Machine learning applications for seismic processing and interpretation

During the past few years, exploration seismology has increasingly made use of machine learning algorithms in several areas including seismic data processing, attribute analysis, and computer aided interpretation. Since machine learning is a data-driven method for problem solving, it is important to adopt data which have good quality with minimal bias. Hidden variables and an appropriate objective function also need to be considered. In this dissertation, I focus my research on adapting machine learning algorithms that have been successfully applied to other scientific analysis problems to seismic interpretation and seismic data processing. Seismic data volumes can be extremely large, containing Gigabytes to Terrabytes of information. Add to these volumes the rich choice of seismic attributes, each of which has its own strengths in expressing geologic patterns, and the problem grows larger still. Seismic interpretation involves picking faults and horizons and identifying geologic features by their geometry, morphology, and amplitude patterns seen on seismic data. For the seismic facies classification task, I tested multiple attributes as input and built an attribute subset that can best differentiate the salt, mass transport deposits (MTDs), and conformal reflector seismic patterns using a suite of attribute selection algorithms. The resulting attribute subset differentiates the three classes with high accuracy and has the benefit of reducing the dimensionality of the data. To maximize the use of unlabeled data as well as labeled data, I provide a workflow for facies classification based on a semi-supervised learning approach. Compared to using only labeled data, I find that the addition of unlabeled data for learning results in higher performance of classification.. In seismic processing, I propose a deep learning approach for random and coherent noise attenuation in the frequency – space domain. I find that the deep ResNet architecture speeds up the process of denoising and improves the accuracy, which efficiently separates the noise from signals. Finally, I review geophysical inversion and machine learning approaches in an aspect of solving inverse problems and show similarities and differences of these approaches in both mathematical formulation and numerical tests