12. Machine Learning for Integrating Well Data and Seismic Attributes

Training data preparation
User can prepare training data for supervised classification or reservoir property prediction from upscaled well data based on well type or interpreter’s prior knowledge. User can label training data either automatically based on either facies log or polygons.
Data analysis
User can analyze training data by using the following tools
  • Histogram distribution
  • correlation table and
  • cross plots between upscaled well data and seismic attributes
Attribute dimension reduction and selection
User can estimate the “best” seismic attributes that are sensitive to the prediction target property by the following algorithms.
  • Principal Component Analysis
  • Neighborhood Component Analysis
  • Step-wise regression
Unsupervised classification
Seismic facies map, facies grid, or facies volume can be generated by the following algorithms
  1. Self-organized map (SOM)
  2. Hierarchy classification
Supervised classification
Facies log, lithofacies or oil and gas “sweet-spots” map, strata-grid, or volume can be generated using the following supervised classification algorithms.
  • Waveform correlation map
  • Deep-learning neural network
  • Bayes classification to integrate well facies log and seismic attributes
Reservoir property prediction
User can use the “best attributes” obtained in “Attribute selection” step to predict pseudo-well log volume through the following algorithms
  • Multivariate linear regression
  • Deep-learning neural network
Facies volumes or maps can be generated from neural networks or Bayes classification
Machine Learning Performance analysis
A set of tools to analyze the performance of neural networks are provided in a “Performance Window”, where user can visualize the partial dependance of each input attribute to the target variables, as well as the relative importance of all input attributes. Contributions of individual wells can also be visualized. Facies classification accuracy can be viewed for each class.