GeoSci.xyz SimPEG, 26 May 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: SimPEG Meeting May 26th

Description

Weekly SimPEG meeting from May 26th, 2021

A

Nice to see you guys all um add yourselves to the quick reports. If you got something, there's not really too much on the agenda for me today or things to talk about.

A

I don't know how about today. Let's just do a quick round of everybody and see what everybody's.

B

A

It's been a while, since we kind of done that sound good.

C

A

C

You're having my screen first, if you don't mind yeah, I'm.

D

Literally, just wrapping up all the loose ends. My project ends on monday, so I hope to get a the version, one out for amg 3d and I revamped docs and website and all the things. So I can leave it a bit less attended after that. Hopefully,.

D

That's it tuesday, I'm going to start more in the direction of data assimilation for reservoir modeling and in the delphi consortium and with some other things, I'm not yet 100 sure with terrace is a small company from germany. So we'll see how that new phrase goes interesting.

A

You're still going to be able to take it keep up every week with us.

D

I think so so it's going to be three days a week at theo, delft just another section in the department and two days a week for this company, but it's remote. So I I stay. I stay where I am yeah. I hope.

A

So cool john: what have you been up to.

E

Yeah I have uh been tackling the merging main into simulation mt right now, yeah, ultimately to get tiled all brought in or made and all brought into the simulation or the tiled simulations um yeah. So just picking away at that slowly here and uh I went on a little bit of a tangent and looked at that pi amg solver or that multi-grid solver uh yeah for the big problems yeah, it doesn't take up too much ram, but it it's still much slower like a couple or over a couple orders of magnitude slower than party.

E

So so I don't know, maybe there's some options that you can play around with. I was just a quick and dirty implementation, but uh yeah good on ram, but not so fast, not much faster, interesting yeah, and then I got a request at work to look at tiled tensor meshes um they don't really. I haven't really been able to sell everybody on the off trees fully, so they like the speed with the tiling, but they won't kind of want the tensor mesh stuff, so I'll maybe be picking away at that over the next little bit.

E

If it's possible or yeah, I might bug somebody to uh if I get lost in the title code just to see if.

F

E

F

What's the issues they have with octree, I'm just curious of: what's the, why they they have of their skepticism here.

E

I guess it's: when the the cells start growing, it's like yeah, there always just seems to be a distinct boundary like it's distinct between the layers or like in the octree levels. uh Yeah I've I've tried- uh or I haven't, tried too hard, but I'm sure if you discretize properly enough, you can kind of minimize that effect, but uh everybody that I got to play with it they're just like. Oh.

B

E

Always looks you can always see this boundary where it's, where the cells change, um but yeah, that's mostly on dc, though I've had really good luck with the mvi stuff. um No one can no one really complains about the octree on the mvi, but the dc. It seems to be a little bit of an issue.

F

Yeah no like this, the potential field works great, but yeah. I can see the I can see the what's happening with the uh with dc. It's true but yeah like maybe there is some good practice to implement for that for the old tree, yeah well,.

E

I think like dom- and I did a test set together here or you were involved there tebow on that um on that monitoring, uh the dc one and those audrey models. Look pretty good. I thought, like didn't really seem like there was too much boundaries at the entry levels, but yeah, I guess playing around with it enough people are just like hey.

G

I think, what's uh my feeling is that uh they're always pushing too hard uh the misfits and then that's when you start seeing the uh the actual changes that make sense, yeah yeah I mean I can tell you right away like the tiled on dancer is, is like it's pretty good. It's going to be messy because you can't nest anything right, so start doing partial volumes and all that I don't know yeah.

D

I would vouch.

G

For them to to get a better handle on actually, instead, I mean.

E

Yeah I'll still keep kicking at that horse for sure, but were.

B

They not like impressed by the by the savings you know, because you got sort of where those those corners are in the mesh and in octree it gets to be not very populated with cells, but when it's tensor, you got these long skinny like prisms, yeah yeah, and they weren't, I think also the for for em. Maybe the you kind of run into that aspect, ratio of the cells- and you might not be solving it as as accurately there. Somebody correct me if I'm wrong, but I thought there was. I found.

D

Plenty saw some of that in our cscm comparison that she just couldn't get it down with the oak tree to an to a comparable error level. As with the tensor mesh. Because of this two one ratio.

F

Like I, I think I've said it before, but the ubc dcip octree ubc code, like the commercial one, cannot handle non-cubic cell, like you, can create a mesh with non-cubic cell but you're gonna. It's it's gonna create massive artifacts and I think there is maybe something in the gradient or I don't know, what's uh what's really happening but like yeah, it's like the the like the ubc code, with the when you put like cells that are non-cubic, you create.

F

I had experience, whereas it created massive artefacts that disappeared when I was putting cubic cells, but how how non-cubic did you go? Oh, like a ratio of like even like 1, to point five like something like ridiculously small ratio and was like yeah? No, that's like I could not use it with non-cubic cell. Basically.

B

Like I might try out some some testing on that, because we made some improvements to that code, nothing related to what you're talking about, but um yeah, I'm kind of in it right now, so I might be able to just try something. Was there a particular geology? Was it real data? Was it synthetic? Oh.

F

Yeah, it was real data and, like I inverted with like cubic cells, was, and uh it was a very big survey- that's the thing too so I was trying to, and I was trying to to gain on like lateral thing extra, like my electrical spacing was 50 meters and like line spacing, was 150 or something like that, and I was trying to get it some a little higher in term of like the across line to gain a bit because, like uh with a cubic cell, I will only I could only go for even on the cluster.

F

I could only go for 25 meter cells, so only two cells per extra spacing so like the short ones were not that good. But I could not do more and I was going over over memory and the inversion was already even with the 25 meter cell.

F

That inversion was taking like two weeks to run, so it was like a big one, so I was trying to to increase a bit just since, like I think I went like 12.5 and like 20 meters, something like that and he was just putting massive con and and suddenly massive conductors appeared between lines, and I mean this was not and those features were not in the cubic cell response in inversion. So.

B

Okay, yeah, that's interesting because I I inverted mount isa with the latest code. I guess it must have been like eight months ago or something- and I remember having I'm trying to remember if I made cubic or non-cubic cells when I inverted that, but I might go back to it.

B

Yeah cause we've been playing around with sensitivity, waiting.

F

But then I I I I want to go back to what you were saying for the tensor and okay, like what you said you cannot. You cannot do something with tensor that you can do with oxford or something about the volume. But I could not understand what you were saying. I'm sorry.

G

I think you can do it. It's just. uh You need to be careful on the uh when you do your projection right from the global to local, um and I think joe, maybe with your new enterprise that can they can possibly do it. It will be a bit slower to generate them, but it might work yeah. It might work, it's just it's a bit messy because your your nested meshes are gonna. Have some uh you know it's not it's not a one-to-one interpolation, because the cells are.

A

G

To be overlapping.

A

Well, right, yeah, the volume average operator should work for it. If I get tested against it, it should work for that in general. Okay, though I mean, I don't think, it'll they'll probably be some small, artifacts and stuff from the overlapping cells. It's not as clean as the one-to-one ratio from like option nesting like for nesting them. As far as timing, it'll work it'll still be volume averaging.

E

E

I guess that's yeah about all that. I'm working on and uh yeah doll I'll be getting on to the subtitle or sorry the freak, the vtem styled survey, uh subtitling um yeah. I just we're moving offices and I just got my cluster back up and going so good to go.

G

Yeah, I still need to finish that example again.

G

Not to worry ben is working on. His work is working on it right now uh using the new implementation, so we'll be testing it on our site first, but uh I'll try to get it to you by the end of the week.

B

I got a few things: um we're gonna do a big release of the proprietary proprietary software stuff that our our group does. So it's not really related to any of your things, but um some some stuff that might be interesting is that we've been playing around with different sensitivity, weightings for dcip octree and we've also been investigating our time domain, em octree code, so um by by kind of identifying, maybe issue, let's say: imperfections: we're really kind of trying to zero in on uh kind of.

B

What's the best way to, I guess, define sources or kind of like these little details and uh because we're we're not worried about being able to solve the system a bunch of different ways. We just want to find the best all-purpose approach for uh for for simulating. So uh some interesting things have come out of it that maybe we'll end up getting adopted into simpeg.

B

We've been working on this mdru project with mag and grav tebow and myself, and that's been pretty fun and uh and playing around with uh kind of trying to remove a background or a regional uh contribution and then inverting locally. So we still have some questions, but the mag seems to be giving the good results. The gravity is very sparse and kind of difficult to work with, but we're moving along in that and then when I can find time more docs drinks.

B

So I've got enough to keep me busy.

A

Anything anything exciting uh not too much relating to simpeg um yeah, working on dom's, refactoring, dom's geo, apps uh package and uh yeah. I did finally kind of touch some syntpag code recently but uh yeah just on our local child simulation branch, which maybe one day we'll get pushed, but keeping it local for now. Yeah cool.

A

Have you been up to something.

H

uh Last week, I just uh I did some minor reveal for the depth sweetie. According to you and the dom's comments. I think it's a it's good to be it's good to merge in the frame main branch. What do you think.

A

um Yeah, it looks pretty good. um Do you mind? I think the only other thing is. We need to write a test that does it like a test function that runs it over runs it through, like 1d, 2d, 3d mesh things like that. Just simple make sure it works.

H

Okay, okay, I see so and uh test the test files in the test. Folder.

A

Here you get it.

H

Okay, I can do that.

A

Down here next on the screen.

G

Yeah, well, we did the yeah. We worked with xiao right and yourself. I tried to get this general distance waiting function, looking pretty good. Otherwise, I started looking into uh regularizing the uh you know different the different, like functions of the model in terms of like amplitude right now, it's a bit more complicated than I thought, and just because the way cell weights are applied right, we're doing the mapping the mapping of weights and then we kind of want to have two layers. It's not just applying the weights on everything.

G

Anyway, it's a little bit more complicated than I thought uh some tinkering around trying to see how we can you know, maybe add another layer or reflect whatever we have that's what I'm up to.

A

um Did we could you give any more thought to the so with regards to the depth weighting stuff, when we, you know we're talking about the the volume normalization part of it whenever you give any more thought to that like in context?

A

um So when we have cell weights on like the regularizations right um on all the whenever they're implemented, I believe, like we've multiplied by volume afterwards, because we're doing the integral of that function.

A

It happens everywhere, except for when um you have the update sensitivity, weight thing, a sensitivity weight all like includes the volume turned in so kind of you have to take it. You have to use like a special regularization function. I think, when you're using the that, because you can't use like the t-connectorization this stuff because of the volume like the volume gets multiplied essentially twice there.

G

So it's a bit worse than that, unfortunately yeah, because uh the simple sorry, the uh the small right on the in the uh canal, basically always multiplies by right match volume, even if you have if your cell weights are normalized or not.

G

So we need to change that priority because it's different behavior. Basically, if you use simple or think enough, um it's big of an issue, so we should uniformize uh everything they should all behave the same and then yeah. Basically in the directive, denormalize the uh or you know divide by volumes beforehand, so that when it goes into the regularization it multiplies it back in.

G

I think that's probably.

A

The easiest you just wait, that would be my bet as well just to have everything come in and not expect things to be volume and volume we'll handle the volume here just make everything yeah.

G

Yeah and make it clear that the weights should be volume normalized right. If you put anything anything in there, it should be volume normalized. Maybe that john, that's something you should look into to write like in the in your code that you're, using uh which regularization you're using basically because that might affect you big time.

E

Yeah well for the I've got everybody over to sparse for the for the auction but yeah. I remember we had this chat with ticking off and yeah. It was pretty bad and taken off yeah, so we should create an issue I'll create an issue now to make sure that we have a flag.

I

Like a there's, a subtlety because, like let's say, if you use the sensitivity weighting for instance, then that sensitivity actually includes the volume term. I think yeah, that's what we're saying: yeah yeah! I don't want to re-like.

G

Multiply the volume that's right, so we need to make sure that we need to assert that if someone puts in cell weights they need to be volume normalized all right, so that that way we handle the volumes ourselves inside the regularization.

I

I see so in the regularization. You always multiply the volume, so that's the user's job to like uh divide your weights by volume exactly every time.

A

Or if it's like, for instance, for someone uploading a function like you, you don't multiply that with the volume very easily, because that's just the function.

A

So it's just you have to make sure that your things are volume normalized and it's not dependent on the volume of the cell. That you're.

F

uh Not much much in term of code, I've been mostly focused on uh several case studies at the moment and uh dealing with uh real data sets in very great in very green area that we don't know much about and see what I can do with pgi in this area.

A

A

Song we've just been going around, obviously.

I

Yeah, I don't have nothing particular um yeah, I'm actually trying to invert the remote sensing data using syntax. um We actually wrote a small uh flow modeling code and it's actually basically 1d it's very similar to airborne inversion. So you like take account the remote sensing data, which is the time series that's like as like a one-time version. But you got zillions of points, so you can set it up as a 1d problem and the goal is actually recovering the the head variations in the aquifer system. So I'm probably I'm pretty promising.

I

I'm not sure whether that's gonna work or not, but if that's working that'll be actually quite impactful, because you can actually get the head data from the from this remote sensing data which actually you can get every six days so.

A

I

Working on quite different problems- and I am slowly moving on with the cscmt inversion, so we're going to test the emg 3d implementation in simpac using the cscmt data. I got one dataset from that and I'm going to get another data set from nick.

I

So yeah it'll be interesting to test that code and one thing that we're going to try a little bit new with the mg3d is the anisotropy. So we're going to invert for uh like at least like a horizontally isotropic, but an isotropic resistivity.

I

B

I

Makes a big impact to the cscmt version or not, but anyway we'll test it. So that's sort of what I'm doing.

A

Okay and then me, I just I've been going through um working on one of my own small projects, simulating electric fields in the skull which is kind of interesting, and then I've also been going through the m1d stuff as well, trying to look through it and make simplifications where I can kind of see what's happening and working towards getting that time domain, one working, much quicker, I'm working towards getting it.

A

I just kind of uh my only handle that I don't have on it is I'm not sure how much memory it takes up already in regards to okay, if I want to store a few extra things that are going to be that impactful on the memory footprint.

I

I think it's pretty minor uh joe, but the memory consumption is pretty minor and actually increasing memory. Consumption is kind of fine in the sense that if you can actually improve the time yeah like usually even my laptop- is pretty fine to work like invert, the large scale at one year, data so nah.

F

It's not even using.

I

Like I got eight gig, it's not using eight gigs, so it's uh it's pretty small.

A

Okay, then I'm going to keep chugging away at it. um I just I've been making simplifications to things like you know, just looping like getting all the loop through the source receivers calculate all.

B

A

um If all their offsets all of their coefficients ahead of time, just do that once and go through and describe all the unique frequencies and call that art that um rte function wants things like that.

I

Are you thinking about like uh let's say you got a thousand soundings and are you thinking about handling those ten like a thousand sounding ones? Is that is that what you're, thinking or you're still are thinking at the level of a single sounding? I'm still the level of like a single.

A

Sounding but also like what you're talking about like ten thousand sound menus like things like that, um I mean, if they're all over it'll right now. I just have it kind of going through and looking at the unique source, receiver, offsets and just modeling those for each sounding. So hopefully we can do that, um but at the level of like you know the multiple like the time, the stitched 20 stuff, um you can at least take advantage of, like you know, calculating all those coefficients ahead of time. Instead of repeatedly calling.

C

A

Just getting stuff.

D

If you ever have- or anyone has a huge airborne, inversion, job or project or something, I think it might very well be worth to try and design a filter for airborne stuff, because I don't think any of the filters we use was designed having airborne in mind. So we should. We could probably get away with much or to filter specific for that topic, which would then uh weigh in heavily on the runtime.

I

We talked about that with joe.

D

I

D

I

Probably sort of the best part that we can actually make, because if you can reduce that, I think we're using 101 and I'm not.

F

Sure, what's the time.

I

But we can add 101 filter to 50.. It's factor.

B

I

Yeah it'll be it'll, be.

A

So, anyway, that's what I've been up to check it out here. That's.

H

A

H

Just wondering it is a good idea to end something in the and the depth waiting results in the tutorial gravity.

H

I think I can show the results further with different exponent exponent exponent parameters and.

I

um Speaking of that show, did you like initial depth waiting implementation?

I

I think they actually solved like a simple, uh like least square problem, because the you got couple of unknowns right, like you, don't know what exponent you need to use. You also don't know like that. There's a constant factor that not needs to be into into that depth weighting, so I think they actually just think in the beginning. They they got sensitivity, kernel and then they kind of solved a simple problem to to figure out what that obama.

I

Is that what you have done in in in your implementation or it's it's basically just an input like that set, not an exponent.

H

I think it's uh it can. This function can do the both. We can use the sensitivity curves to determine the exponent and the constant parameters, and we can also improve some other parameters. What is.

I

One that's cool cool and uh related to that. I found that I thought that what the project like projected gauss mutant that we're using it's kind of fine, but still I like uh I compared that that that uh the raglan inversion result with the upc code and I found like a it- was pretty good like, but the especially where uh you have.

I

I think it's actually happening when your update are negative, but then projected gaussian newton kind of put that into a positive value and I think it kind of generates a weird, uh very sharp boundaries. uh So I'm not sure, like. I think uh we spent quite a bit of time to actually like modify that that projector gauss newton, but still it seems like it's not like as good as the ubc code, which uses the slightly different method. I think they're using the log barrier.

I

I guess so I thought maybe that's something worthwhile to think about implementing, especially for that positivity constraint. I think in general that that the projected gospel is kind of fine, but I think if they're like, if the projection has a lot of like uh invisible sets, then uh then that could still make a problem. I don't know that was just a thought, but uh yeah yeah sure you're implementing the old kind of style thing. So I thought it'd probably worthwhile to think about that.

G

You just project it to zero, the right psych. It doesn't move up to positive territory.

I

I I think that, are you talking about the log barrier method or the the projected yeah, if you're just going to just like project it to zero.

G

I

And then it uses if, if the values are on the bound, we use the first order for those like uh values, on the bounds so which we, which we follow whatever the methodology is so I don't know like we can make a big improvement there.

I

The another option was the. I think that that ubc code used his death uh log bearer method. I guess not.

G

Anymore, I'm not sure yeah ben did a long time.

I

Ago, oh, I see so that it uses the projection gaussian now.

G

I

Yeah, oh okay, then.

G

Never mind yeah logia has been abandoned, probably like since, like v3 or something.

A

Yeah, maybe then yeah that's not worthwhile, though well, what I know from the log barrier just like it slows down like the light late iterations, just don't update very much and then it's hard when there's a bunch of stuff near the boundary you get like small updates for stuff. It just slows down when there's a lot of stuff, with the bound.

G

Okay never mind, then.

A

I mean there's no reason we couldn't have like a log barrier method in there. People wanted to use it like that organization function in there like there's no reason we can't have it as another option, but.

G

Isn't it just like inverting like susceptibility and live, would not do the same same thing, yeah just to avoid to go negative.

I

No, I think it's it's.

A

More the barrier term that essentially makes artificial walls on the boundaries you can imagine you know thinking about a minimization function as a whole. It adds these like really bounds and then, as you get closer to it, it goes up and then it's.

G

Harder and harder to cross over.

A

Right especially harder and then like as you iterate, so it starts as like a smooth boundary and then it starts like you reduce the parameter and it goes sharper and sharper.

A

That's what happens like it's just a nonlinear regularization, that's added to it!.

D

If there is time I thought I quickly bring up the question, I asked about plotting in discretize. If my assumption is correct or if I missed something, if we could change it potentially or not, or.

A

Yeah, so I took a look through it and I kind of went through the tensor mesh code. At least, I think what we could do like. What should probably happen is that when you're plotting like those face values and things, we should do a slice. We should slice it before we average it to the cell centers, because right now everything gets averaged to cell centers and then sliced along the axis, which is what's taking up a lot of memory.

A

When you do that, so it'd probably make better sense like to just write it where it slices those faces and edges first and then averages it or it slices those faces.

I

A

Faces edges nodes first.

D

But yeah that would definitely alleviate the memory consumption, but does it make sense to plot the edges at the edges.

A

Yeah, so that's one of the other things that I talked about that garage uh shading. So if you do that like it should have those values.

A

Just be correct and then it's kind of like linearly interpolated through the.

D

A

D

A

D

Yeah yeah yeah.

A

That's great, it's the correct value, but it just interpolates it through the cell as part of the plotting routine.

D

A

So that's. Why, like I think that probably the best route, but then I don't know what to do about it on a tree mesh, because you can't just slice it ahead of time.

D

You cannot, you cannot take out a no.

D

I also know what would happen at the cylindrical mesh. I don't know how.

A

D

A

I don't know I'd.

D

Have to look at it either, but we could at least get it better.

A

For tencent, I'm actually gonna improve it. For that, like pretty well the beyond tree mesh rate like it's, you could do the nodal. I think you could I'd have to look but like you can't because things aren't regularly shaped, so you can't take advantage of those slicing operations.

A

D

It's nothing urgent. It just came up because yeah some plots with it and then compared.

D

I think I I created a p color mesh manually and then plotted with discretize, and then you can see there are tiny differences because yeah one is interpolated to cell center and then yeah. I just remember me that I had a couple of times. I could compute the electric field, but when I tried to slice to plot it with the slicer or plot slice, then it would say memory error.

J

A

Yeah, we should definitely just slice it ahead of time and the tensor.

A

A

Anything else: what's the chat about.

A

If not, I don't mind keeping things lighter.

G

Sounds good! um I'm still curious, though, about the uh the editor of solver that you talked about joe right. Why is john getting such a slow slow solve compared to what you, because you're resolving for three million cells? I mean.

A

Yeah, like 2.7 million or so I don't know in like in like 18 seconds. That sounds pretty remarkable. Like I don't know, I have to look at what he was doing and what this was.

E

Yeah I did have like it doesn't take the multiple right hand side, so I have to loop over the right hand side so that right there kills quite bad, of course, but that's the thing, though, when you look at the par diesel, one they're actually looping over the sources as well. So I thought, oh, maybe I can get away with it, but it's.

D

Not the same though yeah, but they probably don't have to make the full, even though the loop over the right hand, side right.

G

They just do it as a matrix as a as a matrix operation.

E

Well, I thought that yeah there's a for loop right through it through all the sources. If you look at in piment silver, I thought it was financed over.

A

You know you can.

E

A

Like it does it at the sea level, I don't know there might be a way that we can yeah.

A

We have to see. Is it like it's the iterative solvers that kind of yeah, let's probably check it out,.

E

I'll push it to this branch that I made for that neo guy uh I'll, just add it to there and you can take a look. Yeah dieter.

A

Have you messed around with that pie amg at the algebraic, multigrade solver package.

D

No, no there's also the pi amg, or is it this one? The algebraic, multigrade, solver and.

A

D

I

Yeah I tried like kind of it works like for, if you're, using homogeneous value, it works quite well, but once you introduce an air cell, I think it just like gets so poor. So, like yeah, I think I kind of gave up at one point like because you have to do it more sophisticated, like.

D

That the thing is with multi-grid it's: if, if it works, fine, it works fine and if it doesn't, it doesn't so like the stretching the the thing that is implemented in emg 3d. There's this line relaxation and semi-coarsening, because otherwise, with an isotropy or stretched cells, you would very quick become very slow.

D

So there needs to be just the pure multi grid will only work for very, uh very uh restricted.

G

D

A

They have anything to do like how is it related to the condition number of the matrix.

J

I don't know I.

C

Just have a matrix.

I

The condition number gets bad, so I think it's one reason why, let's get so poor.

D

Like there are many things like in emg 3d, if you make the air very resistive or the frequency very low, you will very soon go to never ending land because convergence becomes very slow. uh So you can, for instance, at the moment, not use it for dc. We would have to implement additional things so each I think it's and I think it's a with a lot of iterative solve for the case that you have to to tailor them to your problem, whereas with direct solvers, they are much more robust and flexible.

D

So yeah, I'm not surprised that with just a quick try, it's probably not very quick yeah.

A

And is that using um multi-grid as the solver or a pre-conditioner to.

D

Either but usually.

A

C

D

I can use it as a precondition for the for any or many of the sci-pi uh trill of subspace solvers, but often I use it on its own yeah.

C

A

I had always heard that if yeah the best performance comes, if you can finely tune the multi-grid method. But if you can't don't have the time or can't do that that using it as a pre-conditioner can get around some of the.

D

Yeah, yes, particularly strong, and I thought it should be your stretching things like this. It can help, but then again it's it's so difficult. I I wouldn't have few rules of thumb because you just have to try it. Sometimes it helps.

C

D

It doesn't yeah, it's really tough.

I

Yeah surprisingly, if I read the literature, people still use the jacobi and I'm yeah, which is uh surprising, but it's probably working.

I

I mean I think the convert just gets poorer as the system goes large, but I think people still use it somehow, which I don't know. Oh that's working.

G

Okay, second party, so nicely all right, see you guys, but.

I

Make the domain small and use the direct solver.

G

Yeah exactly yeah so far. That's that's the only solution.

D

Cool all right well, but when, if you get the csa mt working, that could then also work for john and and some mt work right.

I

Yeah, I think so the potential um john is actually transferring emg into an mt code may not be a huge hurdle. I think, because, as long as we use maybe the secondary field formulation, I think there's nothing to change. We just need to change the right-hand side in the emg 3d, so I'm not going to do that, but the I'm just going to test whatever existing code and the cscmt example, which is much simpler.

I

But I think what probably worthwhile to generate is a receiver class that can handle the impedance so because it seems like the typical system is contact system and what they basically provide is the apparent resistivity and the phase from either zxy or zyx. So yeah. I think that's probably my goal and uh as long as we can implement that- and I think from there actually updating that into an actual empty code.

G

I

Be a big big kind of work. I guess so there's a chance and I think that eater's code is working like relatively high frequency. Let's say I know 0.1 I guess for sure, but I think that's probably mostly the case that you would like to invert for for a lot of mining unless you go really really deep, like a mental, so I think it's in a practical sense for neural exploration. I think that could be quite useful, but anyway, that's uh just the thought that we have at the moment yeah.

E

I could uh be into that for sure. That's that's cool, um we're actually doing some csa mt ourselves. Now um I've got some good data sets that might be able to use like if we we've like got all the wire paths taken out like we've collected, like all the details, so that we can simulate it in a way. Eventually, I was going to get that oh great, so you're actually going.

I

To look like kind of record where the source locations are.

E

Yeah exactly because we just had a project where we thought, like the the client thought it was going to be far field, but it turned out. The ground was more resistive than we than they thought and it was more into the near field. So you kind of need the the source to find, um but yeah. That's just.

I

Yeah one of the ones we have here that we were looking at uh yeah. I was just kind of laughing because people don't record at source locations because.

B

I

They are using is the 2dm keypad.

B

So they don't really need to.

I

Know where the search conditions are it's a lot of historical data doesn't have the source communication, which is like. I felt crazy and yeah.

E

It gets better yeah exactly far field example or far field assumption all the time. So.

A

Sounds good? Well, then, I don't know see you guys. All next week have fun. You have good stuff.

A