeah. So that's that's governed by the ERP system or system. Well, we'll ramp up the voltage till it gets enough current to major potentials at all of them. 00:36:30:15 - 00:37:08:13 Unknown And for this application, they were, you know, anywhere between, I would say 50 and 200 milliamps, somewhere in there. Okay, Thank you. Oh, one more quick one. How easy or difficult is, is it to use the joint inversion software. I guess just generally. Um, and which joining Virgin Software is referring to here. But there's no clarification. Okay. So I will say this. 00:37:08:13 - 00:37:33:09 Unknown There is an Pflotran, a capability to do joint inversion, and that might be what this reviewer is asking. And it's it's computationally intensive and it's fairly difficult to use. And that's one of the reasons not difficult to use, but it's just computationally intensive. And there are issues with getting stuck in local minima. That's one of the reasons we tried the machine learning in comparison. 00:37:33:09 - 00:37:54:16 Unknown The machine learning analysis in comparison to a true deterministic joint inversion is orders of magnitude easier to implement and get useful results out of that. What? At least that's what we're finding for now. That type of joint inversion is something we, you know, people have been trying to do for a long, long time. It's very, very difficult to do. 00:37:54:20 - 00:38:03:02 Unknown And this is another advantage of machine learning as it's making that type of analysis much easier.