ОБ ЭТОМ ЭПИЗОДЕ
In the inaugural episode of Beneath the Subsurface, we delve into the exciting realm of AI and Machine Learning as a blossoming new part of the energy industry. Arvind Sharma and Robert Gibson discuss and debate the impacts of disruptive technology, the importance of robust data libraries when building AI solutions, and the future of our industry with AI and ML solutions. With your host for the episode, Erica Conedera, we explore the factors that pushed our slow moving industry to this tipping point in technology and where it could be leading us.
TABLE OF CONTENTS:
0:00 - Intro
1:03 - Factors that brought AI to O&G
5:32 - Job creation with AI
12:05 - Career paths and team compositions in the industry
15:30 - Industry pain point solutions with AI and ML
21:32 - Clouds, open source and democratization
24:24 - Kaggle and crowdsourcing Salt Net
30:51 - Kaggle challenges with Well Data
33:58 - Catching up with silicon valley
36:49 - Approaching solutions with AI
44:18 - Disciplining data and metadata to get to the "good stuff"
EPISODE TRANSCRIPT
Erica Conedera:
00:00
Hello and welcome to Beneath the Subsurface a podcast that investigates the intersection of geoscience and technology. And in our first episode, we'll be diving into the dynamic field of AI and machine learning as it relates to the oil and gas industry. We'll be discussing the impact of disruptive technology, the importance of robust data libraries when building AI solutions, and exciting possibilities for the future oil and gas. From the TGS software development team. My name is Erica Conedera. And with me today are Arvind Sharma, our VP of data and analytics, and Rob Gibson, our director of strategy, sales, data and analytics. Thank you gentlemen for being with us today for our first episode.
Rob Gibson:
00:48
Glad to be here.
Arvind Sharma:
00:49
Thank you Erica.
Erica:
00:51
So let's start our discussion today by talking about the factors that brought the industry to AI and machine learning. Why now? Why not sooner? Why not later?
Rob:
01:03
Well I'll start. Um, so thank you for the introduction, my name's Rob Gibson. I've been with TGS for almost 20 years now. And in that time, the thing that I have kind of seen over the 20 years in this company, , and probably another eight or nine in the industry, is that we've always been a little slow to adopt technology. And I come from the IT side of the world, software engineering, database design - so from my perspective, it's always been a little bit slow to bring in new technology.
Rob:
01:34
And the things where I've seen the biggest change has been fundamental shifts in the industry, whether it's a crash in oil price, or, or some other kind of big disruptor in the industry as a whole, like the economy, not just our industry but the entire economy. But in middle of 2014 with the current downturn, that's really where I finally started to see the big shift toward AI, toward machine learning, towards IOT in particular.
Rob:
02:00
But it seems like it took a big, big change in the industry where we lost hundreds of thousands of people across the industry and we really still needed a lot of work to get done. So technology has been able to kind of fill in the void. So, even as the downturn happened, we kind of started to level off at the bottom of the downturn and that's when companies started to see that we really needed to inject some more technology to get those decisions made. So generally speaking, I would say that this industry has been a little slow to move to adopt technology even though the industry has got a lot of money to invest in those kinds of things.
Arvind:
02:34
Um, so thank you Erica for that question. And, I'm going to slightly disagree, more broadly, I agree with rob that um, oil and gas industry is historically a little slow in adopting technology, but, the reason I think is a slightly different, I think a oil and gas work in very difficult area where we need to have very robust proven up technologies to work. And in general, we wait a little bit for the technology to prove itself before adopting into, um, more difficult areas. So if we look at a little bit historical view, um, we have been on the leading edge of technology for a very long time. Um, some of the early semiconductors were built by your physical, um, companies. Um, then, as we moved to, PC revolution, we started actually PC, um, we started to actually pick up PCs into office very quickly, not as good as the silicon graphics people, but, soon afterwards, and then when the technology evolution started happening more in the silicon valley, then we started to regress a little bit. We continued on the part of what we were doing, whereas there was a divergence somewhere between mid nineties where silicon valley started to actually develop a little bit faster and we started to lag behind. And I think as Rob said, that, 2014 was a good time because at that time there was a need for us to adopt technology to increase our efficiency and, fill the gap that was created due to capital constraint. And as well as fleeing of, some of the knowledge base, employees - from our sector.
Rob:
04:39
That's a good point on the technology side because you said that we kind of diverged away from where silicon valley really took off in the mid nineties. I entered into the industry in '94. So for me, my entire career has been that diverging process and just now it feels really good. Like we're finally catching up, not only catching up, but we've got customers, we've got employees who are sitting inside of the top tech companies in the world sitting at Google's facilities, even though they're an oil and gas company, sitting and working with Amazon, with Oracle, with IBM, with all these top names. And yet they're doing it in collaboration with the industry. Where in the past, it was almost like the two things were somewhat separated and now they are on a converging path. They've got the technology, we've got the data, at least in our space. And those two things coming together is kind of the critical mass we need to see some success.
Erica:
05:32
So on that note, what kind of jobs do you think are going to be created in the future as the industries continue to convergence?
Rob:
05:40
You know, that's a, that's a great prognostication. I mean, it's kind of interesting when you look back at like airbnb and Uber and those kinds of things. Nobody saw those coming and nobody knew what that was going to look like five years into their business, not to mention 10 or 15. I think that's what we're looking at in the oil and gas industry as well. We still have to find oil and gas. We still have to explore. We still have to be technologists, whether it's IT technology or G&G technology, we still have to operate in those spaces. But the roles may be very different. I'm hoping that a lot more of the busy legwork is a lot easier for us to work with and it has been historically, but we're still going to have to do those core G&G jobs. I just don't know what they're going to look like five years from now.
Arvind:
06:29
I mean the way I see it is that it will be high-gradation to, like it will be more fulfilling jobs. The future jobs hopefully will be more fulfilling. So because a good portion of the grunt work, the work that everyone hated to do, but they had to do it to get to the final work, like final interesting work. Hopefully all those things will this machine learning and AI and broader digitization will help alleviate that part. And even whether you are technologist, whether you are a geologist, whether you're a geophysicist or whether you're a decision-maker. Like in all of those, um, you will start moving from the low value work to high value work. The technologist who was looking into log curve, they will actually start evaluating the log curve rather than just digitizing it. And that's, in my view, it's a more fulfilling job job compared to just doing the mundane work. And I, so that's the part first part is that what kind of job it, my hope is that it will be more fulfilling.
Arvind:
07:43
Now the second is how many and what type of job, um, as Rob said that, the speed at which this is moving, we, it will be very difficult for us to do the prediction. Is that like if we sit here and say that they are, these are the type of job that will be created in five years, we'll be doing a disservice. We can actually make some guided prediction in which there will be need for geologist or geophysicist or petrophysicist and other people to do in what form will they be a pure geophysicist or a geophysicist who is a has a lot more broader expertise, a computer science and geophysicist working together. Those are the kinds of roles that will be needed in future because for a very long time we have operated in silos because it's not just technology is changing is the way we work is also changing is that we have operated in silos that we develop something, throw it over the fence. They, they catch it most of the time and then actually move into the next silo, and so on and so forth. Is that what-
Rob:
08:58
You hope they do anyway.
Arvind:
08:59
Yeah. I hope that they do anyway, but so that's the sequential process now. Some of them will be done by machines. Some of them will be done by human. And then you have to actually create a workflow which is like fulfilling as well as efficient for the capital investor.
Erica:
09:19
Perhaps less siloed off?
Arvind:
09:21
Less siloed off. So there will be team of teams and the team will actually move very frequently. So it will be almost like a self organization is that these are the four people needed to solve this problem. Let's take those four people and work on that problem. And then when that problem is solved or productionized, then they actually go solve the different problems.
Arvind:
09:43
And so it will rather than back in the days or even today, hi- fully hierarchy of system, it will still be there, will be CEO (Laughter) and but there will be more, um, team of different group and different expertise, um, very quickly building and dismantling and those, that's the agile methodology that will be needed to take this technology and use it for, like basically doing things better.
Erica:
10:18
So to kind of hone in on where you're saying, your background is in both geophysics and um, software engineering, correct?
Arvind:
10:26
Okay. So sorry, I didn't actually talk about myself. (Laughter) So, um, I joined the TGS a little more than a year back, um, started as a chief geophysicist and then moved into this role. But before that, most of my career has been with BP and before that for a software company. So I have worked as a software engineer for some time and then got my PhD in geophysics and then worked for a little more than 10 years in BP all the way from writing imaging.
Arvind:
11:01
So basically fundamental imaging, algorithm writing to drilling wells. So, in my short career I have seen a lot of things and what I do see is that, there has, there is a lot of silos in BP as well as in TGS. And BP is also working on it - breaking. I have a lot of friends there who are saying is that there is a significant effort in technology and modernization is happening in changing the culture rather than- it's not just about changing PC from going from a laptop to iPad. That's a- that's a tool. But the fundamental change will happen in the thought process. And if we want to actually use machine learning and these kinds of digital technology then it needs to be very integrated and the silo mentality is not going to work. You have to look at the problem as a holistic to solve it.
Erica:
12:02
Yeah.
Arvind:
12:02
So, so that's the background. So that's my background.
Erica:
12:05
Yeah. So I asked because I wondered if you think that your career path is going to be the future of the industry, do you think that there are going to be more people with a dual background in both computer science and geophysics?
Arvind:
12:19
So that's a very polite way to say that. My, I am actually looking at that my career is the right career. So, no and yes and no both. I do think that people will become more generalist and they will have deep expertise. And it's counter intuitive - is that generalist and deep expertise is not the same. Like we are used to someone who has a very deep expertise and that are not generalists about other topics
Erica:
12:57
Narrow and deep.
Arvind:
12:57
So very narrow expertise. But very deep and they have shallow expertise, very broad. Those are back in the days I think we are moving towards a deep expertise in several different narrow fields. So you need like, so to truly get good collaboration and innovation, you have to have deep expertise in several different fields to integrate them together.
Erica:
13:27
So Rob, it looks like you're chomping at the bit here. (laughter)
Arvind:
13:30
Deep and broad. So like what we need is deep and broad.
Rob:
13:34
Yeah. When, when Arvind was talking about, kind of the career and, and some of the other topics, two things came to mind on the technology side of things. If you look back at AT&T, they had a choice and they did investigation and some pretty deep research on whether or not they needed to move into mobile cell phone technology. And they made the choice. They did a big expensive study and spent hundreds of millions of dollars or tens of millions of dollars to identify that they needed to be prepared for an industry of say, a million cell phone users by a certain year. And that number was, I don't know, 150 times wrong. It was way, way higher than that. And you could use the same thing with Kodak. They invented the digital camera and then lost the digital camera battle. And struggled in the industry. We want to make sure that we're looking broad enough to understand what's coming down the pipe and can adapt and change to that. Not just from the individual roles in the company, but the company direction as a whole.
Arvind:
14:34
To give a concrete example is that , I have a background in geology or physics and computer science or Rob has background in geoscience and computer science and the data analytics team. It likes our TGS data analytics team. They have, we have people who have the um, physics backgrounds. They have PhD in physics and then they have worked in geophysics and then working on well logs. Then, the other one, Sathiya - he is a geophysicist who now is working on more of a deep learning problem. And a Sribarath is the team leader. He is a geophysicist. Who is it more of a computer scientist who is working on these two problems. So, our team composition itself, the TGS data analytics team composition itself is built in a multidisciplinary fashion.
Erica
15:30
Yeah. So I'm glad that you brought up are our current team here cause I kind of wanted to pivot to the problems that we're using AI to solve for right now. You know, like what, what are the pain points in the industry and how are we using AI for that?
Arvind:
15:46
So, so the pain point in the industry, are I'll talk about one, is it one which is very close to my heart. I was a, so in BP I did a lot of salt interpretation. So anything which requires a lot of human intervention is a big choke point because our data set is getting bigger, larger and larger with a lot more volumes to it are a lot more information to it and we have limited human resources and we want to actually take those human resources and mobilize them to do more high value work rather than doing a lot more um, grunt work. Salt model building is an example. And where we, we actually, our data analytics team started working there. So I'll, I'll work, I'll talk about that later. But that's an example where a lot of judgment call is made early, which don't require a lot of human judgment call early interpretation. Is the true place where automation and digital transformation can actually help.
Erica:
17:04
Rob, what's your take on this?
Rob:
17:06
Well, the Nice thing about what we're doing with salt picking is we're really helping us and our clients reduce the time it takes to get to the indecision. On my side of,of the house, my background with TGS has largely on the well data side of things. So it's not so much about reducing the amount of time of processing the data as it is getting a higher value data set in the hands of our clients. So historically, especially in the onshore U.S., there's a significant lack of data that's reported to the regulatory agencies. So we source that data as do a lot of other people. We source data from our, our, our customers, our partners operators. We process that data, but the most important thing that we can do with that is take that huge volume of data, the largest commercially available in the industry and add more to it so that the operators are able to get to that decision making process. So like Arvind said, if we can avoid the grunt work and get them to the point where they're actually making business decisions, that's what we're doing with our analytics ready LAS Dataset. We're in-filling the gaps in the curves because they either weren't run or weren't reported. We're predicting what the missing curves would look like, based on an immense volume of data. So it's not so much about getting the product created faster, although that is another goal that we've got. Of course, we're a commercial company. We're trying to get products to our customers and make money like anybody does. But the ultimate goal with our current analytics ready LAS product is to get the most complete dataset available so that the operators can make better decisions in the subsurface; drill less wells, drill more productive wells, drill wells faster. All of those things go into why we chose to go down that that path.
Arvind:
18:50
So, looking at a higher level. The question that you asked was like what are the choke points and how we had actually using digital transformation in machine learning and AI to help that. Um, I think we published something like our CEO talked about that in the um, few months, a month back, Norwegian Energy day. There was a nice plot that, shows that most of the time we are acquiring data for a purpose. Like we are acquiring data to solve a geologic problem so that we can actually make a decision whether to drill somewhere, or not drill somewhere whether to buy acreage or not buy acreage by our clients. So when you take that data, you have to convert that into information, that information need to convert it into knowledge. And that knowledge is what enables our clients to make better, faster and cheaper decisions.
Arvind:
19:51
And that cycle converting from data to knowledge to decision and enabling their decision is actually is the big choke point. If you want me to say one, this is that your point is that how to actually take data and convert to knowledge fastest way and cheapest way. And that's where most of our effort is. So salt, model building is an example where we right now it takes us somewhere between the nine months to a few years when we acquire data to provide the clients with the final image that they can do interpretation and make decision. This is too long of a time. In this day and age it needs to be compressed and a good portion of that compression can happen, by better compute. But some of them cannot happen without doing a deep learning where humans are involved in like for example, salt models building where like you can actually throw as much computer it as possible. But since the cycle time requires human to drill that model, it will be the limiting cases that, so there we want to actually enable the interpreters to take our salt net, which is our algorithm and accelerate the early part of it so that they have more time to do high quality work and build and build that model faster, reduce that cycle time so that our clients can make better, faster and cheaper decisions.
Rob:
21:32
It's been interesting to watch the transition too with our industry and the technology at the same time we've moved to the cloud, right? All of our data's now sitting at a cloud provider and if you would have looked at the oil industry five years ago, there's a very security minded mindset around the industry that says, I need to keep that data because it's a very, very critical and I want to make sure the only, I've got access to it. So there was a lot of fear about putting data in the cloud several years ago. Now you look at the cloud providers and they're spending literally billions of dollars on things like security and bandwidth and access, things that didn't exist five, 10 years ago. So that transition to be able to go to the cloud, where all, where, all of our data sits today. More and more of our clients are going there as well. And the nice thing about that is you can ramp up your needs, on compute capacity, on disk capacity, on combining data sets across partners, vendors, other operators, and collaborate and work on that data set together to come up with solutions that you couldn't possibly have done before. So it's, it's fun actually to watch that transition happen.
Arvind:
22:43
It is going a little tangent to the question that you asked her, but, because there's a very important point about the cloud services the the biggest cloud platform is Kubernetes by Google and that's actually open source. So Google developed that and made it open source available for anyone who wants to build a cloud infrastructure. They can have it. That's the, the most to use open source, platform that, available today. So that's changing the way people work. Like red hat or Linux, Unix, Sun, Sun, microsystem or Microsoft or apple. They are very, like, even in technology sector, they are very controlling of what they are providing to their consumers. They control that environment. Whereas now things are changing in which the open source systems like, which is publicly available is becoming one of the most dominant form of a software platform. Um, if you look at android for machine learning, it's tensorflow, Pi Torch. Those are open systems software that is a democratizing the technology so that anyone and everyone can, is able to take that next step and the solve complex problem because the base is available for them. They don't have to build the base. They can actually focus on solving the high value complex problem.
Erica:
24:24
Speaking of both Google and open source and democratizing, problem solving. So TGS recently had a Kaggle challenge, correct, can you speak a little bit about that?
Arvind:
24:35
So, yeah, that actually, so when I joined TGS, I had, one data scientist that we, we were working with, like we were still building the data science team and we started working on the salt net problem. We had an early, um, success. We were able to do some of those things and then we realized that there is like ocean of data scientists who are across the world. We don't have actually access to that Google actually open source and they have, they're working on their problem, they're working on Apple's problem, they're working on very interesting problems. So why they're not working on it at two different reason. One is that they don't have access to it in a second, the problem is not interesting enough for them. So Kaggle was our effort to make it accessible to everyone and make it interesting so that people will work on it.
Arvind:
25:30
So just for the, um, description of Kaggle, Kaggle is the world's largest, data science crowdsourcing platforms. So crowdsourcing is a, um, where you put the problem and it's a platform or website where the, um, the problem description is given and data science scientists to work on their like on their spare time, nights and weekend or that's their hobby or that's their job. And they solved that problem. They submit to submit on that platform and they get instantaneous result that, how a good their solution was. So that's the Kaggle is the one of the largest world's largest platform for that recently acquired by Google. So we actually approached Kaggle that- can we actually put the one of the complex problem that we have on this website or this platform and they worked with us. And so we partnered together to host the oil and gas first serious problem for the automatically building salt model. And we actually, so to Rob's point, um, the hardest problem was getting the data rights that are convincing our management that it's okay to release a certain portion of data. We had to work really hard to create an interesting problem and that once we released that data, um, this competition was very successful in the sense that if they were around 80 plus thousand different solutions, just think of the scope of it
Rob:
27:06
From almost 3000 different teams
Arvind:
27:09
3,800. So close to 4,000 people. Oh yeah. 3000 team and comprise of almost 4,000 data scientists across the world work on this problem for three months and gave us more than 80,000 different solutions. We would have never got anything like this working day and night with whole TGS working on this problem.
Rob:
27:32
I, I found it interesting because I like did a search on Google for our, TGS salt net.
Arvind:
27:39
Yeah.
Rob:
27:40
And if you look at the results just on Youtube, you'll find probably 20 different videos of PhD students, data scientists getting their master's degree who are using that problem that we posted out there as part of their thesis or as part of their Grad student work to show that, that the data science process that they went through as part of their education. And now that's out there for everybody to use.
Erica:
28:02
So this is a major disruptor isn't it, to the industry because we have basically non geologists, non geophysicists solving problems for-
Rob:
28:12
Yeah it's, it's definitely, we, there was a lot of teams, right? So there was some that had geoscience backgrounds, some that didn't, but most of them, they just come from a data science background, right? So they could have stats or math or computer science or anything. And when they applied this, it was interesting to see the collaboration on the Kaggle user interface where the teams were out there saying, hey, I tried this. What did you guys try? And the whole idea of crowdsourcing and, and the idea that we're kind of in somewhat of a unique position where we can do that. We can, we own the data. We don't license it from somebody else. Um, it's the data that we own that we can put out there. So we've got a huge volume that we can leverage and put it into a community like that where we can actually see some of those results come in.
Erica:
28:57
So to kind of put you on the spot-
Arvind:
28:59
Can I- one thing to say after that to is not just about data owning the data because there are several different companies who own data, even oil and gas company, they have their own data library. I honestly think that, it says volume about TGS, that TGS was willing to take a bet on this kind of futuristic idea and like go on a limb. But, and this is, I'm just giving credit to the senior management here, that they were, they're allowed us to actually go with this. That was one of the bigger hurdle than just to owning data, that management buy-in
Rob:
29:39
Second only to data preparation for the challenge itself.
Arvind:
29:42
Second only to the data preparation, it took us a lot of time to build-
Rob:
29:45
Yeah
Arvind:
29:45
an interesting problem. It's not just about like you have to create an interesting problem to-
Erica:
29:51
to attract the right talent.
Arvind:
29:52
So the winner was a group from a Belarus and the Japan. They have never met. They have never seen each other other than the Facebook.
Erica:
30:02
Wow.
Arvind:
30:03
And did they actually met on this Kaggle platform? They were working on this problem. They found out that there they are approaching with the two different ways and they actually teamed up so that they can combine this to create a better solution. Combining both of their effort and that that's actually happens to be the winning combination. But a traditional method won't allow us to tap into this kind of resources or brain power. That to someone from Belarus and Japan working together whom we don't know solving our problem and that is going to be a disruptor and we have to be ready to capitalize on it rather than be afraid of it.
Erica:
30:51
Right. And that's why I wanted to go to rob, not to put you on the spot here, but as someone coming from the well data side, do you see any potential future Kaggle challenges using well data?
Rob:
31:05
Yeah, the, that could absolutely be in our future. I think at this point we're really trying to frame the problems that we're trying to solve for our customers. And if we decide that one of those problems deserves, some time in the public, like on Kaggle, then we can absolutely go that direction. Not a problem whatsoever. At the moment though, our real focus is trying to figure out where can we provide the most value to the clients and we're kind of letting them steer us in a, you know, a way we have got our own geology department internally so we know what we need to do with our internal well data in order to high grade it to the next level product. However, we're really taking direction from our clients to make sure that we're moving in that direction. So yeah, I could see us having a problem like that, especially if it's starting to get into a Dataset that, , needs to be merged with another data set that maybe, we need support from, somewhere else in the industry. We're in a different industry.
Arvind:
31:59
Just a few minutes on that is,the next problem I think that Kaggle need from oil and gas is a more on the solution side. So the knowledge to- like information to knowledge site in which you are all taking very different type of data set. For example, success failure database for the basin. And building a, prospect level decision that requires a, as Rob said, that collaboration, that the TGS collaborating with one of the E&P company or someone else, like those two or three companies and now bringing their data together because at the end of the day, this integration is what everyone is looking for. Can we actually create an interesting integration problem and put it on the Kaggle competition. So, any listener, if they're in, they have a good problem, they can actually contact Rob, or me. That, because we are always looking for good partners to solve complex problems. We can't solve all the problem by ourselves, neither other people. It does require teams to build the right kind of Dataset, interesting problems in to, to get into the board.
Erica:
33:22
Okay. So we've talked about how we got here to this point in the industry with AI machine learning and we've talked about what we're doing today with the, um, let's move on to the future where we think AI will take, um, the industry. So to follow up on something that Arvind had said earlier, so you had said that we sort of fell behind silicon valley at some point. How, how far behind do you think we are right now in terms of years if you can make that estimation?
Arvind:
33:58
Oh, that's a tough question but I'll try to answer it in a roundabout way. Is it that when I say that we lag behind, we lag behind in the compute side of it, like the AI side of it and some of the visualization and web-based technology when it comes to high performance computing, we were still leading up to very- probably in some of the spaces we are still leading. So storage and high performance compute which is both, oil and gas defense and Silicon Valley. All three are working. Um, we are not that far behind actually we might be at the cutting edge of it. And that was one of the reason that we didn't actually focus on the AI side because we were solving the problem in more high compute way and we are using bigger and bigger machine solving, more complex problems more physics based complex physics based solutions.
Arvind:
35:04
So when it comes to solving physics based solution, we are still, at the front of the pack. But when it comes to solving a heuristic auto machine learning or AI based solution, we are behind, we are behind in robotics and things like that and we are catching up. So when you think of a mid midstream and downstream where there's a lot of the internet of things, IOT instruments, so things are getting is like instrumentized and there are a lot of instruments which are connected to each other and real time monitoring, predictive maintenance. Those are happening and happening at a very rapid rate. And that will actually, we'll, we'll catch up in a few years in, in midstream and downstream side or mostly instrumentation side where we are truly lagging is subsurface because it's not the problem that Ian, and like, silicon valley was trying to solve.
Arvind:
36:05
A subsurface problem are complex. They are very different type of problem; that someplace you have very dense data, someplace We have very sparse data. How to actually integrate that and humans are very good at integrating different scale of information in a cohesive way, whereas that problem is not the problem that silicon like, technology sector was trying to solve. And so we are trying to actually take the solutions that they are building to solve different problem and integrating it or adapting that to solve our problem. So that's where like I see like, so I think it's a non answer but that's what the best I have. (Laughter)
Erica:
36:49
It was a very good answer. So how does this change the way that we're building our products then our approach to getting our products out there?
Rob:
36:58
Well, one of the, one of the things I'll start with is we're actually seeing our clients adopt analytics teams, analytics approaches, machine learning. there's a lot of, there's a lot of growth in that part of the industry. and they've gotten past the point where they don't believe that a predictive solution is the right solution. You know, with our ARLAS product, we're creating an analytics ready LAS dataset where we're predicting what the curves would look like, where there's currently gaps in the curve coverage. The initial problem the customers had was, do they believe that the data's accurate? We're starting to get past those kinds of problems. We're starting to get to the point where they believe in the solutions and now they're trying to make sure that they've got the right solutions to fit within their workflows in their organization. So I think the fact that they've actually invested in building up their own analytics teams where they've injected software engineering, geology and geophysics, a data science and kind of group them all together and carved them off, or they can focus on maybe solving 20% of the problems that they actually, attempt. That's kind of where the industry has gotten to, which means we now have an opportunity to help them get to those levels.
Arvind:
38:10
You see that a change in conferences, and, meetings and symposiums that, like for example SEG Society of exploration geophysicists and, that, conference three years back there was one session about machine learning and last year, machine learning has the largest number of sessions in that conference. So you're looking at a rapid adaptation of a machine learning as a core technology in oil and gas and at least in subsurface, but most of them is at the very early phases, people are trying to solve the easier problem, the problem they can solve rather than the problem that need to be solved. So that's where there's a differentiation happening that everyone wants to work on machine learning and most of the people are actually taking solution to your problem rather than taking problem finding solution for a problem which is relevant. So,
Rob:
39:21
I think that's pretty fair because,you've got to get some sort of belief internally and if you can prove that you've got kind of a before and after, here's what I did to make this decision or the wells that are drilled in the production I've got and here's what I predicted was going to happen. And you can start to see those two things align. Then you start to get belief in something. If you just use something that's predictive only and you've got nothing to compare it to, it may be the right solution. But do you have the belief that your company is going to run with it? So that's why I think we're starting to see them solve problems that we know can be solved initially rather than the big problem of say, if I shoot seismic here, I can predict how much oil I'm going to produce. That's a big problem and it's at different resolutions and scales than we believe we can solve and, and be definitive about it today. but I think that, I think I agree with you that they're, they're really focused on, on proving that this technology, that analytics that AI/ML is going to work for the problems that they know about.
Arvind:
40:24
Agreed only up to a point is that, the reason and why I think it ML/AI solutions are different is because, in physics, one of our basic assumption is that, if we solve a toy problem, you can scale the same way is the same solution will apply on a bigger problem. That's not the case for machine learning solutions. The solution that is applicable for a toy problem is not going to scale. You need to actually retrain the data and the solution becomes different as the scale of the problem increases. So although it's, interesting to see that a lot of a small problem are very easy problem people are taking to- people are solving a lot of easy problem using machine learning. To show that machine learning works, that's good. But to truly take advantage of machine learning, you have to actually solve, try to solve one of the complex problem because you already have a solution for those easy problems.
Arvind:
41:40
Why do we need machine learning? So for example, ARLAS is a good example. Our analytic ready LAS in which we are predicting well logs from the available, well logs. Now if I have only one well, or a few wells then I actually want my petrophysicist to go through the physics based modeling and solve that problem. I don't need AI to solve that problem. I have actually solutions which works there. If the solution that I need is that how to solve this problem on a scale of Permian basin or a scale of U.S. So like what we have done for ARLAS that the first basin we started was Permian is where we took all the data that we have as a training data or actually a good portion of that data as a training data set. We build that model, which is actually based in scale model that can actually ingest all the like 320,000 wells we have. So we used thousands and thousands of well as a training build a very robust model to actually solve that problem and now that solution is available for the whole basin. That's the kind of solutions that are problem that AI is good at solving and has actually best potential not for solving few wells. Learning about AI by solving a few wells is great, but as a product or as a true application of AI, we need to actually look at tackling the big problems.
Rob:
43:11
Yeah, I agree. There's been a lot of, shall we say analytics companies that come out with a claim of being able to perform some sort of machine learning basis and they've got a great interface and everything looks really good. And the story behind it is that it's been taught on five wells or 10 wells in our learning set was in the tens of thousands of wells, which is why I believe in the data set that we've built.
Arvind:
43:40
At a very high level, machine learning is like teaching a kid, like someone has come out of graduate school and they want to actually learn something and you are showing them this is how we actually do. The more things they see, the better they will get, the more experience they will have and the better their capability or work will be. So it requires the, the whole concept of machine learning or AI is that you want to actually train with massive amount of very high quality data set and that actually solves more complex problems.
Erica:
44:18
How do you discipline data?
Arvind:
44:22
So you are saying that did- have you talked to our lead data scientist and he calls him to himself a data janitor, that most of the time he spent is cleaning of the data and organizing the data so that he can actually do the high quality like the machine learning AI work. So if he spends his time like out of a hundred hours, 60 or 70 hours- so he's actually organizing, categorizing data set so that he can do the fun stuff in the last 30 40 hours. I mean that's actually, that's better than a good, most of the places where people spend 90 hours doing the curation and 10 hours doing the fun stuff. And that was one of the reasons why we had to build the data lake because one of the thing is that we need all the data to be readily available in a kind of semi usable format that I don't need to spend time learning about the 2003 data is different than 2015 data versus 2018 data.
Arvind:
45:34
I need to actually consume it as one big dataset. So last whole year we spend actually considerable, considerable amount of time and effort in building our data lake in which we actually took all of our commercial legacy, data set and moved it on cloud. The two things that we did is one we standardized the data set so that lead data scientists don't have to spend on doing janitorial of data janitorial work and a second is creating metadata. So what Metadata is that aggregate information.
Arvind:
46:06
For example, Arvind Sharma what is the Meta data about Arvind Sharma um, that he is five feet 10, I don't have a lot of hair. (Laughter) He drives some car and he, he has gone to- he has a PhD like so some aggregate information like out of her, like rather than cell by cell information about Arvind, what is the minimum, set of aggregate information that you can use to define Arvind. So that's the metadata about any data set. So what we did when we are moving this a massive amount of data set into our data lake for each of these data set, we extracted this aggregate information that where it was recorded, when it was recorded, what are the basic things done to this data set? What is the maximum amplitude in this volume? What is the minimum amplitude in this volume? What does the average amplitude in this? So those things we actually use it because a lot of analytics is that some of the higher level analytics will be about integrating the information about data set, like Facebook uses information about people to make some of the decision. We are not that creepy as that Facebook, but (laughter) it's, it's like taking the information about the data set and actually learning creating knowledge about the basin.
Rob:
47:37
It's interesting when you were talking about the data janitorial work and how we've kind to standardize our data set on the, on the cloud because it kind of brings it full circle back to something you said early on. And that was that we want our customers to be able to get to that decision making point sooner without having to do all that data, janitorial work. I've been going to data management conferences for 25 years and I hear the same thing every year for 25 years. I spend "fill in the blank" percentage of my time, 60 70, 80% of my time looking for data and the remainder are actually working with it. That's what an analytics ready data set it's going to allow us and our customers to be able to do is not have to do all that janitorial work, but actually get to the point where I can actually start interpreting what that data means to me to make decisions.
Erica:
48:30
So looking towards the future of the industry, do you think we're going to continue to ramp up in terms of speed and getting to the good stuff, the fun part? Do you think that's going to continue to logarithmically increase?
Rob:
48:44
Probably faster than we can ever imagine. I think the, I think the change that we saw with companies moving to the cloud companies going toward, service based solutions, companies moving toward high volume, normalized consistent datasets, all of these things have been moving at light-light speed compared to what they were, the, the past 25 years. Up until today, every day about probably about every three weeks. We basically, have got some new technology that's been released that we can start adopting and putting into our workflows that wasn't there three weeks, three weeks prior, open source. It comes back to that topic as well. More and more of these tech firms are putting the data out as open source means we could leverage it and get to solutions faster. So to answer the question, absolutely faster than we can possibly imagine.
Erica:
49:28
Well, awesome. I cannot wait to get to this future, with both of you.
Erica:
49:41
Well, thank you so much for talking with us today. Being part of our first episode of Beneath the Subsurface, it was an absolute pleasure. If our listeners want to learn more about what TGS is doing with AI, you can visit TGS.com You can visit our new TGS.ai platform and, we'll have some additional show notes on our website, to go along with this episode.
Arvind:
50:06
Thank you Erica.
Rob:
50:07
Yeah, thanks a lot. I appreciate it.
Conclusions and plugs:
Check out the newly launched tgs.ai to dig deeper in to the data with subsurface intelligence. Gain detailed subsurface knowledge through robust analytics with our integrated data and machine learning solutions at tgs.ai Discover Geoscience AI solutions, Cloud Computing, Data Management, and our Data Library. Learn more about TGS at tgs.com
