HorizonMachineLearning

Machine Learning Horizon Guided attributes for reservoir properties prediction

General Concept

• Seismic attributes are extracted along interpreted horizons
  • A window of, let's say 20 ms, is used for the extraction of various attributes.
  • These attributes are then exported to an x y z flat file. Let's say we extract 10 attributes.
• Well petrophysical data, e.g. porosity, permeability or net to gross:
  • These are upscaled to the equivalent of the 20 ms window and supplied as a csv file.
  • A dataframe (a table) with all the attributes is generated:
  • Each attribute is listed in a column while the rows represent the individual locations, i.e. trace locations

Workflow I - Data Munging:

• Format the horizon files into one file
• Scale the horizon data
• Create a well file with all the attributes back interpolated at the well locations
  with the last column being the petrophysical attribute, e.g. permeability
• We are ready to apply Machine Learning

Workflow II - Data Analysis

• Check data distributions and statistical ranges
• Check for linearity between various predictors amongst themselves and with the target
  • Generate a matrix scatter plot
• Check for feature importance using using RFE (Recursive Feature Elimination)
• Check for feature contribution using PCA (Principle Component Analysis)

Workflow III - Model Fitting and Prediction:

swattriblist.py has many models to attempt to fit to your data they are all based on sklearn package.
CatBoost is installed and used instead of XGBOOST

Clustering

• KMEANS is first tested to identify the optimum number of clusters
• Once the optimum of clusters are found then KMEANS is applied to the predictors
• The resulting clusters are then one hot encoded to be added as predictors for further model fitting
• tSNE t distribution stochastic neighbor embedding. Attempts to project all your attributes to 2 components
• umap Uniform manifold approximation and projection. A powerful clustering technique to project data on to 2 or 3
  components

Regression

Below are the various regression techniques that can be applied  

• Linear Regression
• SGDR : Stochastic Gradient Descent with Lasso, Ridge, and ElasticNet options
• KNN : K Nearest Neighbors.
• CatBoostRegression
• NuSVR Support Vector Machines Regression
• ANN Regression Artificial Neural Network using Keras

Classification

Below are various classification models that can be used:

• LogisticRegression
• GaussianNaiveBayes
• CatBoostClassification
• NuSVC Support Vector Machines Classification
• QDA Quadratic Discriminant Analysis
• GMM Gaussian Mixture Model
• ANN Classification Artificial Neural Network using Keras

SemiSupervised Learning

Imbalanced Classification

Most of our data is imbalanced. These correction techniques apply to all classification models

• ROS Random oversampling
• SMOTE Synthetic Minority Oversampling
• ADASYN Adaptive Synthetic Sampling