Machine learning and seismic interpretation

Machine learning uses supervised and unsupervised learning methods to recognize and identify a similar pattern in the geological or geophysical data. This figure shows the unsupervised and supervised machine learning workflow. Supervised classification depends on a labeled training dataset. It uses a regression technique that predicts a continuous measurement for an observation. Unsupervised classification does not depend on labeled responses in order to draw inferences from a dataset. It uses a clustering method to group similar hidden patterns in the data together.

Machine learning in seismic interpretation uses computer algorithms to help geologists understand the relationships between large amounts of geological data or information. The computer algorithm is trained from an input data and then adapts independently to produce repeatable and reliable results that can be used for seismic interpretation. Machine learning is useful in seismic interpretation because it solves two significant problems that seismic interpreters face: interpreting large volumes of data and understanding the relationship of various types of data at once. The two types of machine learning methods used in seismic interpretation are unsupervised classification and supervised classification. Supervised classification has limited application in seismic interpretation because of the earth's heterogeneity, which can make interpretation of geologic features away from wells more difficult.On the other hand, unsupervised classification is less limited and does not need the desired answer to be known in advance in order to understand the pattern in the data.

Supervised Machine Learning

In supervised machine learning, the correct or desired answers is known based on a training dataset that has known results and then the training dataset is applied to another dataset with unknown results in order to predict its output. Optimally, the learning algorithm will be able to determine class labels for unknown or unseen instances.

Types of supervised learning algorithms:

1. Multi-layer perceptron: A neural network where the data flows from the input to the output layer and there is one or more layers between the input and output layer. Multi-Layer perceptron is used in seismic interpretation for pattern recognition, approximation and classification.
2. Probabilistic neural network: An important method in seismic interpretation used for pattern recognition and classification that computes the distances between the input vector and the training input vector and then produces a vector that shows how close the input data is to the training data. Then, the second layer of the algorithm creates a sum of the contributions for each of the input classes to output a vector of probabilities. Lastly, the maximum of the probabilities is decided and classified as 1 while the other probabilities are classified as 0.

Unsupervised Machine Learning

Multi-attribute machine learning interpretation steps.

In unsupervised machine learning, the unsupervised algorithm is used to figure out and understand the patterns in datasets that consist of input data without labeled responses. It can be used to show subtle geological features that were missed using conventional analytical methods and can be applied to predict fluid properties, lithological variations, wells analysis, and for selecting the best seismic attributes to be used for seismic interpretation.

Types of unsupervised learning algorithms:

1. Principal component analysis (PCA): A very important classification technique in seismic interpretation. It is a linear mathematical method that reduces large sets of seismic attributes to a smaller set of attributes that still have most of the variations of the larger set. To find the most useful attributes within a large dataset, the first PCA accounts for as much of the variation from the dataset as it possible could. Then, the following components find the remaining variability. So, the seismic attributes that contribute the most to the initial principal component and the next few succeeding components indicate best seismic attributes that should be selected for seismic interpretation of the data.
2. K-Means clustering: This is an algorithm used to partition data into distinct clusters based on distance to the centroid of a cluster.
3. Hidden Markov models: This is an algorithm that uses observed to recover the sequence of states.
4. Hierarchical clustering: This is an algorithm that builds a multi-level hierarchy of clusters by creating a cluster tree.

Seismic Interpretation with Self-Organizing Maps

This is a self-organizing map created to explore the mineral deposits in an area.

Self-organizing maps (SOM) are a type of unsupervised machine learning technique that classifies data into clusters, categories, and pattern based on their properties. SOM is significant in geology for identifying important geological features for oil exploration such as thin beds, DHI features, seismic facies, and sweet spots for shale plays.

The interpretations steps to use for SOM:

1. Define the geologic problem that is needed to be resolved such looking stratigraphy, DHIs, or bed thickness.

2. Choose the appropriate variation of seismic attributes is selected based on background knowledge and experience on the types of attributes used for the proposed feature of interest or based on principal component analysis which reduces the large set of seismic attributes down to a smaller set of attributes with the most of the variation using a linear mathematical technique.

3. Run the SOM with the different combinations of attributes to create a 2D colormap that can be interpreted to resolve the geologic problem. The 2D color map can be interpreted by isolating various combinations of neurons to in order to identify different geologic features. In general, self-organizing maps allow for a rapid and accurate comparison of large sets of seismic attributes to reveal seismic anomalies in the data which helps in the process of identifying hydrocarbons and geologic features.

Supervised Classification for Seismic Facies Analysis

Seismic facies analysis important method for understanding important information about subsurface structures, lithology variations, and distribution of reservoirs based on seismic and well log data. This method is important for decreasing risks associated with drilling and for finding economical traps and can be used during different stages of petroleum exploration as development, optimization, and management. First, in order to produce or construct an accurate seismic facies map, similar seismic traces are classified based on frequency, amplitude, phase, and seismic attributes extracted from well logs and seismic data. Although past methods have mainly used unsupervised methods to classify similar seismic traces and create seismic facies maps, machine learning and supervised classifications has been shown to map reservoir and seismic facies distributions with less relative errors and with more efficiency.

[1] SeIsmic facies analysis of carbonate reservoir from oil field in Iran using K-nearest neighbor classifier.

With supervised classification, training datasets such as well log-facies and several seismic attributes are used to extract a volume of seismic facies. When selecting attributes, it is important to evaluate the extracted attributes by using methods such as backward feature selection and forward feature selection to ensure that relevant attributes are selected. As a side note, the cons of training datasets is that collecting labeled datasets is expensive because it is obtained through electrofacies analysis. Electrofacies analysis is usually performed using gamma-ray logs, sonic logs, neutron porosity logs, and bulk density. Furthermore, supervised classification trains a learning algorithm that uses known/labeled input data and desired output data in order to create a classifier. A classifier is a discriminator function that maps sample on class labels. Some of the popular classifiers include support vector classifier, multilayer perceptrons, Fisher, K-nearest neighbor, and Parzen. After the classifier is trained on the facies information from the available seismic traces and logs, it can be applied to the seismic data to predict and analyze the facies between the wells. In conclusion, using a supervised classification method for seismic facies analyses can be effective by choosing the choosing the best classifier and seismic attributes.  

See Also

Facies Classification using machine learning