GeoSci.xyz SimPEG, 8 Sep 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: SimPEG Meeting September 8th

Description

Weekly SimPEG meeting from September 8th, 2021

A

Hi welcome everyone nice to see you all again, it's another week here, a few of us are back in the eoas area.

A

Just here, anybody wants to join they're, welcome to join at the big table.

A

I've got my first class today, so I'm excited yes.

B

A

About where liz was joking with me this morning about taking one of those like first aid school pictures where you like have those uh like people are holding up their signs and we're.

B

A

Oh who's, your teacher this year and I'd be like me.

C

Yeah we could get a uh eight foot by eight foot blow up and put it in the gif room. That would be cool.

A

Oh anyway, so as always, if you've got some quick reports go on, go ahead and add yourself there a couple agenda items just we have another syntax seminar tomorrow which, how long at 2 pm pacific time from correct pretty sure that was the time um lindsay, got the.

A

The newsletter sent out, so you guys all probably saw that this morning or yesterday morning, yeah we all saw that come out. So just a few quick things from there.

A

Lindsay, I think we could probably.

C

A

About the um you had sent out an email about them from mtnet about the csamt code.

D

A

D

Home for that I.

A

Will talk about that too? If you don't, if we can, I don't think we'll have too much else today.

B

That sounds good joe.

A

As far as me for some quick reports.

A

um Not too much sim, peg ish related a few little things here and there with the um with the tetrahedral mesh stuff, the simplex mesh it'll just little pieces here and there let's implement for it. It should be pretty good to plug into simpeg. So I'll.

E

Be excited to get that.

A

Plugged in and trying it a few things like a few more little interpolation matrices, a few more um like I need to implement like a cell gradient stencil, so we can do some inversions with it just simple things like that, but it should be nice. Interpolation thing is just a little bit more work than everything else like being able to arbitrarily interpolate it.

A

I've got ideas about how to do it like the nodal base. Interpolation is easy. Edges.

B

And faces take.

A

A little bit more.

B

A

Self-Centered interpolation is can be challenging to implement properly with, like second order, convergence.

E

A

E

But joe like do we really need an interpolation for that mesh because I think the like the beauty of that unstructured mesh. You can actually define the node at the data location. You want it. Yeah.

A

So hopefully not, but it should. We can support it and it's like a lot of the code and simpeg uses that like uses those matrices under the hood to like interpolate to receiver locations right yeah, it.

E

A

Not be super important that they're, like second order, convergent or anything, but we just need to at least have them there. I need to be able to call get interpolation matrix right right.

A

Than that I don't really have too much synthetic related over the last week.

A

Does anyone else have any other quick reports things they've been up to.

D

Not too much but maybe just uh check in on. If we're gonna try to merge the uh discretize api stuff.

A

Oh yeah, um let's so we keep talking about it, we're gonna get it in. We need to set up a way to should we just do it on twitter or something I don't know where else to do this thing, where we just choose a simple logo.

B

Well, do we have like a menu of logos or a place? I mean because I think we could do it in slack like and also put it up on twitter, but I don't know if there's like a document right now that we have that everyone can look at and see like this. One is number one. This.

A

B

A

I mean I just have those three options that I put out before.

A

um Anyone has any other thoughts or things they'd like to see. Let me know I'll put it on there. We can start that up this afternoon,.

D

You do it as a google form, just say like which one do you like a b or c, and then we figure out what the favorite is. Then we choose that.

E

Yeah democracy sure.

B

Do you want to do it as a google doc? So if somebody does feel creative and inspired this week, they can add something there and we could do a little poll on slack.

A

That sounds good to me. I'll, put up a google doc and put those three things in there that I've got so far and go on.

A

A

Anyone else I know it's just if you want to give us some quick updates on what you guys have been up to too.

E

I don't mind that.

B

It's been a while.

A

Since we've seen some of you get some of you.

F

To try to deal with the normalization of the cell volumes, but I think mike is going to be uh it's going to be touching it because I didn't look at the uh at the pgi part of regularization um and then yeah, that's going to give you the starting point to refactor the whole the whole classes.

F

But it's a start. So if you want to take a look joe or anyone else, do our review.

A

I feel like we need a nice way to just ask it for.

A

Like we should have a thought about where that volume normalization thing comes into play at because, depending on like the maps right, it can be weird where it should actually be.

A

It's like sometimes like it's. If it's an inactive and active versus inactive cells, it can be odd if it's a. It can be an odd thing if it's.

A

I mean, I guess, the regular.

F

Volumes the rag knows about the active volumes we just.

A

G

F

Yeah, look at the pr and tell me what you think yeah.

A

Yeah, okay, then.

A

So that should be enough, then just having the reg mesh, pull it from there, because that should be that's the mesh that we're doing the regularization on right. And it should know it even in cases of like if we're doing parametric, inversions and.

B

A

It should know what mesh we're on.

A

F

A

F

Israel in the sensitivity weight directive, where now it's just normalizing it on on the spot, rather than expecting it should be.

A

E

Okay kind of speaking of that would it be worthwhile to actually write up them, uh because I think I just still don't exactly understand what the sparse inversion like does because, like there's a subtle point, whether you take into account the volume or use a stencil, um maybe I think this is probably an opportunity to like at least write that up what it does at this point and make a clear distinction like how it's different from thick enough, um because yeah, I think uh yeah, maybe most of the user, wouldn't know. I'm not sure.

E

That's that, like they really required to know. But I think it would be nice to have some place that we kind of document that and what provide what it does.

F

I mean we talked about this uh multiple times already, but uh basically the idea is that we can collapse all those organization functions into a one and then the only difference with tknav and and sparse right now is just that auto normalization for alpha s right, so we can collapse it all and just you know, throw a warning: if you use norms and on normalize, then you should throw a warning to the user.

F

Otherwise the t can have is just all normalized and the other one is normalized. That's that's a it's. The only difference so there's no reason to have multiple classes, uh it's redundant code and it's all pretty complicated right. So that's that's sort of the point.

E

Like uh then, if that's like uh generalizable, then, uh let's actually like, would it be kind of writing up that that, as a like a small document.

F

Throughout the class so that it's obvious what's happening.

E

A

F

Classes and then make one regularization function. That's it! It's gonna, be much simpler for everyone to understand. Yeah.

A

Yeah I see or if anything like these subclasses are just minimal changes, just minimal, necessary stuff to debate to the base class.

F

Gi, you know, as is a subclass, and I'm probably going to do a vector regularization instead of having to build it every time, just you know, stack multiple regularization on top for vector models, but yeah those are just subclasses. I mean.

E

Yeah yeah, okay,.

A

A

G

Getting that is my microphone working now here. Come here: okay, yeah uh yeah! I didn't get to anything. While I was on the field, I thought it was going to be a little bit of a break, but I ended up having to lead the crew so that didn't uh work out, but I'll get uh I'll get back into the prs and the cold cup stuff here shortly. But I was back here at the beginning of the week and started diving back into the parallel computing. I got a keyboard.

G

We have the kubernetes cluster now set up um yeah. We want to compare that to the mpi solution. I got the mpi solution, like kind of what we talked about there um before I left there, joe and now the slow part is having all the, I guess, all the tech or all the pros processes, loading in the data at once from one access point over the network. So yeah sorry.

B

There's always.

G

A bottleneck somewhere um yeah we're just going to see what the kubernetes stuff, just by passing it around not normally just how we have it set up just to see what see what it does should be interesting, but yeah more to come on that now's been back, for it should be for a while.

A

I hope everything did go well down there.

E

G

Well enough, it was a lot steeper than I would have hoped.

A

A

With that, I think that we can have a lindsay if you want to chat about the joanna mt or csa emt code that they sent out on emptynet.

B

Here I don't have too too much to say on this other than I think it really. So I don't know if folks saw that first, I guess. Let me just pull up that email.

B

um So the what prompted this is richard smith sent out an email saying um he has peter weidelt's code for a thin plate, a thin dipping, dike and sort of you know. Some of these solutions are really useful for just being able to test things and also maybe even perform kind of simple inversions, um and you know we've got kind of the em1d um components in simpeg and and geoanna and just kind of wanted to prompt. This is like what other solutions do we think that we should have.

B

That would be really useful to the community. I mean we can come up with like a laundry list and prioritize. I don't think we need to start like a massive effort to just be implementing everything today, but I think it would be useful to have just a bit of a list of like what what are the solutions that are sort of classic and really useful and things that we should have.

B

um And then we can kind of as time progresses and as the you know, we have applications that could make use of these solutions just go ahead and implement them and um yeah. But I was thinking just just something to track that. We could even start a github issue or something like that, um but mostly just wanted to prompt um yeah prompt, a bit of discussion and thoughts on any solutions that we should have.

A

I do like having these like simple solutions in there. um It's it's! It's nice, it's good to test them against it.

A

Some of the other things I thought about adding to it are like potential field solutions for like simple bodies. I think those would be useful because.

B

A

Just kind of like.

A

And like rectangular prisms are great they're, simply simple. These fears are simple, like just simple bodies that we can, we can use, and uh my other thought there is that, since we've we've got a bit of compiled code inside there now for when we can't like when we make use of it, we can do the same thing for some of those like, like. Let's say we have like a.

A

We can do a fast potential field, modeling like we give it a bunch of node locations, and it calculates the components for those instead of doing them in python right now and then, like the specialized function, would just be okay. Well, this it knows how to use those things to do the cube for the rectangular prism we just have have it called that and have like more general things that we could potentially call from simpig. That would call that compiled code to generate those things faster.

A

Let's, like try to move more of the analytics out of its instant bag to there.

B

It would be worth taking a look at what's in fadiando, so that we're not duplicating efforts. I think they might have some simple prism solutions, and so we might be able to either reuse or or even sort of contribute upstream there, but yeah totally agree.

C

On the em front, uh always that conductive permeable sphere in a half space is useful both for frequency and time, as well as for plane waves and a dipole source.

A

I have a I know I had. I had a full like a half space, anisotropic dc resistivity or layered, half space and isotropic dc. This is typically modeling, um so I think that I mean it could be useful. I just had like one vpc resistivity modeling in there. I don't.

B

A

B

A

I'm pretty sure it's not in there, so that could be something useful to add.

B

Yeah so maybe in terms of next steps, we can start an issue and maybe link a google doc might be easiest just so that we can sort of capture links and just have it be pretty editable right from the start, and then we can maybe turn that into something. That's a little more actionable later on.

A

Anyone have any other requests for things to put into geoanna.

E

The thinnest solution is actually pretty like appealing to have it's actually complicated, I'm not sure it's relevant for geoana. You actually need to solve this integral solution into the equation, to have a general thin solution so, but I think would be. One of this is one of my bucket list tool like something that I want to have, but it would be great to be in uh joanna, but the question that I had like there's some like, so it's actually complicated and there are some old fortran code.

E

um The question is: do we really want to rewrite like get that from scratch and then turn it into python and kind of store that as like archive as a joanna or just just gonna, just want to revive that old fortran code and compile it and use it whenever we want um yeah?

E

I think that what like that email, it also brings up the same question richard smith had like okay, we've got here's fortune code that I wrote um do we like if we want to rewrite that and kind of have a fresh python code, kind of doing the same function for that or just kind of wrap up his fortran code and somehow compile it and use it in joanna.

B

That's a really good question sargeant. I think the only potential challenge there, which is probably can be navigated, is how it's licensed, and it's probably, and if we sort of have permission to distribute it or not.

B

um It certainly sounds like we can get access to it, but I don't know kind of the ins and outs of us like distributing that with an open source package when it was yeah different licensing conditions, but that's something we should we. I mean it would be useful to get that code anyways from richard smith, and so we could see if there's a way to prompt that conversation.

C

Yeah well, since it comes from richard he could he might have some. uh You know comments on that, but basically this code is kind of freely distributed. I'm sure peter would have liked that, uh and most of these codes are really old. I mean we're talking 30 40 years, at least for some of these these codes so yeah.

C

I can't see anybody would object, but I guess that probably need to dot eyes and cross t's just to make sure that nothing, you know unexpected happens, but I I can't even foresee a possibility of that for the for these really old fortran codes.

C

Yeah can't see it.

A

It is getting like it's something like we could think about just writing four times like wrappers to call the fortran functions. I haven't seen the code. I'd have to look at it to see what we could do with it or what would be best to be done with it, but if it is all old, 1980s fortran code, it's likely you know 4chan of 77 based. So it's just very older, like you, just get older versions of fortran, make it harder to interface with newer things.

A

A

um So, like you do bring up another point that it's like an integral based solution, identical equation, based solution right.

E

E

Yeah, like I have some connections with like some of the old like korean em geophysicists, who actually wrote like some of this like analytic functions and they probably like archived on their old fortran code and somehow. I hope that I can capture that because once they're gone, that's gone so and actually it's a lot of work to rebuild that that analytic function is not necessarily trivial. Like.

A

E

Even harder than writing this 3d code, so um yeah it was my hope, but not exactly sure, what's the best way, but the first time I think basically asking them, and can you actually provide the forefront code and accessible to everybody? That seems a good start and and if needed, we can pro. We can rewrite that maybe. But uh that seems a good starting point. Collecting and archiving.

B

Well, that seems like a good step, I'm happy to write an email to richard and just kick off the conversation and see if we can we'll first get access to it and then be thinking about distributing it.

A

I mean I'm fine with including integral equation stuff in there in general, because we've got like other semi semi analytic based things.

E

Yeah yeah as we go, I can also contact some of the people that I know and uh they probably it's old code. That stuck said like at one point: it's it's probably licensed to something, but at this point they're probably willing to open source it. I guess in general. So I could. I could also send a couple of emails too.

C

I'd actually be surprised if it was formally licensed by anybody. You know it's just you know you just have this code, you keep it, you keep it under wraps, you don't distribute it, but uh yeah license. Agreements were probably not really very involved. Those days.

A

Okay, well, I will I'll be posting links that at google.com, probably post it I don't know, is there a deal? Do you have a doana channel on slack.

C

A

G

It's also in the agenda.

A

There is a link to it for now.

A

Anything else anyone would like to chat about.

A

If not, I think we're good to go until next week.

A

So have a good week: everyone have a good weekend.

E

A

E

C

Have a good lecture this afternoon enjoy remember you're the teacher and.

A

C

E

Good luck, joe bye, all right.