Chapter 1 - Course Introduction

01-01 - Course Introduction (20 min.) Sample Lesson Quiz: 01-01 - Course Introduction

Chapter 2 - Introduction to Well Log Data and Petrophysics

02-01 - Introduction to Well Log Data and Petrophysics (14 min.) Quiz: 02-01 - Introduction to Well Log Data and Petrophysics

Chapter 3 - Understand The Fundamentals Of Petrophysics

03-01 - Introduction to Machine Learning (18 min.)
03-02 - Workflow of an ML Project (14 min.)
03-03 - Coding with AI (4 min.) Quiz: 03-03 - Coding with AI

Chapter 4 - Training vs. Test Data In Petrophysics

04-01 - Training vs. Test Data in Petrophysics (17 min.) Quiz: 04-01 - Training vs. Test Data in Petrophysics

Chapter 5 - Python And Petrophysical Data

05-01 - Python and Petrophysical Data (16 min.)
05-02 - Checking The Python Setup (4 min.)
05-03 - Reading LAS Files (13 min.)
05-04 - Some Basic Examples (16 min.)
05-05 - Data Display With Commercial Software (7 min.)
05-06 - Pre-Processing Data (8 min.)
05-07 - Pre-Processing Examples (20 min.) Quiz: 05-07 - Pre-Processing Examples

Chapter 6 - Clustering Methods For Petrophysical Data

06-01 - Clustering Methods for Petrophysical Data (13 min.)
06-02 - K-Means Clustering Exercise (17 min.)
06-03 - K-Means Clustering Exercise - Visualization (16 min.)
06-04 - Key Takeaways (6 min.) Quiz: 06-04 - Key Takeaways

Chapter 7 - Classification Methods For Petrophysical Data

07-01 - Classification Methods for Petrophysical Data (9 min.)
07-02 - Ensemble Methods (14 min.)
07-03 - Random Forest - Net Pay Exercise (16 min.)
07-04 - Net Pay Exercise (Continued) (12 min.)
07-05 - Lithology Classification Exercise (13 min.)
07-06 - Lithology Classification Exercise (continued) (10 min.)
07-07 - Exercise Overview & Key Takeaways (8 min.) Quiz: 07-07 - Exercise Overview & Key Takeaways

Chapter 8 - Regression Methods For Petrophysical Data

08-01 - Regression Methods for Petrophysical Data (6 min.)
08-02 - Common Regression Methods (8 min.)
08-03 - Regression Exercises - Overview (3 min.)
08-04 - Regression Exercise 1 (22 min.)
08-05 - Regression Exercise 1 - Key Takeaways (5 min.)
08-06 - TOC Prediction - Exercise 2 (4 min.)
08-07 - TOC Prediction - Key Takeaways (11 min.)
08-08 - Regression Summary (3 min.) Quiz: 08-08 - Regression Summary

Chapter 9 - Model Deployment

09-01 - Model Deployment (14 min.) Quiz: 09-01 - Model Deployment

Chapter 10 - Course Wrap-Up

10-01 - Course Wrap-Up (14 min.) Quiz: 10-01 - Course Wrap-Up
10-02 - Bonus Lesson on LLMs (10 min.)

Practical Machine Learning for Well Log Data / Chapter 1 - Course Introduction

Lesson 01-01 - Course Introduction

Access All SAGA Wisdom Training content Subscribe

Already a member? Sign in

Access All SAGA Wisdom Training content Subscribe

Already a member? Sign in

Transcript

01. Lesson 1.01: Course Introduction 02. Course Overview 03. Course Outline 04. What You'll Learn 05. About Me 06. The Business Case for ML/AI 07. Productivity Increases 08. Get Answers Faster 09. Get Better Answers 10. More Consistent Answers 11. ML as a QA/QC Tool 12. Course Outline

01. Lesson 1.01: Course Introduction

Hello everyone. My name is Cameron Snow, and I'm pleased to welcome you to my course:Machine Learning for Well Log Data.

02. Course Overview

So I think we've got a great course lined up for you here. And this course is going to cover a wide array of material. What I want to do here in this first lesson is introduce you to the course and to myself and what it is that we are going to be covering and what you can expect from this course.
So let's start with the topics that we'll be covering in this course. We're going to try to cover a very broad array of topics that will give you all the tools that you need to start to apply machine learning to your well log data.

03. Course Outline

So here's some of what we're going to cover. We're gonna start off by giving you an introduction to well log data and petrophysics. Now if you're an expert, so if you are extremely familiar with well log data or petrophysics, then you can feel free to skip this part and move on to the next section where we start to dive into more of the machine learning side of the course. However, I do think this is a good review and that you'll benefit and you'll make sure that we're on the same page with respect to terminology as we start to progress in the course.
Once we cover some of the basics of petrophysics, then we're going to proceed onto machine learning techniques. And I'm gonna start off by giving you some general definitions here and covering some background material. So that's what we'll do in the second section.
We're then going to go on to kind of a special mantra of mine, which is how we do training vs. test data in petrophysics. Now this is something that'll seem very obvious once we discuss it, but this is something that I've seen a lot of people get wrong in academic studies and out in the wild. So I feel like it's worth specifically covering.
Then we're going to go into Python and specifically Python for petrophysical data. So how do we read in well logs? How do we set up our Python environment? How do we make sure that everything is going to be up and running?
And then we're gonna dive into the machine learning techniques. So we'll start off with clustering methods. Then we'll move into classification methods. Then we'll cover regression methods. Then we're gonna talk about how you can deploy those machine learning models and some of the considerations around those. And then we will have an overview of the course.
Now I did add one more section here, which are LLMs for petrophysics, and we'll be covering that as a bonus section. And the reason I say that is a bonus section is because this is a field that is very rapidly evolving, and perhaps by the time that you watch this, even just a few months after the recording, it's possible that this could already be out of date, but I still wanna cover that for you in this section.

04. What You'll Learn

Let's take a look at what you will be learning here. What I really wanna cover are what are some of the business drivers for machine learning. If we're not applying our machine learning to real business problems, I personally think we are just doing science projects and we're doing vanity projects. So I think it's very important that we always keep the business case in mind. And as we're going through real-world examples, I'll try to make sure that we always put this and frame it in the context of business. And I'll try to make it seem like a real exercise that you may do.
I want you to take away the fundamentals of machine learning and AI approaches. You're not going to necessarily understand the exact mathematical internal workings of these methods, but what I do hope you take away is the underlying concept that you can then use to explain these models and justify why you were using each specific one.
And of course we're gonna cover the application of those models to petrophysical data. That's the whole point of this course.
As we're doing this, we're gonna talk about model performance and how you can evaluate that. It's very important to understand how your model is performing and to put that in the right context.
And as we're going through, we're going to be doing projects, you'll be gaining hands-on experience.
Now, we're going to be applying this specifically to well log data. Now my reason for choosing well log data is for a couple of reasons. First, as a geologist and petrophysicist I work with well log data every day, and I think many of you probably do the same. Now, I think it is important that we try to apply these techniques to the data that we're working with the most. Second is, well log data is especially well-suited to machine learning in a lot of ways. It's very dense numerical data that is tabular in form in a lot of cases, and it really lends itself to machine learning techniques. So I think there's a lot to be said for applying this to well log data. However, the techniques you can apply to really any kind of data. Now in some cases you'll find that you don't necessarily have a large enough quantity of some types of data to really make meaningful models, but in general the techniques that we'll cover here you can apply to other datasets.

05. About Me

A little bit about me. Why should you care about what I have to say about machine learning and petrophysics? So I am a geoscientist and petrophysicist with about 20 years of experience. I've had the opportunity to work very broadly in my career. I've covered a number of different plays in different geographic regions. So I've worked across the US and Canada, South America, Africa and Europe. I have also worked on projects that are in a broad array of stages, going from development drilling, like you'd be doing in the Permian Basin, to frontier exploration plays in deep water and little-drilled areas. I think I have a good background to bring context to what we'll be learning here.
Also, as the co-founder of Danomics Petrophysics and the GeoRiskX platforms, I use machine learning on a day-to-day basis. We use it in both of those platforms to help our users accelerate and enhance their workflows.
And if you recognize me from somewhere else, it may be from my Applied Petrophysics and Geological Risking courses that you can also find here on the SAGA Wisdom platform.

06. The Business Case for ML/AI

Let's start off briefly by covering the business case for machine learning and AI. So what is it that you really want to get out of machine learning for your data?
So I think first of all, what you should look for is a productivity enhancement. Whether or not you're a geologist, a petrophysicist or reservoir engineer, there's always more work than you have time to do. And I think machine learning and AI really gives you a way to apply a multiple to your personal productivity. So I think that is the first area where you should be looking for gains.
Now the second area here is very similar, but slightly different. It gives you a way to get answers faster. So let's think about what an example of that may be. Let's say that you needed to do an interpretation for, let's say, total organic carbon (TOC). The way you would typically do this is you would go into a well, you would select a method, like let's say one of the PASI methods, and that means you'd have to set up a template to do an overlay. You would then need to adjust the clay resistivity for that well. You would need to set the maturity for that well to generate an answer. Now, with machine learning, there is a possibility that you could go in, interpret one well, interpret a small handful of wells maybe, use that data to train a model and then apply that model across all of your wells. So this takes you from the interpretation phase to the QC phase immediately. So I think that is an area where maybe your answer isn't getting any better, but you're getting it a lot faster.
Now the next thing is to actually get better answers. Now you may be wondering how can I get better answers from machine learning that I can generate myself or that my petrophysicist can generate for you? That's an excellent question and it's one that I have some internal back and forth with all the time. So when it comes to getting better answers, there's two ways that I like to think about it. The first way is when we look at who is applying the method and who is doing the interpretation. There's the question of, if you have an expert petrophysicist who has gone through and they've carefully analyzed a handful of wells to use as the training data for this model that they'll be making, and then they give that to a more casual user, let's say a newer geologist who doesn't have the breadth of experience, there is a high likelihood that that model will generate more accurate answers on wells it is predicting than the geologist would be able to generate themselves. And that's not saying that that would be the case if the expert was looking at it themselves for every well, but when you combine it that you have a number of different people that will do each of these exercises, you can either make your petrophysicist faster or you can make your geologist better and maybe in some cases you can do both. And I think this is one of those cases. The other area is, of course as petrophysicists we're always trying to fit our models to core data, and I do think that there is an opportunity if you have enough data to train your petrophysical models on core data, and then use machine learning models to generate predictions that are truly calibrated to core in a very quantitative way as opposed to a more qualitative fit as in, Oh, this looks like a good correlation on a crossplot or a good correlation in a log profile. Now there are lots of challenges about that and we can talk about that coming up, but so far you can get faster answers potentially, you can get better answers potentially.
And you can get more consistent answers. So one of the things that will change is when you are generating interpretations manually. So when I as a petrophysicist am going through, I'm having to set parameters, I'm making crossplots, I can get tired, I can be off a little bit one day, and maybe my answers just vary a little bit. Maybe I work on a project, I go away, I come back 3 months later, I don't quite remember all the details and my interpretation from a well I did in that first batch and a well I did in the second batch are not fully aligned. They're of course gonna be very similar because I'm still applying petrophysical techniques, but they're not really identical analyses. They're not apples for apples, so to say. And this is where I think AI and machine learning can be quite powerful, is once you have that model locked in, it will give you exceedingly consistent answers. So you can know that every well that has been ran through that model has been treated the same way. So if I do an interpretation, it's going to have my biases all the way through it. While the same can be said of a machine learning model, except its biases don't change day to day, week to week, and month to month. So I think it has the potential to give you much more consistent answers.
Those are all about answers. What about outcomes? I think machine learning data can help you avoid bad outcomes. And you'll hear a lot about this actually on the operational side of things. You'll hear about models that can help you predict when pumps may go down based on vibration or things like that. So you can go in there and you can preemptively stop that before anything happens that's negative. Well we could do the same thing with well log data. We could try to avoid bad outcomes. Now what's a bad outcome inpetrophysics? Well, a bad outcome may be drilling into an overpressure zone and not recognizing that, and that causes drilling hazards. So we may be able to use real-time data coming from the rig, feeding into your model, to help you recognize quicker what is potentially going to be happening in your well. So I think there's some potential to build models that can help with that.
At the end of the day though, what we wanna do is we want to create value. So I think all of these things, all of these potential use cases that we have here, are value creative for your company. And if it's not creating value in some way (now there are certainly probably other ways that you can create value other than what we have here), then I think your model is more of a science project or a vanity project just so you can say that you applied machine learning. So I want you to keep this in mind as you're going forward: how do you create value? How do you translate your model into a better outcome for your company?

07. Productivity Increases

So let's go through and we'll just briefly touch on a few potential examples here. So let's say we want to increase our productivity. Well, what's a way we can do that? So one of the ways that I do this on a day-to-day basis is through repairing washout intervals. So here in this well log, what we see in the upper portion where you have some pink shading in here, these are washouts. And we've identified those washouts, and we're using machine learning (either in this case a multilinear regression or a random forest model) to predict the curve response that we would expect based on the other data here. So we're using it to repair the bulk density curve. Now, typically the petrophysicist would go in and do this by hand. But with a little bit of smart programming and machine learning working hand in hand, we're able to do that automatically. So this can be done across thousands of wells without having to take manual intervention.

08. Get Answers Faster

What about getting answers faster? Like I mentioned, machine learning gives you a way to scale your expertise. So here we are training, as in the example I gave you earlier, thisTOC model. We are then running through this mode, we are generating predictions, and we can see our actual vs. our predicted both in crossplot and in log profile. So once again, that is giving you a very good answer.

09. Get Better Answers

And if we choose to look at this also for getting better answers, like I said, this model is performing very, very well. And so I would say this means that the model is accurately capturing the model of your expert and allowing you to distribute that knowledge. So now your geoscientist or your reservoir engineer, who's not as familiar with the techniques that we'd use for prediction, is able to generate an answer that is on par with the petrophysicist in this case.

10. More Consistent Answers

I did a fun little experiment just for this course. What I did, I asked a couple of my petrophysics buddies to go through and interpret a couple of wells for me. And what I wanted to do was to look at the variation from their interpretation to my interpretation and compare that to the model to my interpretation. And what we see is the model generates an excellent correlation here, starting at the origin at (0, 0), and really progressing on a unity slope. What we see here, when I compare my work to another expert petrophysicist, is we actually see that there is a misfit here, where even though we generate things on the same trend, what we have noticed is that one of us has an opinion that this system has higher TOC. So there's a way that, even though we don't know which answer is right, we can certainly generate a more consistent answer using the machine learning model.

11. ML as a QA/QC Tool

And one last use case here for you is looking at machine learning as a QA/QC tool. So maybe this you could classify in avoiding bad outcomes as the category. Here we have an area where we have some organic shale, so this may be an area that we'd be interested in evaluating for an unconventional play. And what we see here is that the red curve displayed in the right-hand track is what the machine learning tool is predicting. So it's predicting a relativelyTOC-rich interval. But in this case the manual interpreter has under-called TOC relative to the model. So we could use this as a way to flag the model to draw our attention to this well and say, hey, this is a well that we should take a deeper dive into. Once again, I think that is another excellent use case for machine learning and to help us (in this case) avoid the bad outcome of missing out on a potential new play.

12. Course Outline

That is gonna be one of the longer lessons that we have here, this course intro. There was a lot that I wanted to go over in terms of examples to get you ready for what is coming up.
And next we're going to be moving on to an introduction to well log data and petrophysics. I'll see you in the next lesson. Thank you.