DoodleTrain

In 2026, machine learning appears to be everywhere. You can run apps that will identify a weed in your garden, a bird from its song in your tree, and even a rock in your hand. Although we are excited that machine learning may automate repetitive tasks in seismic interpretation (like horizon and fault picking), we are fearful that the machine will eliminate our jobs. In my opinion, the best way to address this fear is to dive in and spend a day learning what today’s machine learning algorithms can and cannot do.

We will do this by driving some state-of-the-art machine learning software ourselves. To do so, we will need to look at the seismic data like a computer does. What features in the seismic response differentiate one facies of interest from another? Seismic attributes measure most of these features. Next, how might we describe those differences to a computer? Then, we need to manually define these zones in the data to generate a training data set. After training a given algorithm on the training data, we then apply it to the target area, typically between two horizons, but depending on the objective, sometimes to the entire volume. After displaying our results, we need to calibrate them. Most commonly, we use our understanding of geologic processes to determine if the results make sense. In other cases, we may have well control or manually interpreted seismic lines not used in the training. In the first case, we are using ground truth, in the second, interpreter experience, insight, or hypothesis testing.

There are two different types of questions we can ask a machine learning algorithm. The most obvious request is “I have defined and labeled a facies I am calling X. I have defined training data that are X and training data that are not X. Please examine each voxel in the volume and tell me the probability that it is facies X.” We call this process “supervised learning.” An alternative request may be “I don’t really know what kind of seismic facies I have in this data volume. I’ve defined training data that are randomly selected from the data volume. Please partition the data into clusters representing statistically distinct seismic responses, give them a computer-generated label (with a distinct color), label similar facies with similar colors, and assign the most likely facies to each voxel. We call this process “unsupervised learning.” Not surprisingly, we can bias the training data for unsupervised learning to favor specific facies of interest. Here, we may have 33% of the training data represent facies Y, 33% facies Z, and 34% randomly selected from the data volume. We call this process “semi-supervised learning” or “unsupervised learning with interpreter bias.”

In this short course we will evaluate all three of these machine learning approaches. We begin with attribute-driven machine learning and use crossplotting and principal component analysis (PCA) to determine which attributes are independent and which attributes best represent the variation in the data. Then we use the first two or three principal components and proceed to the simplest unsupervised machine learning algorithm, kmeans, to define a finite number of clusters that represent seismic facies in the data that are self-similar. The first two eigenvectors used in PCA analysis define a 2D plane that captures the maximum variance contained in the N attributes that define an N-dimensional space. In self-organizing and generative topographic mapping, we deform this plane into a curvilinear manifold that better represents the data. By construction, clusters with similar colors represent facies with similar seismic responses. Mapping interpreter-defined seismic facies onto this manifold generates a probability density function (PDF) of those facies; applying this PDF to the mapping at each voxel provides a level of supervision to results in a semi­supervised clustering algorithm. We complete our tour of attribute-driven machine learning algorithms with a supervised learning algorithm – random forest decision trees, where we will explicitly define a suite of seismic facies and then ask the computer to assign a probability of each facies at every voxel.

We complete our driving by examining two deep-learning algorithms. While some deep-learning algorithms

use only seismic amplitude, often supplement these data with seismic attributes. They also require a very large amount of training data. Because of this, most algorithms come pretrained, where the interpreter can add additional, survey-specific, training using a process called “transfer learning.” In this course we use this deep-learning workflow to pick faults. A more recent innovation is a process called “segment anything” that is much more interactive. Here we will use it to map channels and other features on time and horizon slices.

At the end of this course, you should have a good feeling for what ML can and cannot do. It is quite good at extracting patterns in the seismic data. Whereas color blending allows us to combine the information content of three, and sometimes four patterns, or attributes, machine learning can combine the information content of four or more attributes. In contrast, today’s machine learning algorithms know very little about geologic processes, or limitations in the input seismic data associated with acquisition and processing. To address this limitation, the interpreter needs to skillfully bias the ML algorithms to provide answers that are geologically meaningful and supported by the measured seismic data. In short, ML is another tool that with experience and care can be used to better interrogate Earth’s subsurface

Hands-on Exercises

Pre-course preparation

1. Software installation, seismic data loading, and seismic data display

½ Day One

2. Interactive multiattribute analysis using visualization
3. Kuwahara filter attribute image enhancement
4. Multiattribute dimensionality reduction and projection techniques using principal component analysis (PCA) and independent component analysis {ICA)
5. Unsupervised clustering using k-means

½ Day Two

6. Unsupervised clustering using self-organizing maps (SOM) and generative topographic maps (GTM)
7. Semisupervised learning by biasing the training data and using a posteriori analysis
8. Supervised learning using random forest decision tries
9. Convolutional neural network fault prediction
10. Interactive analysis using Segment Anything