Forecasting Petrophysical Trends Based on Well Logs and Seismic Attributes Using Supervised Machine Learning Algorithms in the Teapot Dome Field, Wyoming

The objective of this research is to forecast petrophysical trends at the Teapot Dome field, Wyoming using supervised machine learning algorithms based on a combined use of well logs and seismic attributes. Thirteen instantaneous attributes and two reservoir properties, porosity and water saturation, were selected to set up the training structure after depth conversion and data re-sampling. Three algorithms, including kernel ridge regression, decision tree, and artificial neural network were proposed to build machine learning models. The optimization was done by minimizing the errors between the model-predicted results and the values of the targets. The performances of the models were quantified by the mean absolute errors, and each individual trained model was compared by a cross-validation approach. Further evaluation was done by three testing wells that were not used in the training. The tests indicate that, for my data and test parameters, decision tree is more useful than kernel ridge regression and artificial neural network in terms of the testing accuracy and effectiveness. Feature selection in decision tree was tested as a useful tool to reduce the data dimensionality, so that the redundant features and information can be removed without damaging the training accuracy. Based on the trained model, the distribution of porosity, and water saturation were simulated effectively and efficiently in the Teapot Dome field.