GeoSci.xyz SimPEG, 6 Feb 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: SimPEG meeting Feb 6

Description

Dikun Yang presents how he would like to use SimPEG for Marine CSEM inversions

A

Yeah, for some reason, um by on my screen, I can see like everyone. Clearly there is something yeah I, don't know. It's maybe just interested see.

B

That just eat one, it's just because, like it's just a tiny webcam, so it's not I'm gonna have everybody. Okay,.

A

Yeah, no, it's better I! Think just for one moment you spent like retards yeah.

B

Yeah, just so, you might see it's pretty soon minute.

A

Okay, those are showing estar en are yes,.

B

Sounds good, do you want to stir this up bye I know, you've been looking a little bunch of things and you got some specific mine.

B

B

A

So maybe I made a couple slides. Maybe I can share my screen and see some stuff. mmm Where is the.

B

Mouse over on the left part of the street, you should see a side bar pop up. Are you sure.

B

Yeah, no we're just putting you on the HDMI audio. So let's know if it's just echoing like crazy on your end and I. Will okay.

B

A

Can you see something.

B

Are you hearing yourself here is that okay I.

A

Think I'm, oh my.

C

A

So, do you hear me better.

A

B

A

It sounds better yeah, yeah.

B

A

Yeah, so can you see my screen? No, it's.

B

Flat right now, my trying.

B

B

We can see that.

D

Give a comment: are you sharing the other screen that can you move your thing or make a qualitative screen.

A

E

Right, we see it under the movie presentation.

A

So now the in full screen right. That's.

B

Right yeah, so if you don't.

A

Okay, yeah, that's fine! You can see this right, yeah, yeah, okay, so yeah! So this this.

A

You see this stuff on yeah yeah, so basically our heroes have been trying to do some rhenium stuff together. Some you decide, inverted like IP or links ready. So the only cover here is the picture. I took from ESP and.

E

A

So the era so yes B, is for Empire State Building.

A

Yeah, you are in same place and that's where I got most of inspiration.

A

Okay, so in terms of solving some marine says he and problem, so here's the basic diagram showing the basic ideas, for example, have a ocean bond receivers, so those are the electrodes for a field and also sensors on the yellow for the magnetic field. So in particular, for example, you, if they're doing like a large-scale subduction zone or Mayo, should really study. They may want to use a empty source because it can't penetrate much deeper at the same time, becomes an empty source and can only do like no freaking stop for shallow or near seafloor.

A

They have to do all the control source here. That's why they have also have the boat touring of transmitter am transmitter and for the hydrocarbon exploration is mostly just the CSEA yeah. So, basically, in terms of the demand you need like a CSI, a modeling immersion and also empty for investigation at different scales.

F

F

A

Which one the tongue, which transmitter or receiver.

F

Another legs, a Lexus, the square- that, besides the long waistline navigation, is that the shoes to right at the surface. Yes,.

A

Oh I think it's a.

B

A

Have no idea, maybe yeah I, don't know what it is. I just got this picture from Google. Yes,.

E

A

Could be some navigation system? I don't know anyway, so for this kind of problem. So basically we can talk about three difficulty or challenges. First, first, first of all, it's a tow system so usually is covered, like hundreds kilometers, like that kind of scale. So it's really large.

G

Those large scale.

A

Serving and also, we need a jolly version because for.

F

A

Shallow partner can only use this as a cm, but for deep part also need to include empty information, so, ideally should be able to MPN, see SEM Johnny version and the for CSUN part that can be freaks to main or time to me and so the in reality. People have track the folks of Rick's command and have to make them very boring.

E

Yeah, it's usually frequency domain right, there's system.

A

Okay, so 95 percent freaks domain- yes yeah, right and maybe uh I- don't know five percent, so they're expand experimenting. Some time domain system, though, carries doesn't make some test, so you were able to see the tandem. A system is it's better to can have better tsukumo's ratio or anything because for some reason, freedom a system may lose this information for high frequency or after so the phase information is basically missing.

A

Do you have to do some time to my system so anyway, so there, the yeah in the future. It could be time to my system and also brings environment is highly stratified. So, of course anisotropy is a big problem need to think about yeah. So this is especially true for doing the global and local Mazda company in the forum only match the little match. We can have a large sales to represent a small sales in an immersion match.

A

Tell you now into the lair in the seafloor is highly stratified, so the obscured connectivity has some anisotropy.

A

So next is framework track. I've been thinking about I caused a D cubed approach, because that means decoupled this cubed and it is synchronized. So here's diagram I born from the impact framework, so I'm sure most of the stuff on these slides has been taught about or even implemented using this impact framework, so I'm just repeating maybe some existing things and just saying something: I'm thinking and if you guys have any idea, for example, here after done this or is going to do this, maybe just feel free to interrupt me.

A

Yeah, of course, are all gonna have a decoupled for the modeling version. This has already been done in the simpad framework to have fourth simulation universe elements, these parts perfect, and so also importantly, here's have a universal and the cross-platform interface, so I think I've seen has done pretty good job in doing making the interface very universal. You can change different for the modelers by just plug it in and use the same immersion energy. So that's good.

A

Another salt, in my mind, is maybe have to think about how cross-platform interface you will have a, for example. Your inversion entry in Python, maybe one day you want to use a Ford modeling code written in different languages. Is this especially true here because most of our carries code of an element code, the one to the or 3d or unfortunate, so changing everything from a fortunate to Python? It's not going to happen in a short time. Yeah.

B

So since the port simulation actually is very well decoupled from a virgin, Python is a very good language, meaning that it interfaces nicely the other software. So it's like that that step would require a light wrapper to interface to port, handles. Ok,.

A

B

A

Good good here, yeah, it's good and yeah, so just uh ideally I want. It is pramook to be able to communicate with all basic existing for modeling code as long as they obey the same interface or same designing or like API whatever. So you can talk using the same language, so the yeah I think that's good and something next I.

B

A

H

I think they do have everything all the necessary.

A

Ingredients, oh yeah, but the thing is that these things are kind of native unfortunate and the things are account like only connected internally, so there no external API to be called. So that's something need to work on yeah and the next is what if we have some Cordia coolers, for example, if we have a match last nominee, we don't necessarily have a match or yeah.

A

Just for example, I recently have been looking at using machine learning to do the for the modeling swinging like hey so where I was now actually solving the PBE search base is just a training set network and eventually network can um take any input model and spit out response. So there's no PD solving, so all kind of like exhorting modelers, maybe just make a uniform interface, so they can be called.

B

Problem with problem I can send you the link I'm, not playing once that's implemented, will be.

B

A

That sounds good, so the next item is atomic decompositions. This is something of telling my thesis as I. Consider this is the most important finding is that if we want to the best efficiency we have to break apart of the whole problem as fine as possible as fine great as possible, which means we Malo transmitter-receiver pair at only one frequency or or at only one time channel for each subproblem for for best efficiency.

A

So this might not be difficult to implement. Maybe it's just yeah I think for the model is just only model one frequency or one time, probably simple. An accessing is called a fully non conformal global, a local transform. So this cannot become necessary because, if you think about like urban emc problem, so trans means you see the same location. So one making a local mesh it's around a table is 3/4, but for like Marines, yeah or CIC M, the transmuter is sometimes it can be like a four kilometers apart.

A

So while making the mesh me like a locally, we find the saying or yeah, so the the global local transformer can sometimes it's not straightforward and also, if we take into account and start to be so, the I believe there must be some sort of work need to down. In order to do this, because the existing seeing a formula formula for me is to do the volume based averaging the water based averaging, doesn't take into account the geometry of the fine cells or fine /, fine structures.

A

So, for instance, if it's horizontally layered, so there's anisotropy. We need to do if we're going from global to local.

B

So we're the volume average global local set up- it's not even a course in paid, yet that's sort of been thought through and so extending that to the case, we actually want to upscale into an anisotropic, but we can sort of show you where that contribution would be. Okay,.

A

B

A

That sounds good and.

B

Then could and isotopic modeling is hooked up. So as long as you know, like long, we create steps.

B

A

That's good and now another thing is the pricing non conformal, which means that the global logo can be in different coordinate system.

A

There might be some rotation, especially if we sing, for example, this picture the transmitter up here, receivers here with an offset post horizontally and vertically so, depending on what kind of mass you're gonna use. For example, if you use the oak tree or like unstructured mesh, probably you don't have to rotate anything, but if you use a tensor mesh or structure mesh or they have to do the rotation to minimize number of cells in the local mesh and another point here is to invert out data without warning of them yeah.

A

So the idea is simple, so you just the early iteration you only. You were maybe model 10% data later on you're doing the fine skill structure, you gradually add more modeling to to to the inversion. So this is correct, multi-level or multi-resolution see so, but if I think before I've done the decoupling here, so we have a full control over which data we want model.

A

So this is only the small part of the scene, so I'm just laying out everything actually thought about and then I'll do the priority later.

A

So what I'm saying here, just ideally I wanna code like this and but to make it happen, it's another story. I'm just doing my imagination,.

B

Places for all of us and there won't that mean they're modular pieces. So it's not everything from life services, okay,.

A

Good yeah I figured for most of stuff. People have things all about it, so it basically just working on existing scene already just to make some Corrections or making some changes to make it happen, so it won't be too hard. So the next one is I. Think is more so it's already a big deal so so now on the right I'm comparing two different paradigms here. So why is a server-based another p2p network? So the current framework of a parallel computing is at least for eme. Worthless.

A

Gme inversion is for this kind of server based thing to have a master thread, dispatching jobs to different agent workers and the ones work it down. The local mesh or local problem to send back the results to the master and the master gonna do a model update and the send of updated model back to them and they work on just iterate. So there are lots of overhead in the computing. There are lots of issues in the load.

A

Balancing, for example, won't do the dynamic scheduling or static scheduling the reference counts, all consideration, so another thing I've been trying to do is do like a p2p based inversion.

A

So all the p2p agents are equal, there's no master, so the job is done collectively over the internet work. So the big motivation for doing this is because a synchronization is the bottleneck of scalability in 3d Alaska inversion, just because it's a master need to send all the information to workers and the work and the once is the information sent out the master. Just wait, they're just waiting for everyone to return without before some reason a couple of the workers slow.

A

It just slowed down so everything and just because everything needs to be perfectly synchronized and so how to do this quench early so for move everything to a p2p based network, so every agent p2p to the agent or is possible and replaceable. So there may be some duplications here, so this one is for maybe are pretty much doing same thing and they are just updating the the mod or updating version without any gap. So once this one is down job, it can start working on the next scene, and so the nice thing is about this.

A

Nowadays it can be highly redundant and the resilient. The first marine, a couple notes, get got crushed or still emerging still Ronnie, and if you later on, I got an additional 500 notes from Amazon. I can use them as an additional resource to speed up my work and if you later on I, have a tight budget, so I don't have a ghost anymore. I can take them off the inversion, store, Ronnie somewhere else. So it's basically just making version, especially this is a geared towards mostly for last killing version for a really large problem.

A

So the in the next time, I'm gonna show you example. Just for example, only world a look I mean her Canada. So what's what's going to happen, I'm showing here is a map of a West Canada Post, alpha I, don't know everyone. Young data across in her Canada, so also a lil agent, are only gonna. Do a regional inversion, for example, I'm now picking area around can loops. I won't do the inversion around can loops. So this agent gotta look around because it can read the information for this turning lot so the current model.

A

So it's just fun of oh there's, a agent working on prints charge agent working on Vancouver agent working on Carberry, so the Ken loops, just communicated with those three cities and get information in this region. Then the Ken loops can start working on updating the model around Kennex, so the ones that canvas monitor is updating, and so the agent that was working on category it's already done at the Calgary. So maybe this agents now moved to Edmonton so for the ad.

A

So, in order to get Edmonton word, he don't need to look around for okay information from a4 McMurray is the information from Carol? Oh that guy just done kokujin values. I can just really talk to the calyx and get updated model and admit the can. Can you know working so you can see this one can happen crossing an hour. So when those ages are working on the West Canada, there are a bunch of other working, Toronto, Winnipeg or some Jones.

A

So on this way, so the network, basically working on the inverting, the entire Canada and there's no waste time in waiting.

A

So the load are automatically balanced, so air, all the activity can be lodged in the matter in the in the matter very similar to proxy, so everything can be received.

A

This is just a very brief description of what can happen in this kind of framework.

A

So that's about it, but maybe just move for.

A

P2P is a decade's old technology in like computer network, but I don't know any applications, your physical inversion, but there's tons of application in just transferring fire files across networks.

B

A

Okay or try the gate idea, your yeah.

B

Do you have an example or something.

A

No I don't have any now, but but I know the technology has been existing for a really long time. Maybe there's some thing we can just from them.

E

If I think is just a synchronized problem, is that basically saying it's like anything think about a certain base? Also synchronize for a synchronic is like that. If there's one worker, it's not working, we can force them to work right. So if there's no option, you can be it's a knife or a synchronize, so that is that still super beast or that's that we do those.

A

Yeah, it's ok, repeat your question again: can you speak slowly because uh you're basically I'm hearing a lot of the echo here, there's.

E

Two option or a synchronized so like like: if you run a parallel algorithm, there's an option like synchronize me: okay, you can ask one worker or multiple workers. We need to wait until all the workers on the job right. Okay,.

A

E

The other option we synchronize like, if you just say one of ours, just do whatever you want and then boom stop this work right as.

A

So the most important see in PvP is there's no master, so everyone contributed to an hour and the the whole model existing only on the network. So any single agent here doesn't have the entire picture. It only has the part of the picture, but if you can somehow together, they have the whole picture. I think this is the most important part for p2p. So this allows them to communicate like from A to B or to see they don't have to all talk to the central master. I guess.

D

The question, because is how you end all that you see somehow stored final answer right like it's like chain, you don't even like yeah,.

A

I think I think it's for the broad chains of the way to record who that's what or who did, what and the actual model are. Actually dispute is across an hour.

A

Those nouns of the agent here has an entire model, just because of all those large just talking about a really large a conversion and the numb of the worker. Now agent here has linked our data set.

D

Yeah, given what happened if you started with the hundred word, carries you ramped it up a thousand workers. You mentioned the end of the tour. Here's now on everything is living on two workers like housing and building at Imperial, City avatar through yeah.

A

So there are not duplications, for example, so can loops. It could be like a five workers working in their area. Close candles, we're gonna, do their own model update. So while we can do this to use a temp steps to make the priority or like maybe just a you see Marie this case, just average them.

A

So there are lots of. You only need to be figure out.

D

B

C

Motivating eliminating example that would have cropped up. You know Atwood solar, something that kind of an anchor ourselves is to what we'd like to do.

C

A

Think maybe some large large air brownie MD set, for example, has a wider pepper.

A

That was very large. That's could be example, and also I know. There are some like Skycam data from damn are basically, however, I have the country, or something like that.

A

That could be good example here, I'm just making a very extreme example showing like a hurricane, our West Canada, but this is just a scale, different skill, soul,.

A

So here are some questions, so if we want more for it, my first question is: was a current status of an PRC SEM for assimilation?

A

Okay, something in these columns.

B

We have empty code, sorry inscription like we have empty and we have the ability to simulate, see SEM hooked up, there's a bit of cleanup that needs to go on in the MT code. It has a bit of a different feel right now, when the frequency doing am code not too much, because it there's a few things that needs to be cleaned up there. These are hooked up on tension, wretches authoring rushes at like any match, that's in disparate eyes. At the moment.

B

Right now, we wouldn't necessarily like we're working on getting to larger scale. The code is not currently paralyzed, so that really does limit the scale of what you can do with it right now. To put it I'm sure, you saw last week's meeting that we're now sort of in a place where we can start to think about paralyzing, and how do we go forward.

B

We have infrastructure in place to do joint inversions. That's not something that's been hit on too hard, but.

B

A

Because I've seen some the the the documentation web page I've seen lots of a 1d scenes, but I don't see 3d example or like the the stuff Salzer through the exam, because I know tebow sent me something he did for the time to me, yeah that just he boasts his own work or like is there gonna be merged one day or people are actually on the development branch or artist.

A

F

What I send you was good because you, you asked just close your burrito if we can use the operative separately, so I send you caller ID I didn't, and that was only you lose your burger for these pitayas, but time-domain en easing this impact. So I don't choose my own legal project. Oh.

A

So he for for myself in terms of getting started, I, basically just need a it's just you miss has been done. Omg RCS, you Evelyn for our sensitivity, calculator.

E

Big place to start is it. There was some problem, specific I think sin that was and a developer a warned in the beginning, so we got most of stuff for the born and then maybe go to pit of lava source, but not me for a cassette, there's some specific things right. You can take like that and the primary field, probably a to later area and the see maybe see same as well as a primary field- I'm, not hungry sin, sure how that's been not. There are quite a bit of details of policy.

E

Canisius iam works in the network. Simulation sense. I think that details may not be in need impact, so I think that you probably could map out how that really works and then I think making that for simulation. Former easiest, it really works. Its impact might be a good starting point. Okay,.

A

Public becomes on the new to CPAC, so this can be different. So I have been using the Julie package, a Julia package to solve something like the arrow and Marie.

B

A

Pablo so there's a framework of a Julian code. They are very similar, so you just implement a different source.

A

Yeah either just find out a line source or just a loop source, so loops are introduced where four-sided Alou, instead of just one segment four segments in the Julia package. It's just that kind of difference. Do.

B

You use a primary secondary routes or you just forward simulate with the finite line source in without whatever space you're in are.

A

You being the true they are in.

B

General or in carries codes.

A

I believe it's total field, I.

E

Think, oh so kind of check, because I need to choose the right way and I'm not quite sure. What's the best way to choose that part and then at my penance time.

A

Okay, so existing seeing the current code base already incremented like that, the SEM can be either airborne whatever so something I can take a look, maybe just change something so it works. For said a Marine says: yeah, maybe that's the fastest way to get since tower. Yeah.

E

That's true and then there is a code, as you remember, the line, currents in fact there's a code that you can use it. It's only working for tensor mesh but they're easy. So you can kind of use that or miranium.

H

E

H

A

Which I say it again? Yeah.

E

There's a line current source, so BC you can do the same thing that you do with Julian same time. So you don't read your million to redevelop that.

A

Okay yeah, so there is already some 3d code control source code, don't necessary for Murray, but just too generic yep, okay.

E

Yeah I think, but we haven't quite tested that that was my point.

C

Playing at you.

C

B

Do you have an example you could send to the ground.

E

E

Yeah, isn't you an example? Okay,.

A

Sounds good yeah, so that'll be really helpful. You have examples, maybe templates or like even just menu or something I can just follow so I'm sure so make sure I understand. If I do this type of coding is it's kind of like English impact fashion or like it's, um it's not violating any assumptions or like not violating them. The style I'll.

B

Send you there's another example: we have that uses. It uses the cylindrical mesh, but you could swap it out for a 3d mesh. Okay,.

E

B

Explain, step by step what's going on, so that would give you a feel for the terminology and like how we set up a second illusion and then that, coupled with the example that saw me sends you will hopefully get you pointed in to how to set up a 3d example. Oh that.

A

B

A

Already exactly what I wanted and.

B

Then you see the sympathy, examples right, I know, there's not.

E

B

But there are a few yeah.

E

Like you see has incredible people: oh, the hill is pretty similar, like problem will be changing, but I would've actually set up on the surveys that have problems it up. It's quick place in one good.

A

Good yeah, so maybe some tablet or example. Another thing I need to get used to the terminology here like because I realized, there's some yeah so like an inner product, see.

A

I wasn't sure what it means by inner product, beginning and I realized. Basically, something I've seen before is integration of the to do the basically outraging matrix that right.

B

Yeah physical property, matrix, okay, Halle G comes from okay.

A

B

A

Yeah yeah I read them yep good.

B

We outlined some of the terminology as well: yeah.

E

G

A

The last question is like the hat here. So, if I divide something, how do we determine is of the simple code base, or maybe is just like a new standalone package with a label like a simple certified or seam pack like a faced hiking shoes do.

B

That really depends on the scope of what you're doing keep on. So, if there's something like developing a new source that, if you want to do a primary secondary approach for marine C SVM, that would be developing a new source and that fits right into sente. If you want to try and build a p2p distributed parallel solver and that's probably like on top of sim peg okay, so these I think are just like, as, as things come up, we'll chat about it and we'll find the right home for it. Okay,.

A

D

F

A

This, if we proceed no code in Sturgeon area or may be likely to be recycled reused by lots of other people, maybe it's good to be part of a simple code base yeah, if it's very specific, for example, I solve a very particular problem, it's better to be a standalone package, so.

B

I guess like the and I think this is what you're saying, but it's basically like if it's reusable it's like you're writing script. Action solve the specific style of inversion, then that probably sits outside okay. We can figure that out this weekend. Are you on the same page organization on github? It's not I. Will.

A

Yeah, let's double check I'm, not quite sure, okay,.

B

So if you just sending your github handle all add you so that you have push access to sin peg.

B

So we don't let anybody push directly to master you'll, just branch off and build from there and then because we do like anytime that there's new code coming in will peer review it and make sure it's sort of testing standard style standards. All of that and then before it comes in. So that's a bit of the workflow there there's a write-up in the docs about a bit too dragon. Okay,.

A

B

A

Alright, that's poverty. That's here, yeah.

B

That's exciting.

B

Do you have any other questions, Deepa.

A

No really are you guys still gonna continue the meeting or yeah.

B

Every Tuesday I mean.

A

Today, I mean I are gonna wrap out. Today are oh.

B

These are the other thing.

B

Yes, it's there for you and you say anything else. People want to go can't.

C

We just serve sorry son right. What the the simplest problem at this point is that people would like to do it. Is it basically a grounded source in Walker yep, the way the fields from Athens yeah.

A

Yeah I think that's uh on looking at now do.

B

You want you having a simple sketch of maybe a 3d problem that we could help you set up. Oh.

A

You mean the sir er yeah.

B

And the model, just like you know, there's maybe the canonical sort of marine cfcn problem, but then make it 3d. Do you have an example of like the right scales and a simple survey like a line of data? Okay,.

A

Yeah, maybe my son do something: maybe tomorrow I getting to office me.

B

Great, especially, actually, if there's something that carrie has used to benchmark its codes and then we can make sure that we're all on the same page and getting what you expect to see. Okay,.

A

Even just some sketch there are that's okay, right, yeah, absolutely, okay, all right yeah, I'll figure out.

F

C

Because it actually, if we talk about the 1d case, then there's actually quite a few benchmarks that could be performed on that that Peters and yeah and put a book or something yeah.

A

So I'm closing my chéri or okay, all right. Okay, all right!

A

Thank you! Guys, bye.

A

Thank you, dami. The.

A