GeoSci.xyz SimPEG, 12 May 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: SimPEG Meeting May 12th

Description

Weekly SimPEG Meeting from May 12th, 2021

A

So again welcome everyone to the weekly meeting nice to see you all uh and, as always add yourself if you've got some quick reports to go through would be great.

A

I don't have any particular agenda items again this week. If there's anything, anyone else would like to discuss with everyone. As far as like big picture stuff,.

A

um Add it and then we can go through it after the quick reports, if you like, if not, let's kind of go nice around everyone's going in um so I see soggy has added things, but he is not on yet.

A

Lindsay do I go nice, give him a little bit of an update on anything governance stuff. It looks like you got something in there.

B

Sure, mostly just a quick um request, if folks could take a look at the governance pull request, um it would be great if I mean, if you have any suggestions for changes, please please feel free to make those suggestions or comment on the pull request.

B

If you're happy with it it'd be great to just get an indication from folks so either like approve the full request or thumbs up the pull request, or something like that, just to get sort of an indication of broad consensus in uh in the community so um yeah whenever, whenever folks have a chance, but it'd be great to get that sort of formally in and we can move on to the next steps.

B

A

You might go next time.

C

uh Yeah I put in some some work over the past couple of days to yeah fill in the the the doc strings kind of do the auto generation of uh api stuff for discretize so got my way through all the utilities um made examples for for all them as well to to show how they work, and uh so that's yeah. That's a pretty good final draft for that and uh try to move on to documenting the classes in discretize. Next.

A

uh What branch are you working off of.

C

I believe I'm working off of something api docs, something let me find yeah numpy docstrings underscore api uh and I have I've kept up with.

C

I guess the the d prop that you had so I haven't done anything with pi data retheme, and we can. We can talk about maybe making sure that I'm I'm up to date with whatever the latest is. For uh I mean the package, there's no point in documenting things that don't exist anymore or are changed.

A

No, it's fine. I think I I got everything merged in relative to the new like the most recent release of discretized, so it should just be working off of it, but it looks like it's working off of that now. So.

C

Okay, so I'm I'm up to date then looks like it: yep yeah.

A

Cool um I'm looking forward to getting that in it'll. Be a nice nice, nice thing to have all the classes like fully documented it'll, be great.

C

Yeah well, and once once that happens, uh there's still uh tutorial stuff. That's been been hanging on so several months ago. We kind of went through a review of that and then realized. Okay. We need to make changes to the code and then that's kind of hanging around yeah. A lot of work has been put into that. So I.

A

Think the necessary code changes are there now to kind of finish off the tutorials like there was a that those.

C

Like longer tutorials that you had okay, I mean that that might be something I'd consider working on. I got a review um and there was a there's. A couple functions that I couldn't do to kind of make a comprehensive tutorial of certain things, but it sounds like I could do that now.

C

So it seems like something that just needs to get pushed over the finish line and I'd I'd like to bring that in pretty quickly as well. Yeah. Okay, um do you see any? Do you see any big concern with kind of working on those two things simultaneously on the same branch? Or do you really want to keep them quite separate.

A

C

A

Them separated, and one is one- is like directly modifying the code right. The other is just not creating tutorials outside of it, so let's just keep them separate for now. It's already.

C

A

A

Okay, soggy, I see you popped in here. Do you mind going through the quick report items.

D

uh Yeah so uh I yeah last week I worked on that 1d code, refactoring, so yeah. I finished it on my end, so I passed that over to joe and dev and yeah. So now it's actually using the existing source and receiver class for the 3d code. So like you can actually run about the same.

C

D

In a very similar way, so yeah, if anybody interested have a look, I think it's not it's not very elegant at this point because I had to dump all the variables into that uh the existing em surveys, so um probably a couple more iterations, but uh the mechanics are there uh and yeah. So initially, I probably need to like I thought like we need to sacrifice the efficiency but uh yeah. I didn't like that.

D

So I spent some time to actually optimize under the hood, so uh it's actually yeah it's actually better than than the previous version. So it's actually nice um anyway. So that's uh that's what I worked on and that's done on my end, so I think uh joe actually have you had a look at that. That's a variable offset issue in the rta function.

D

I think that once that's figured out, I was.

A

When I was thinking we're trying to think a little bit more about it and just kind of I think we could implement some smart broadcasting information like smart broadcasting there to kind of make it work in both cases.

D

Right, I think once that's uh solved, I think then uh the current 1d code can handle like. I think that all uh all airborne systems, I think, because that was the only exception. Yeah.

A

So right now like you, can definitely have it. Do multiple offsets and that's true. Well, you can't, but I think the sim like the simulation there is that it can do it'll compute right now, it'll compute, like everything needed for everything, but not ever, but it's not all needed right like you only need a certain certain numbers of them. So it's like.

D

Yeah, the the the thing that we're talking about in 1d, um like that the reflection coefficient is not depend upon that. um Your.

A

D

Offset as well as the so if your source and receiver are at the same plane, it doesn't matter so that doesn't change. So if you actually can combine all of the offset into a same place, you can just compute that reflection coefficient once then. uh So I think in such a way you can actually optimize the code. So that's uh that was what's related to that.

A

Yeah um yeah, I think we could, if you're talking about that, I think that we can work that together and figure it out.

D

Yeah, but it's pretty minor though I I think that resolve. Actually it's all same so, I think uh probably the most cases uh it doesn't really matter, but I thought that could be nice to have that that option um and a couple other things I met up with uh nick about the csmt, so I was working on theater about that frequency domain code and one of the potential applications that I thought could be the cscmt, so nick sims pretty excited about that.

D

So he's going to provide us kind of good csmt data, and what what's I thought interesting is uh the the setup for csmt is set, is basically using the mt codes.

D

So they don't really like uh the typical vendor seems like the contact and that it seems like people, don't really keep track of where the source locations were, because your interpolation interpretation method is the mt code, so they don't really care where the source flows.

D

It's a lot of historical data doesn't actually have the source location, which I thought interesting, so it. This is another case where your interpretation method kind of constrains your design of the survey, which is actually not good anyway. So I think we're going to get a data set from nick and we're going to work on that and related to that.

D

I think it was mike who worked on that the wire line source, so we initially implemented that, but still not quite available. So I made an issue and I'm going to work on basically finishing up that up and uh what's actually nice is the once you got that source.

D

Actually, the adjoint is the receiver, so you can actually have both wireless line, source and receiver with the same function. So I'm going to work on that and nick sims interested to using that to calculate the en effect in vip data, so yeah it might yeah. This is.

E

Something that I'm working.

D

On as a side project, but I'm excited that's uh where I am.

A

From what I remember about the line source branch was like it was already. It seemed like it worked like it all passed. It passed all the tests that I had locally, but for some reason I was having some weird memory issue on the when testing in the cloud on their computers. So I think I can go back. I need to run it through my my linux machine and see if I can reproduce that error and kind of track it down.

F

So I rehashed, like I, went into that branch and took some of that code and implemented there for this forward. uh Modeling for cs amt that I did there uh just recently, so it does it hooks in right, like perfectly yeah, it should.

A

It should work like I just don't know like it's: it's causing this random really other out there test. That has nothing to do with this to fail, and it just it just runs out of like it just hangs like the test, just hangs, and then it gets canceled after 50 minutes or something yeah. I'm.

D

Also happy to test joe, so I'll uh I'll, just like take a look and try to use it and see if I can make it that yeah.

A

Like I said it's, not it's, the test has nothing to do with the line source at all. It's just doing it right right.

D

Yeah then I'll focus on actually implementing the code and testing.

D

A

Just need to, I just need to figure out what's making that test available.

D

And potentially, uh if you guys are working on mt problem thinking about using that wire line source to generate the uh magnetic field, because uh in in that way, you don't really have to project on your electrical field into a magnetic field. So it could save quite a bit of projection part uh anyway. Just that it was just a thought.

A

I thought the purpose was to like use the the discrete operators, so it's still like divergence free when we're doing that like because, if it's like an exact solution, we end up with some like numerical dispersion or something.

D

D

Not sure joe, what you're.

A

Actually talking about, I just know, I just remember looking through all the codes like the time, the frequency domain source codes and seeing notes about. Oh, we do this. We do it this way, because otherwise, there's going to be a little bit of dispersion or something like.

C

This, instead of.

A

Like actually computing d.

D

I think that's for the vector potential joe, I guess for the like, because if you use the like a magnetic field directly for your source term, it doesn't uh kind of make sure your diversion is free condition. So by using the vector potential, we actually make sure, and actually that makes a big difference and yeah.

C

A

D

The magnetic source.

A

So it was the only thing I was referencing, then yeah, okay,.

A

See dom what we've been up to.

E

A few things uh but john uh before uh before I go myself so the the line source things that you've pushed on our tiled uh simulations branch. Is that uh exactly what soggy's working on uh I don't like it? It would've been in the branch I thought.

F

That uh mitch and okay, so you merge this in and then yes.

E

Okay, got it yeah, so the style simulations will we'll have to parse it apart and bring it in chunks because uh it's getting pretty stainy. But it's good.

F

I mean we're doing progress, but I've been uh yeah. I've been curious with the csa mt with diaz, so we've been playing around with it, and I just it just was so natural to go with that branch, because everything.

A

So is it like? Is it a matter if it's just that, you guys have been pulling other branches into it? Since that's all it is, then we can just merge those other branches at the master. It'll be easier.

E

Yeah, but also like there's like when you look at the files like there's, there's a lot right um most of it isn't tested, so we should probably just bring it in chunks and test. Add tests, for you know each part of it at once. You know: okay, like we refactor the receivers, there's a lot of desk stuff, that's going to be on its own, but um yeah, it's sprinkled.

E

So I think it will be safer to spend time.

F

Yeah, I brought all the mt in there too, so yeah there's, there's quite a.

E

Bit um yeah, so uh I just started the so I branched the the fork type simulations back to uh to the simpact simulations branch, and then I started adding um an example for for you, uh john, so that you can look at how to break up like the nt in the frequency and spatial. At the same time, I saved a little bit of time. I think you can generalize it a little bit more so that uh you know we're not using like frequency words, but more just like components.

E

You know basically reshaping the data and then slicing it right in a smart way.

E

uh So that's what I'm done now and then another one you guys that might have seen my rant on the grab my channel, but I reviewed a paper that will really could really streamline the mdi, um the sparse mbi. So I will be probably implementing a mapping to do this as be able to keep it in all in cartesian system.

E

uh So that's pretty exciting. That should that you really ease the uh the run time for the mbi. That's gonna be next week, though, and uh that's.

A

A

John getting right into that.

F

Yeah, nothing too much more to add. I modified the tipper receivers so that they can take a roving zed and a stationary xy. um That's all pushed onto actually that's on tiled right now, um not on the mt one, but I can pass that over soon yeah, it's already starting to be a mess and yeah. I was talking to dom yesterday and yeah, we'll probably start breaking the the tile simulations apart and get the test passing and get that hopefully pushed in the near future uh and then yeah I'll post, the kubernetes cluster instruction.

F

I just have to uh find the instruction that I used for the rsa keys just to make sure that I have it right. It's been a few months now so of course a little foggy. So I just want to double check that, uh but yeah.

E

That's about it, hey john. uh I tried either the uh with the um the station the. How do you call that? Not not the rover but the uh oh stationary, the reference, the right of a reference station right, yeah and the inversion really pulls to it towards it. Today, it's kind of like a.

F

It's a big sensitivity uh exactly.

E

F

To do that with our kind of doing the full mt with the airborne, and uh but we have that stationary e and x.

G

F

Yeah I've I've made like one or I just take the sensitivity from one frequency and use that as my weighting and that's helping a little bit. But I don't know the inversion at that point was still going pretty slow. So I haven't revisited that in a while, but yeah.

C

F

Sucks it right to it.

E

I guess I will need to test right to you if it really changes the the solution. If we put the you know if we compute x y at every point, so we really use the base station and I don't know.

C

It's like you seem to think that.

E

The fields were smooth enough that it wouldn't matter. I don't know.

D

I'm pretty sure that is the assumption yeah, but that is the assumption in general, like if you think about 2d code, it's impossible to put a right reference location, so their typical assumption is magnetic field is smooth, so they just attach magnetic field whatever I receive location and compute that deeper, I think maybe yeah like uh that. They have may have better idea, because I think they have.

G

D

This problem with the zam- I guess so yeah so I I don't know it seems pretty clear that if you have a same reference point then you got really high sensitivity to that point.

D

But I think the reference location is usually pretty far from your core domain. I think so it probably doesn't really matter. Even if you put like zillion structure there, I guess.

E

Why yeah, not necessarily it can be in inside the survey too, I'm not sure.

D

Oh really, oh, I didn't know that.

F

Yeah, I think they planned for far away, but it never really like it's still close enough that it affects things. The tests that we did. We thought we had it far enough away and no definitely not.

A

Okay, a little update from my end.

A

Update stuff for the dc problem, um I have working robin boundary conditions for the nodal dc problem right now as well, so it actually handles whole receivers very well, like I've just kind of looked at its little. You know it handles it much better than it did before like technically. You couldn't actually solve it with the pull receiver because it just violated me and violated the assumptions that you have to make. So it didn't work, but it shouldn't have worked like it worked, but it was not right, but now it's it works pretty.

A

It works really well with low receivers and just kind of in general, so it'll be a good thing to have is like the default boundary conditions travel to the these, for both the cell centered and nodal problems.

A

Okay, it looks nice, it has nice solutions.

D

That's that's awesome, which branch.

A

Are you working.

D

A

The one that I have a pull request up on it's like.

D

It like that's a simpac or um discretize simply.

A

Simply um yeah it then I added a few things to kind of cache some uh some of the matrices in td, so they build a little bit quicker.

A

Because they're dependent on the spatial wave number so just to be little inside things that happen that cache it there.

D

Awesome so then, that joe is that dc code is check. It's done like uh it's good enough to be published. Yeah yeah.

A

Perfect, we'll be happy with it. The only like the only thing that's out there is that it just the nodal problem with that boundary condition like that, that you know robin whatever linear combination, boundary condition isn't set up yet to handle like anisotropy.

A

But that's the only thing.

D

I think that seems fine, I think that's general assumptions, so we have the joe and I talked about writing up uh dc paper. I thought like that.

F

D

In quite good shape um and um yeah, like kind of talking about versatile, multi-dimensional dc code and can handle various types of regularization functions and kind of different types of inversion approach, I thought that could be quite a valuable paper, so yeah we're gonna, write it up. Joe will read, leave that so, where are you? Where are you thinking to send them?

D

I don't need, like uh maybe computers and geoscience, say something like that. That would be better. I think.

A

Yeah, I think that's that's kind of the idea, but it's actually neat because with it so with the nodal boundary condition with a like mixed type boundary robin whatever.

A

It actually has to go into the derivative of the matrix because of the way that it works out. So it's it's! It's interesting, yeah, the.

D

Yeah uh likely we probably kind of emphasize that using filipino arm and parametric inversion and that types of things can then emphasize the like ability to handle all dimensions. Yeah, I think that's pretty unique. I think only pi gimme can handle that. I think.

D

A

I I got the presentation recorded last last week that I was talking about. I think it'll be well, it's going to be given.

A

I think the date is next wednesday. um At that part, parallel cfd conference we'll see how it goes. Just kind of went forward with recording so it'll, be interesting to watch myself back again, always weird.

B

Can we actually see that show or do we have to register for the conference? How does that work? I.

A

Think you'd probably have to register for the conference, but I'll look more into it and see. If I can share it a little bit.

A

A

Tivo, do you have anything? Is he kind of started, writing a name and then nothing.

G

Oh really did not update. Let me check the notes. I thought I would go to something, but but anyway, I just wanted to say that uh I done my cakes talk yesterday. I think it was successful. At least we have sparked some conversation on the simplex like. So that's good, like uh some conversation from uh from nick in the threads, are also like. I saw this morning from.

G

From workers about like her community engagement and uh documentation and tutorial devon.

G

So yeah, that's that was about that's about that's about it. For for me, for this week was a yeah. I think I think it was a a success in terms of like attendance and the bring a message.

F

I think I heard it was the highest attended so far it was. It was.

A

A

There's any anything else, anyone wants to chat about check through some things.

D

Oh, uh I put a link. uh Some of the airborne data that we have worked on is actually publicly available, so yeah. Technically you can download and use. I'm not sure the data is in the form that is useful, but.

A

It's there, so anybody interested actually.

D

Can use it so have a look, it's probably a mess. I didn't. I didn't want to archive that data. So it's.

A

D

But the data is available so is.

A

D

A

It's all of the data.

D

Like all the well data airborne data, whatever the data they have collected throughout that project, so yeah yeah, it's actually expensive data. So anybody interested have a look and use it because it has a lot of money that they spend for this project.

D

A

Well, then, is there anything any anything else.

E

uh Shalom question for you, um so uh how far are you in your cross gradient? I see that your pr is still open and hasn't moved since uh december last year, right yeah. I would be interested in probably starting using the cross cross gradient at some point. So do you need.

C

E

Do you have to push this ahead or.

G

Also, you means.

E

Oh sorry, I finished your the pull request right because, right now your branch is behind. uh It looks like it's, it needs to be updated and then we could help. You uh maybe merge this into uh into synthetic main.

D

E

Yeah that sounds good yeah. So would you mind just updating your your pull request right now to make sure that the tests are running and then uh we can? We can have another look at it, maybe yeah. How does that sound yeah? That sounds good, okay, sweet, because I I don't think I I think we don't want to lose it like. We want to make sure that sky comes up yeah, it's another better. For me,.

G

Yeah I'm just gonna mention too, that the pgi has been passing has been passing tests since december as well. Oh.

E

Yeah, I can't hear an adam, that's crazy, yeah now, they're failing it looks like but uh yeah just.

G

Yeah, I don't know it's like updating master. He was passing test before and I think now it's just like missing like a 0.01 percent of coverage, but he was passing before so. It's like.

C

G

It was passing test before and now updating master is not passing, so I don't know just.

D

I think it's just a coverage of.

G

Like a few 0.1 percent.

D

So like it's do we need some sort of protocol here, because uh I can see I mean it's all of our quality um like what's actually our protocol to to merge a pull request to a main like it seems like the minimum is an approval from a reviewer and uh like, but like do we do. We have actually a protocol to merge something into maine, and this is something that we probably need to think through a little bit more and at least.

A

With the smaller pull requests because it's like- oh it's, just a simple line, change or simple changes here and there's easier to understand easier for me, but it just takes larger. It takes longer with the larger ones for people to look through them. It's like okay, well, there's certain things that certain people have access to. I'm.

A

I think it's at the stage where we should probably just pull in the pgi pull request as well as so that way we can get the cross gradient. One updated to be relative to that one. I think, that'll be that's what we should just do when we'll work on that I'll work on that uh today tomorrow.

A

I think the few tests that were failing on the pgi now are related to the renaming of master to maine.

A

Looking at it. So.

G

A

I think I had that handled in my dc updates. I just switched it around, so we'll get them in this week. We do it.

E

Right so yeah, you're, saying pgi first and then we'll we'll update cross cross gradient, oh yeah and xiao, along with the cross grade, because you you're working on a fork right, uh xiaomi you're, pushing stuff from your from your own repo. Is that correct? Yes,.

A

That's correct.

E

A

Yeah, I think, there's just the the structure of a few folders and things changes a little bit from what I remembered yeah. It was just a matter of updating that keeping it consistent.

A

Okay, then, let's get to it.

A

Two times up, all right is that it that's all I've got for everyone.

E

E