GeoSci.xyz SimPEG, 7 Aug 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: SimPEG meeting Aug 7

Description

Seogi Kang leads the meeting on spectral induced polarization

A

Okay, yeah so just want to share what.

B

A

Working on for the kind of motivation here, it's about spectral induced polarization. What that means is, basically you got a multiple frequency channels or time channels. So sometimes like it's a simple model: they explain that using coal coal, whatever other IP models. So here's an example.

A

Then here's an app energy before so this is the conductivity or I. Think resistivity is a little bit more intuitive in DC survey. So if you look at the curve, so that's a real part, so it started from large value and then decrease so I think that's the zero frequency when you're fully charged up and then you're getting discharge it as for high frequency, because if you got a note, I'm like if that's really fast frequency, then there's no time to build up the charge. So that's what's happening and I also.

C

Not only that you got a imaginary.

A

Part which is like it's like a phase way. It takes time like a to build up. The charge takes the time, meaning you're gonna have some sort of phase information and then.

C

That's actually.

A

In the imaginary part, and that's in frequency.

C

A

We're measuring it's in time, so if I transfer that into time domain, so here, I'm gonna.

C

Convolving with.

A

This step off kernel and then output it. So if you, if you don't convolve it like, if you just.

C

A

It that's an impulse response, but like that's, not what we're measuring measuring is pretty close to step off I, think I'm, putting.

C

That some sort.

A

A

Pulse looks like this thing and then here we're assuming we got in finished a long long time and then turn that off. So that's a step off and that's that response looks like the resistivity is similar to voltage. If you think about the opposite, all so that's very similar to what you're measuring right.

A

So you could measure from like a one millisecond to ten seconds, for instance, and then what you're doing is like you, do something like a stockings and processing, but at the end of the day, what you're using is like that time window and from 800 to 1400 millisecond you integrate this time window. Get one value, for instance, here is like 40 milliseconds: that's your data, which is not power charge really, and that's not so your cost is done now. You take this one point data.

A

Well, you got multiple source and receiver position, meaning you've got tons of data. Although we've got single time channel and then you recover chart ability ball. Well, you first! If you see that, that's what usually do but the interesting thing here: okay, that's what we're doing over ignoring all this other days, which can be pretty useful, but there's.

C

A chance that say: okay,.

A

Let's say we got, we are interested in the trail, have table point one or let's say we got pogchamp so that actually changes pretty bad difference in your responses.

A

So I was explaining.

C

Just that, like.

A

A kind of bit of motivation with the spectral, IQ and.

A

So that one town can change, see it's proponent, one three point four and five: six: seven! No!

A

So actually that it changes a law like it depends upon your polarization character, not look like I, rather than discarding all of those time channel that anyway, to imagining. Let's take that account invert that, rather than recover just a chargeability value, recover all three D parameters and then see if that issues, but that's the additional information. That's anyway we're measuring it. It's there.

C

A

Than just discarding it, let's take that account invert it recover see. If how much in additional information we can get I.

D

A

Don't have like a I think it's it's almost like a machine learning, I guess like a, but that's it what's called Nouman standard. Thank you, baby. A lot of IP survey so, for instance, I think what what that could be. Let's say they are in stirred interest in certain towel. Okay, they're interested in a certain mineralization have a certain towel or she can see. Then there could be a some sweet spot of the time window that they could see that right, but they're, not interest and other things.

A

For instance, then I think that's what it happened. I guess that's my guess. Otherwise, there's no point. It's like! Oh, that pie window could see something they want to see, but other.

C

A

Not quite interested yeah, but.

D

I think so here, like.

A

Okay, that that was the first order and that actually went well, there's other time John. Why don't we actually take that account.

E

A

So that's that's my motivation. Yes,.

B

I think part of that's right because they were, you know, looking for the mineralization sulfides. All of this, the fundamental research all came out of Nevada. That's where that's where the powerhouse was so they were the guys and then other people would come along and they do happen to do different channels and then realize, like okay, we can't calibrate you're doing something I'm doing something. Let's, let's make a scandal. So then they just decided on this.

B

As the Newmont standard, and at least that came, then everybody feel an opportunity to collect data and then compare it. You collect I'm an idiot, I, collect kind of mint. We could at least see if we're I think it got. This is real.

A

B

Motivation, yes,.

A

So that's that is that was the motivation on mine.

F

Because I asked I like a.

A

Whole bunch of other people, who's like at acquiring this year in IP data I, think that's their usual process and then they're processing is really kind of focus on to go. That like to go that a certain time window and get the one single IP data. So.

B

That's how they build up all of.

A

Their processing software and then interpretation so yeah I, think that makes sense them like they at the end of the air can discard it. But I think here.

E

A

Like I want to show some potential as I simply, it seems like a people are pretty pessimistic about. Okay, like getting ten time kennel, adding ten more time. Champ doesn't really add information said that was my impression when I was talking to like that seats, it Visser or other guys, who's acquiring the IP data, but I'm still not sure what like it is that right.

A

That's the paper and we're like you're using a linear solution, so there's a I think that the focus here is like I want to develop. Okay, that's one part for a mineral application. I want to develop special IP inversion code that can extract that right and then the product part of the.

C

Reason I think they didn't.

A

Have a right tool: I guess, to extract that information, so I think I think that's potentially a reason. So I want to develop an efficient code. They can actually take that large scale. 3D additional P data cover those multiple projects, so.

C

That's a mean mean.

A

I think goal yeah and using.

G

The linear kernel has.

A

A lot of advantages because it's actually much cheaper if you take that as a nonlinear problem or very pretty sense like how are those coins are doing, for instance, they're, take like they're computing, all the things in frequency convolving their waveform and then coming back to time to me, like that's fine in 1d or 2d, but like once you're moving into 3d that, like that, doesn't really work, I guess! Alright, you got a really big computer, no problem paralyze.

C

A

I think there's no reason. One would get just roughly it time to me and we can actually use that just a linear kernel, which is actually.

C

Would be developed.

A

Long time ago, okay and the question here, what is the pseudo chargeability here this? How Adri are we derived it because, like a here's.

C

A

Can use coal coal or so that if you choose coal coal, the problem is okay, it's defined in frequency domain. So anyway, you need to compute in frequency and somehow convert in time, but at the same problem like if you want to do that, each single box, all let's say you've- got a million cells and you need to do that transform million times they can paralyze.

B

A

I think it's still pretty expensive and computational. So, okay, can we use some exploit? Is there any other function that I can use it explicitly, then that I'm good to go I think the first thing that I heard was actually from Elvia and he.

E

Was pretty excited? Oh I found like a new.

A

Function and then you only e that he can solve but I think was 1870 is something like that. A German researcher Ashley found his function and use it for a number of other applications, so in polymer study I think that's the main function, though, that there I think they use other functions like cocoa or whatever occurs, or a dielectric relaxation I think it's not like it's not no one. Oh it's known for a while. It has been used for other application.

A

But, interestingly, in that, in our application for IP Cocoa's, the main one has been used for a while. Some people don't.

F

A

If I start talking about stretches potential, but I think we can still use it and by using explicit time function you can you can you can actually do that convert you don't need to do that conversion right, so you just need to somehow plot the same. To here then you're good to go. J is like it's. The VC sense. You've already got that, then you can actually compete yet.

A

Here's a little bit of like a math, so that was the step function. Remember that straight exponential and like it's a step off so use a step function. Then, if you want to write the step function, step off function to start from one subtract, your step function.

A

But the required output is so if this is use, one of you, that's it but jessica's like if people use different terminology. Oh.

C

It's a step function.

A

1000, it's kind of learn event like what's that, like this people start there with their care knowledge, there's no idea what they're talking. So they said it's pretty important to me to really clear what that is.

E

A

They find itself.

A

Yeah, but still.

A

It's not Python stand there. I need to put on I guess: oh is.

B

That right, oh.

A

Hi Connie is actually rather.

F

A

Using just that part of the world I think it. This is fall.

A

B

A

Trade off, so that's a step off function. Right, I get convolving it to step off function, that's the that's the visiray, and by taking it taking the derivative, you can get the impulse function.

A

So that's a recertified if you have a cold core model, for instance, if you directly transform that into time that what you get is a impulse function and then it's like that's what that's explicit form of.

C

A

Exponential in time domain, so that's how it looks like then for when they're like.

C

That's the supercharged.

A

Ability that's impulse. We need to now you to convolve it with a waveform like this.

A

So you need to convulse with that: that's a real waveform and actually in real waveform. That's not just a single period got multiple people. So let's say I think I asked John. John 1 is doing like a East X, for instance, but 40 study, John's.

G

F

Yeah is that right, yeah it's about 80 to actually do generally yeah.

A

So it's in the pantry and then there's a practical like stuff. You need to be careful with. Oh, we.

B

Cannot really used to step.

A

Off, for instance, right you just need to you need to take a count on what is your waveform then like, okay in practice, you're doing 80 steps? Do we need to do at 80 times of this computation or like okay, which is 2, 2 or 1, or you could just assume the step off, for instance, so that's kind of practical questions.

A

C

So that's sort of the background.

A

And then, once you solve those problems, then you're set in terms of for simulation. So, for instance, you can that let's say January this model, that's a portrait model and like I've there there are multiple units, I, think that probably makes sense. I guess I asked Tebow. What is that reasonable portrait model? I think he said? Yes, probably, okay,.

A

George, of course, that I know.

A

So the main structure here is a halo like this orange bag and the mineralization and height there's some porphyry, this purple one and stop okay, and there are three chargeable units which is halo, mineralization and the clay so by just using the typical like dcpip inversion.

A

You can recover is all three of them, but you cannot really distinguish what's why. So, if you can distinguish all three of them, somehow they'll be great and so I actually designed it that I could do it. What is it that example? So what the main difference between the halo and the mineralization is time, constant, so time constants is depend upon the size of the green. So let's say you've got a very fine grain or a fine generalization, because you're gonna get that really small time constant.

A

But if you get like coarse grain, size were like coarse mineralization, which probably what people want to find I, guess and then you're gonna get large time constants and the clay and halo has a different frequency component. So C also determines how, in the case it's not it's more like a distribution and frequency domains. If you go back here so if I change C from point 2, it's very like a very smooth right, like in terms of distribution in frequency domain.

A

But if I change that to it's like that, that's very strong right and then time it's actually became a little bit. I'm, not sure it's, not the Keeley, but I think it actually still makes it faster. Came.

D

A

Right so anyway, so she can make some difference in your data as well, and usually clay has that large C compared to the other, so I assigned that large C for play and then smaller C for halo and the mineralisation. So that's the set up. So in theory, if.

F

I can recover all those.

A

Three parameters: I can distinguish playing Halo and the normalization I think that that was that.

D

A

So could because step in, if you think about the C data, one has large right. You got resistive kind of like a resistive computing. Then you're gonna get large signal, for instance, so clay is very conductive, meaning your actual voltage will be pretty small, so I think yeah yeah.

C

In terms of noise level, play has less.

A

Tense to say, if you got like a chargeable clay, then IP response in terms of voltage, probably pretty small.

D

A

That I'm getting what would step this is.

D

Your volume of improvements, yes, most of your signal.

D

A

That really depends upon.

E

A

So there are like a song challenge and also the measure time band. Is it's pretty it's small because it's designed to see that into a time, Channel right but I, think there are some moved ones? Okay! Well, there's no problem! You just need to increase the higher sampling rate, so there are some people doing it, but I think that's not quite standard in mining industry, so I figure we probably to show the values. Then I think that some people might take it up.

A

Okay, we've problem to increase I, like a sampling rate, I, think that's sort of the goal, but.

A

That's the setup and then that's the connectivity. Okay, so we got a resistive.

C

A

Halo not comply and also conductive me realization zone, not very conductive, but like monthly, it's a charge ability, so you can see any difference so I, just designed, okay I got exactly same charge. Ability all three minutes and Pao is different. Normalization has much greater towel time here. I said: I'm, assuming my mineralization head zone has a much greater course like it. Your life, minimize grain size and see, is different between play and lay low and also mineralization.

A

So the black is a large value, and if you look at the data that looks like that, so.

G

E

A

So here remember: I got the five lines of the data. It's not that big! It's your baby, but it's fine! Well a nice lines of data. Each line is separate. So it's not like each line. It's not connected, so I'm just doing a dipole-dipole in line which is kept typical, I guess.

A

E

A DC data and IP data at.

A

All the times, that's one millisecond, that's pretty early I. Think.

A

C

Date of it, I've got was like point.

A

Pulling on fought like a John, what was your always time channel.

F

A

I think is both 30 milliseconds, so I think one milliseconds pretty Olli in terms of like a mining standard, but anyway this isn't that example work and then IP data or second, that works like that, really see ya. So that's quite different right. It's I think that's a good sign. Well, your data is different, so there's some things that we can actually expect.

A

That is true. Yes, so here I'm in like I'm designing in that way, we could actually extract something but the if you so that kind of like if you see what our disguise doing a thing in there doing less stats. So it's a like. It's always trade-offs right like if you do it searching the et stacks, then it takes a while right, but if you just if you're just doing the for stacks or two stacks, but that measuring longer time channels, then I think of doing the.

A

For second thing: it's not a big deal, for instance, so I think it depends like so that's a question. Okay depends upon what you really want to do or what you want to see. You can actually change your sort of like a measurement system and then well, that's always a trade off. So.

D

A

E

Time right like how much how much time with Peaks.

D

A

Record to the end.

A

Yes, I think here the goal is getting a really high quality, less noisy like a single time integrated yet, but if you're, actually taking a cup like a time like a multiple time channels and fitting that I think it a little bit of noise in that a certain time. So it's kind of fine I think and then you can do that other processing.

C

A

C

Does that make sense tom so.

A

C

Just a trade-off between like either you can stack a whole bunch. So that's what you're, using your time for or you stack less but measure for longer times so.

F

It's like a time versus.

C

Signal-To-Noise trade-off yeah.

F

So yeah we stack that long so that we can ya stack out the noise as best we can, the more stacks you have, the better. You can really beat down that signal to noise.

B

How long the base period you use, because Sagi is looking here at four seconds, yeah.

F

We generally do a two second off time. That's pretty typical place.

A

C

A

A second there using a 16 second.

A

D

Also, the other trick that you.

A

Could do rather than using a half to a cipher, can do I.

A

Have to do cycle is a step up like a turning the car off Yeah right. So you don't really need to do it like if you got a wait to extract the step off from the full duty cycle, so you can actually use a whole theory cycle. Then you can still use just a second and get the for second data.

A

Today there are a number of details here and so now I'm moving to an inversion. So that's a four problem, so I got the data. Data looks like okay, there's some interesting things going on to extract and let's move on to inversion, which is.

C

Just a quick questionnaire using the octree mesh minutes with the tensor transformation.

A

We just got I think that we got the all the pieces, but it's not quite there yet so Wow well push that later.

A

And also I was thinking about the bet, but I think it's not the point of this yeah.

C

I know totally it's just like that's just an implementation question right.

A

I think I hear sobbing okay, so here's a IP and version methodology- and this is your predictive data- that's it still a linear kernel. But then now, in terms of like a death house, see it's not quite a linear kernel, it's a nonlinear, so different, linear and now your P data is a function of conductivity charge, really see the here. Sc e stands for skeptics potential, so I'm, just like I just want to distinguish this from coal coal and to the chargeability.

A

This guy is a function of chargeability, tau and C, and then comma theories known right and we do a DC inversion. We covered that assume. That is pretty close to the true continuity and we can just generate the sensitivity function here. Okay, that's the set up, and now the data looks like that. So we got a multiple time: Channel, meaning you're got a pseudo Charlie really at each time job.

A

Let's say you got a twenty time challenge then you're gonna get to the junction we'll get to it, but that's not what we want to recover right, but if you think about go back and but what the conventional workflow is recovering, it's just one of the Sutra bill. Okay, at certain time, chap and now my model here is that M EDA and tell MC, so I got three parameters that I want to recover and well I get the usual question. Okay, what is my model right?

A

Well, that's kind of the first question that we need to ask, but sometimes that's kind of underestimated. Kids like theirs, are usually a typical choice, but here it's a little bit a new problem. Okay, what's how do you? How do you choose model? It's that a part of it, research, question and I'm, not gonna.

B

Go there but I think here's.

A

An option: okay: are you, can choose either log or than your Wow? Okay? So what's your choice? How do you choose it? So there was a part of my research question and then calorie map my model to actual physical poppy yeah. So here's a detail. It's okay! You guys probably know that. But we've got this yr mass right, which.

C

A

Pretty cool, then, why are you saying it? You can be pretty flexible, which is actually really cool. So here it's very easy to do now. Let's say I, okay, I know pretty well about the top in this region, I'm expecting to 0.5 and so I, don't want to invert that they can easily fix the towel and just invert for chargeability, antsy or a whole bunch of other.

E

A

Is possible by using water, which is actually cool, I didn't mention it here, but it's a bit of details but I think in in practice. That's that kind of cool thanks to do yeah. Sometimes people really care but I think.

C

That's an important.

A

Point do if you a detailed tutorial on that. No.

C

That would be a good thing to put together because it is like a conceptually, it's actually not too too involved, but it is definitely something we get your head wrapped around yeah.

G

If you screw up the sort of the syntax of the process in one step, it's a really great tool that you can't use I, guess, oh so,.

D

You're turning off for the Reg you just green wire, whether I could and for the business function, allows you I'll be using a wire. This.

A

Is just part regularization, just yeah so well.

D

I think is actually explaining a.

A

Bit right, okay, this is basically just an injection, no tricks yeah. That's.

C

Effectively all the wire map does, it basically says like have a giant model and like these six parameters are towel. These are data, see and then, like it maps them to where they need to go. So you.

A

Didn't just sleep and I think I.

C

A

Explain it, but actually that's so portal, for instance, if Tao was the second variable that looks like that. Wait.

G

Doesn't Nestle has to be.

A

Like how can be a very very said if you just said the first variable, then all of this changes anyway, so the choice is either log or exponential a lot more than you, so I I ended up choosing all the log variable, so I could show okay yeah, it's a usual stuff, didn't misread certainty and here's your combo get the function, so I think I just kind of want to mention what what what? What were the pieces that I had to use and how useful it was.

A

So this is the comb objective function right like got the regularization function, smoothness and smallness, but each Pronger it has as bad right. So you need to have each of them each like each parameter with some organization, so I, just kind of looping over so I is a de tau C and we're actually summing that over to a single like a regularization yeah, yes, I think if there are a lot of details and that's if they're warrant that was functionality, it would have been really hard to match.

A

Name from that in practice, because that's actually could be a lot of work and then at the end, we're actually putting some constraint for each three variables and that were minimizing. This functions projected gospel, but.

D

Mister functions are all auditing.

D

A

So that's a good point, so I think I probably need some sort of like a constant here. I guess.

A

D

I think it's just the regularization.

A

D

A

In that case, I passed the college directly to problem. How is a known value, so I can pass tab.

C

And then so you know I think perfectly for me, but I think Dom's a question is just then you wouldn't regular eyes on account. So you would drop that component in the Econo regularization yeah I.

D

Don't know whether it's just right now didn't we. If your model is too long, your assumption would have a gradient that you yeah.

A

So I think I'm doing is smart and in the code. So if it's, how it's not so, for instance, like the mapping for tau, is not set, then I'm not like actually computing. So that's a beauty of the wire maps. I think then you just need to pep, like that. That case actually outputs to zero so and then I make sure I'm not competing the sensitivity or tell so just like a Zillow, so yeah so yeah I think so yeah then.

E

A

A details I guess like if you want to do that type stuff, like a func like a flexible way to okay I, just want to set the top and I'm just inverted for two variables. So, like that case, I could make things can be kind of complicated without.

C

A

A framework or structure which is actually important and I think there's a value because, like let's say now, you want to do even more complicated things like okay, I kind,.

C

A

From I, just don't know that like I know the Tao, but like a pretty shirt, I think I still not exactly sure so. I invert that first invert other data first without the towel, then okay I, pretty close, now I'm going to turn on Tao and invert, all of them together, for instance, then I think there's like a that and that's possible with this structure, which could be interesting thing. So there.

C

A

Be like a other research: okay, what are the good strategies to invert I? Think so that probably interesting.

C

And just like, on an implementation side, there's two ways to achieve what song he's talking about with dropping their identity: ginger so like you could, when you're setting it up, it's not included or if he wanted to do like in his second examples, would be ability to turn it on and off. If you set the multiplier on that then to zero, then it just won't about you.

C

A

So yes, I think that that in Burton part I didn't want to have a touch. So sorry, I did that that's fine, but I have to do it, just like your weave together and right. So when I did actually let the sensitivity function and then weave them together, like that, we can invert that.

C

Actually, that part was pretty fast.

A

Cuz, that's actually it was a cool part, so that so in version methodology part is done. I'm going to kill.

C

You guys some of this.

A

Example: well then, the focus.

C

A

Paper is 3d, but I think anyway developed 2d. It's not going to start from 2d, but it's kind of nice and.

C

A

Going to build up okay, what about is important, what are my models and what are the setup that I need to use cuz ii, the the bottom? I got this problem. We got three variables which could be really not unique. So how do you tackle that? That's could be a part of issue so rather.

C

Than starting from 3d, which could.

A

Be more challenging I'm gonna start from simple, easier problem like T and then build up the set up and then just apply that in three two. That's.

C

A

A

The other thing is so what the opportunity to many conventional workflow but they're so you've all end up did other work. So they worked on spectral by P conversion problem long time, I'm not sure.

A

That's what he dad.

B

A

Same thing, go back to just.

A

So what they did actually like you can just invert each time channel separately right so rather than.

E

Inverting, all of them together.

A

Just got the DC conversion generate the sensitivity, function and invert each to dischargeability separately, and once you got a few key volume of like having multiple plan channel, you can kind of set up by a small inverse problem for each voxel, then recover a chargeability time constancy. That's what they have done right so I think that's exactly what I've done in the same workflow, and so it kind of need to compare with care water, developes, converting all of them together compared to just like inverse M. Please do a separate inversion and extracting some intrinsic croppers.

A

So that's something that I also want to show you the value.

A

So I'm kind of starting from the first problem, and it's actually a this same problem. If you think about that in terms of inverse problem, cuz, okay,.

A

So let's say that's a time eyes attention.

A

So when you're recovering its described right, so you're gonna get ya the first time channel in 340 and.

A

They're doing so I'm going to pick that a single pixel, let's say I've got G. It's still here.

A

Time, okay, so yeah. This duration, then you're going to get minecraft okay, but you got this this guy's at each voxel in 3d or 2d. Okay, so you can set up an inverse problem, basically using the same regularization function and recover all three Crocker's. So now, let's say you got a 2 D 3 and then your data served at the same as a pen, tilde or estimated.

A

So you do a separate internship. You recover that now that's her observed data and then you won't like predicted these fitness data and recover all three powers. So there's a lot of similarity like that. What's changing is that we're changing that serve data and therefore kernel, but basically it's the same problem so here I think that we can learn a lot just solving it and main problem was that K? What is my.

C

A

A

Here's a comparison, okay and if you actually know well about the IP I think intuitive choice is ADA, Logtown and C. So tau is how is varying like kind of logarithmically but ADA. It's like linear.

C

A

They're promising argue, but they think, if you think about how are not working at that site, probably when you and sees linear as well so C changes from 0 to 1 I, think that could be an intuitive choice or you could just take the log all of them together and then, where you could choose other things. What I did is? Okay, so I can compute the gradient and see if how the gradient, like a distribution of the gradient slope, side.

A

Then there was the inversion reserved at the only time, channel late on channel there, the king in terms of altitude, so I did for each time channels now I'm extracting.

A

So here's an interesting plot and if I chose this one actually inversion is like falling into local minima. So it doesn't really converge to my target Nesbitt. That was curious. Okay. Why is that? What's the problem and that by plotting with the gradient, it's actually getting clear. So if you look at the histogram much ingredient and chargeability has much larger value compared to the others right. So you need some sort of balancing here and then you could actually like good, but some constant or you can choose your model. So that's our choice, the mine.

A

What my choice is: okay, hollow well like you're, taking the lock or all the variables then actually like the histogram is getting more compact and actually that inversion converges pretty well. They.

C

A

It quite robust and as actually reading the others paper and that's what they're doing so Pacific does the way but every, but they didn't mention anything but I, think that's probably why they're choosing it or not I'm not sure, did.

B

You try that just keep the see either linear domain dimension stream login, how a ah no.

A

I haven't tried. Sorry, I could definitely try no problem.

A

Yeah I think it's a certain case, so I think that that's for this example. So I'm not 100% sure, if that's really a totally extendable for a whole bunch of other cases, so I.

C

Think there's a okay.

A

There's a cases maybe that actually choosing linear might be better so by that I'm not sure I figure.

C

Log stems like.

A

A robust option, I guess so as a default I think I would choose lock.

A

And I think that's probably like it's too it's.

B

Solo right so and then, when you chose lock, did you go back and kind of think about what your regularization parameter your calculator, ization functionals? So it's now I mean if you're working with you know the log function. It's really different right, yeah yeah! So did you go back and alter that on the basis of what you thought, you're going to try to get out physically, but you're still working with the log of something not quite.

A

Not quite like.

A

All it's kind of like.

B

C

A

And error, so this partner I got some intuition, but when I'm actually working in the inversion to make it robust, like setting up the parameters that anything goes like purely trial and error, I.

C

Wasn't a hundred percent.

A

Sure, like okay, what exactly I had to change but I think I didn't have to change that much side like I was basically used mostly default option.

C

There was some.

A

No I think actually I used like an almost option for for regularization, same kind of like an impact on alpha s. Alpha X was that or something a 4cv problem that was too deep like it for I got.

C

A field example.

A

But like let's say your lines are pretty far then I just has to use alpha Y large, which is kind of typical, so I haven't. They didn't use.

C

A

Fancier parameters here.

C

And so when you did the switch, you then regularize. Did you regularize on the log of those Crowder's? Who did you? Okay,.

A

Regular eyes on the log of the progress so and recognizing on the model: okay,.

C

Cuz, that's also like something that you could try if you were interested, is actually like. Do the mapping and regular eyes on linear parameters, yeah.

E

C

Cuz I mean that actually could be interesting of like what are you doing if you're changing just your day, miss Phinney versus the regular yeah.

A

I think there's definitely yeah so.

C

Yes, I think it here's a that's a straight.

A

Up: okay, what's the question and I think that yeah they did that through depends upon, like your focus, I.

C

A

Like if the question is okay, what is my model? If that's a specific question, we're asking you definitely but.

G

I think that was also part.

A

I think it's not main impact different like it go through. Yes, but I think that's a good good and then here's the cupboard, chargeability Chow and see here what you could potentially see it like. Okay, maybe see not quite I, think it can get tell Valerie's but didn't quite the distinguish now ization. Yes, this approach, it's in terms of resolution, I, think conventional approach. You may lose some right cuz like a year.

A

You can splitting up apart that as a two stems like and then that really depends upon how how kind of like how quality sues, charted, really that you've got from the previous inversion it each time channels right so.

E

F

E

A

In practice yeah, so you need to do a really good job kind of doing Oh recovering see the chargeability at like each of the time channels and that the results is really depends upon that part so that set up it looks like a disadvantage or could be enough that dish, but I think there's a bit of so.

A

C

A

In yeah in theory, it's.

C

A

Robust approach right like a but in terms of non uniqueness, it's actually invert each time channel so.

C

It's less not less.

A

On unique compared to the other problems, but there's also a little bit disadvantage that you need to make sure the quality of those also touch our table 30 time chance put that channels. That could be this advantage. Now, that's d8 when.

C

You say it's less non-unique, I, I, guess the way I think I would interpret it as opposite: you've actually put by I'm bringing them all at once. You've put more constraints like you actually have less free parameters that in between stuff right, because if you invert every single time channel, you actually have more free parameters. Then what you end up interpreting at the end, because you have an eight, it's like you, have a pseudo chargeability at every single time, stuff, okay, that and way more founders than having three at each voxel. Well,.

A

Yeah but the I think an itch conversion like I'm, just thinking about each enclosure, so at a single time, so I think that it like a non uniqueness next to basically the same as the a convert for.

C

The city hood yeah.

A

C

A

Sort of, and then it runs together, even though you've got twenty times more, you got three propers I think that that, like increase of three anything, it's.

G

Not linear increase.

A

I think it that not like it at them actually increase exponentially. I guess like if you just one progress.

C

What are the variables.

A

And to I think that I think that's not quite like a linear increase it like not three times Dominique's I, guess yeah.

C

It did when you're in greeting, though for three you're using all the time towns right, you're, not inverting, for three at a single Timeship right, so yeah that gets rid of through all time. Yeah.

A

A

Yes, in that mind said we were thinking like it, we're just thinking about a single time channel like.

C

In terms of done, units.

A

I think if we were like let's say: I got a single all right, yeah and they're.

C

A

And that that's a 3d variable at a single time, yeah, it's a fair point. I know.

A

Which one is more I'm not unique.

A

A

Kiss you, if you're getting into that level, I think it's kind of tough right like now, then.

A

B

He's pretty big I, don't think it.

B

Definite to working with yokai ciao ciao ciao is that now you put all the issues in the regularization now you're getting you know, you're getting an A Tilda at each okay. So that's that's a non-unique thing right and Nan. Now you do lots of time Council in principle, lots of time Chad was actually going to give you an over determined system right. So over determined is nice from the point of inversion.

B

So sure we get. You know that part of it it mathematically is still non-uniqueness at least an over determined system for three parameters. So their ingredient is arguably not so much as if you're in the function space Bob, then you've got the log data and there because of all your privatizations whatever so I. So that's you know, that's your kind of boot there that's right. If.

C

You just go the.

B

Other way, then, now you're, you still got this huge non-uniqueness, that's associated I, so I I have no idea of like okay, whether.

A

A

So here's the if I just do a spectral IP incursion, so that's a charge, ability and that's town and see there. You can actually see the large tau at the mineralization zone. Together, you can see everything really right. You can get high resolution and you can distinguish probably like play having higher see compared to the halo, so I think that that's sort of what we're saying is I value.

A

So here, if you're, if you're getting the right, pluses and there's not chance to get greater resolution, is very inclusion and then now you got the set up and do a three conversion. So this is a 3d inversion, reserved of DC chargeability time constants in C, so you can get so I think that things worked in 2d. So let's move on and apply this use the same property and use it apply the 3d version. Thank you well and yeah. So here's the kind of gesture fall.

A

So if you plot that recover see and the Tau and crosswalk did you.

E

Definitely see the clustering.

A

Right and we could do something wrong classification, one look why this is like an easy case. We know all that all like primers but let's say I hate their large, see with that small time, constant. That's a clay and small C and the intermediate time constant, halo and if I pick the large toe. That's a minimization, so well, they're, probably whole bunch of different way that you could constraint. But here we're just picking a simple stress moment: those are the parameters of.

C

A

This is a field example and Saudi Arabia. It's the same for free deposit, so this is the DC inversion and that's a charge ability that's time constant and it's do we get the same practice putting in the crosswalk and by looking at here. So these are the time constant frequency dependency and I colored by chart ability. So you can definitely see the Tran.

C

A

Here and there it's a high charge ability, high time, constant, small small charge, ability small time constants. For instance, that's all like, let's say, I fed to a binary and good pencils. Five like this and your model, Iraq model now and three units background, yellow, could be a mineralized zone. Mineralized zone or coarse graining zone in your mineralization and orange could be some sort of mineralization. Don't have a smaller grain size, for instance, it could be okay, actually, there's no.

C

Drilling here about.

A

Whether that's really reasonable doubt, but it's kind of showing the capability of deep code, and actually there are.

C

Some other details: okay, Matt.

A

Mentioned there's that relative waveforms and in there sometimes.

C

It's very important.

A

To take that account, if you really care about the values like a quantitative value, so I actually implemented like it's a it's a linear sum. That's a linear, it's kind of you could think about as a superposition of all the step response, because they got is it as a possible minus plus minus, and then you kind of time delays. So if you kind of track all of those things- and you can come up with some sort of function- a step function looks like this.

A

Then you need to put that in to your suitors, rep and you're. Competing is really tragic, ability also competing the sensitivity. Then that's yourself, that's a bit of details but important.

A

Yes, I think that's it and I. So it's a research research, but the what.

G

I thought kind of.

A

Important here, okay, so we could I'm.

C

Not sure let's say this is the hypotenuse, so.

A

We could actually do it now with the country one, but the what I was thinking is. Let's say there are some pieces right, very important pieces like wires and comb objective function, but there's a there's, a challenge by itself. Okay, like us, we definitely know what's the bell right because we're using it and then we know how helpful it is- are all researches, but maybe outside it's very hard to know that contribution so I'm still not sure, like definitely a power selling that we need to do a high quality research.

A

You should show how much you can do with that. I think that's sweetest way to really sell. Pretty more is important, I think, there's no question but I think.

E

F

Still, we need.

B

A

That part I'm still not sure how how I, how we do it so I think.

C

That's probably we.

A

Need to think about hey, there's, Sam and Ron works, and there are some base functions base boards, which is really important, but it's kind of hard to sell them in the Karman like Society. Yes, so yeah I, don't have a really good answer about that. They think I kind of one to know is such a question that has so.

C

I really appreciate it.

A

Like okay, finish the research and then I, look back but I have time and what kind of pieces that I need it. So there are some important pieces and then somehow I want to contribute that but I think I, don't citing sin that paper or something it could be. But I think it seems like that and I guess so. There's there.

B

Probably something that we need.

A

To think about, but.

B

Within the development of the emerging paper, I mean there, could you know there could be sections on numerical implementation and then just have a section just talking about how things are implemented in how things are connected. I mean that's, that's quite possible, especially especially for something like GG attic publication. Yes,.

A

I actually put that in and I cited, so we could go, look up a bit more details.

C

Because there's a question: if you want to go into that, I think and like just scope of the paper and not that's really up to you um down the road I think if we have a few papers out that are using some of these things, then actually writing like a synthesis or sort of review type paper that talks about actually the underlying structure. That enabled that, like that's, potentially a trajectory forward, yeah because it's hard to write paper just about the combo function. But if there's.

F

C

This that we can point to you and say like this is what it is enabled. Then that becomes a lot. You know more possible I think to get something out there about those things, but.

B

Still with the with the new point of trying to elevate the visibility of what has gone on past I mean one can insert various things if small sections here they're ten minutes.

A

B

There that you can really consolidate that and then it's at least get.

A

B

In a concrete form and then that global paper- okay, yes I, agree trying to elevate some of what you consider consolidation but are actually two important parts trying to give that yeah I mean I'm. The same is true with what you were doing so yeah good thing to keep in mind and then just.

C

Absolutely well and at that point to like I, think it's being specific in like acknowledgments as well like saying you know, these pieces enabled this work like that's, also a lot of bad place to, because then people are sort of less concerned. If, like they actually understand all of it, because if it's in the paper, then you want to go into enough detail that people understand it so, like that combination, I, think it's hmmm.

A

Let me stay there.

C

Cuz you mean that.

A

Let's say a liar naps, for instance: that's like it stating ok, we use the function called wired man, sir, so is that how you I.

C

Guess the I: uh it's not a bad idea! I.

D

Use this website yeah yeah yeah.

C

No I think that's actually totally like. So in the discussion about the implementation like having a statement that says in the implementation, we use the simple barter maps to keep track of world yeah.

A

Like that is a quarterly unfair.

C

Statement right.

A

So we could have a small section that.

A

C

It doesn't even necessarily have to be a small section, like even in your current implementation, when you talk about the P matrix, for example,.

A

Yeah yeah fine I'm work to follow the Kinsey's set your own course impact research, I'm gonna put the LP as well. Yeah, no girls send results and online I think a 2d could be deployed them binder. So people can actually fun that this big. What.

C

I've done with all of my simulations or integrations that are too big to run, is I just do like a download right.

F

Right at the beginning, so.

C

I just say like a rerun, it's false, and so if by default is just downloaded directly.

A

There, yes, so I'll do the same thing.

A

Yeah I think that's actually nice, then Kenneth.

C

A

A good container for us yeah.

C

That's really cool, so I can.

E

Eyesight yeah, that's awesome.

B

John, what do you have for an example over a for free.

F

I, don't think I have one I was looking for one the other week, therefore soggy, but yeah I didn't quite find anything right now. There are nothing that I have access to, though.

A

So we actually started John and I started working on a small dataset that they I think it's a it's a feels more data center right, yeah.

F

I was the field school for the University of Saskatchewan there over a buried target right.

A

So what I thought like in terms of target? Why it's not very exciting but B? It's a metal, metal plate or something like that- is that right, yeah, no yeah.

F

A

Just a steel plate there again but I think in terms of like processing than their standing point of view. It's actually exciting, because that's what that John is interested like it's similar to the that inversion, but that you can apply that those be dead now. So we can consider that a suit of chargeability as a parent charge, ability and now I want to extract some I, keep honkers, for instance, so.

E

That's what John is interested so I.

A

Think I got all the functionality so I could you could set John actually prepared to some small dataset and man that, like all procedures of how they're stacking it and then converting that into a top window, dataset, so here I think it's and there's some interesting things. We could try. Okay, we could take that just that amount like a before the windows data.

A

You can just take that long like not wrong with it, but the kind of stats data then just fit that recover on compromise, or you could just T that we do it data and fit that and recover it. Koco prom, for instance- and there are a bunch of different things that you can think about.

A

Okay now, if you are turning it into a processing problem because well when when theta is 2 these times simple, but then I think it considered a to then that distribution of apparent bility or taozi supposed to be pre smooth if you're making a public if you're a carefully picking the similar sized example and then putting it into a studio session. They're not gonna, get really like creepy dataset or look like.

E

At that part could be just noises so.

A

You can actually set up a kind of like fitting problem and that actually you could do something like awesome, noise, reductions, stuff but I. Think, okay, but I think getting the fill dataset and then showing.

A

And then actually could take the data and invert with this workflow yeah. So yeah.

F

We actually might be just finishing up a pretty interesting, pour freeze survey here so depending on those results, maybe we can look at that. One.

C

Well, thanks Auggie, a couple of it quick updates were little returns will be best. The.

F

C

Posted an issue I think you buddy taking care of this on a different branch: seven, but the vrm tests are failing on master right now and I think you took care of it on the poor request, that's with all of the Dom stuff, but I was wondering if you could just isolate those specific changes and just do a quick pull request and we can get that working again because it's really hard to bring in any form of changes when masters failing.

C

So we can chat about that more afterwards, but it is it's functioning on it's functioning on the branch that you're, like the pretty much acceleration stuff, is I.

G

Have to stop the messed up. The vrm stuff are exactly the same place.

C

Yeah but it's not on a master right, yeah and so there's, there's vrm stuff, that's failing on mastery! So if you can just isolate the fix that you've already done and make a small pool request, not a giant then.

B

We can get it in.

C

Like today, because otherwise it's sort of stall, any other pull requests coming in, because it's it's feeling.

G

Yeah I think it has something to do with octree Misha's yeah, so yeah fix I did with Audrey wishes. For that.

G

But then we'll have conflicts.

C

It will like it should be like a to mine conflict, because I think this fix is like super simple I, think it's just how you're refining so.

F

Hopefully, it's like.

C

Three lines backs and that's a very easy thing to to conflict resolvers so and then other than that Teter pushed some cool applauding code, so that's that'll be fun to look at and feel free to provide input on his PO requests there. What sort of things base so that you can scroll through 3d models, so you can plot different slices of your own 3d model and rather than like hooking up a widget to it. I think he's just I linked it to the house. So you can you scroll and slice through.

C

So that's pretty slick, it's discretized! It's you just started a pull request and then there's a couple other small things: I'm just gonna.

C

One of the things so I started a small pull request and I'll circulate the link, just with some of the plotting stuff, it's nice to have a little more control on, like polish, for example, like to grid the location, Chloe figure like one go in and change the line, width line, color all of these kinds of things, and we just didn't give you hooks and so I think like this is a silly little patch, but I think that we should give some thought to how we actually want to expose that to users.

C

I think we could do something like having sort of line style keywords and because right now, I've just thrown them all in like Kame works, which is not great because then, if you don't know what the cave art is and I haven't documented it that you have no idea and so down the road. Actually, what I think would be kind of cool is to turn the plot. The current, like plotting actually into an object and then so that actually all of the the styling are just properties that you could swap out.

C

So we set reasonable defaults like what we've done, but then actually you could just swap.

D

C

Yeah I agree, and it's just a question of how to close that yeah.

F

C

Right now like, if you look at the plot grid, for example, it's like slightly non-trivial because we've got the idea. I don't want to pass all the same, plotting arguments to how we plot edges faces nodes and the great lines right, because we want different markers or each of those things.

C

And so, if I just pass like blanket, color everything red, then that doesn't do what you're expecting.

C

Yeah so, like I, think going sort of the object, rope. We're then just like edge color edge marker each of those is a property, and then you can just tweak them and like edge size.

C

Might be the way to go, and then we just serialize that and pass it all to them that well Lib. So if that's someone's fault and feel free to add comments on that border list, so anything else from anyone else.

C

That's it good, perfect, okay! Well, we will reconvene next week. So is.

E

C

Paper by the way with Craig, that's awesome, yeah.

D

Definite teamwork, okay, see.

F

Is there that's yeah thanks.