Unconventional reservoir parameter estimation by seismic inversion and machine learning of the Bakken Formation, North Dakota

Abstract

The research reported in this thesis focuses on the prediction of reservoir parameters and their uncertainties. The thesis comprises two studies. In the first part, I focus on quantitative and seismic interpretation problem, where I describe a workflow for estimation of porosity using the results from pre-stack seismic inversion. The second part focuses on the production problem, where I establish a relationship between completion parameters and production given a production dataset from the Bakken Formation. In the first study, I characterize the unconventional reservoir of the Bakken Formation, specifically within northwest North Dakota using 3D seismic and well log data. I employ seismic inversion followed by application of a Bayesian Neural Network to predict total porosity across the entire seismic volume given an estimated volume of P-impedance. The Bayesian Neural Network utilizes Markov Chain Monte Carlo via Langevin Dynamics in order to sample from the probability distribution and to estimate uncertainity. This method establishes a good correlation between estimated P-impedance from seismic inversion and total porosity from well data. By integrating these techniques, a better understanding of the parameters useful for reservoir characterization is possible given a degree of uncertainity thereby improving oil and gas exploration and risk assessment. In this second study, I make use of a production dataset of the Bakken Formation to identify production patterns in the field to establish a relationship between completion parameters and production. A random forest model is employed alongside the Bayesian Neural Network model to predict production given a set of predictive features found through a series of feature selection methods. I then aim to create various training and testing dataset scenarios through random sampling and clustering. I do this in order to reduce the sampling bias and ensure that the machine learning models are being trained and tested on data coming from similar geological regions with similar production rate values. With the integration of these techniques, a better understanding of the parameters useful for optimizing oil production is possible with a degree of uncertainity when using the Bayesian Neural Network.