Multi-scale, Multi-dimensional Reservoir Characterization Using Advanced Analytics and Machine Learning

Abstract

This thesis aims to investigate novel approaches in the field of Machine learning and advanced data analytics that can handle large data volumes and open new doors in the field of reservoir characterization. To begin, a new approach for rock typing is introduced using fractal theory where conventional resistivity logs are the only required data. Fractal analysis of resistivity logs showed that the fractal dimension of these logs which is a measure of the variability of the signal, is related to the complexity of the rock fabric. the fractal dimension of multiple deep resistivity logs in the Cooper Basin, Australia was measured and compared with the fabric structure of cores from same intervals. The results showed that the fractal dimension of resistivity logs increases from 1.14 to 1.29 Ohm-meter for clean to shaly sands respectively, indicating that the fractal dimension increases with complexity of rock texture. The thesis continues with a machine learning application to augment/automate facies classification using resistivity image logs. Given the complexity of the application, a supervised learning strategy in combination with transfer learning was used to train a deep convolutional neural network on available data. The results show that in the absence of other information/logs, the trained network can detect image facies with a testing accuracy of 82% form electric image logs and a proposed post-processing method increases the final categorization accuracy even further. An important step in reservoir characterization is understanding and quantification of uncertainty in reservoir models. In the next section a novel Generative Adversarial Network (GAN) architecture is introduced which can generate realistic geological models while maintaining the variability of the generated dataset. The concept of mode collapse and its adverse effect on variability is addressed in detail. The new architecture is applied to a binary channelized permeability distribution and the results compared with those generated by Deep Convolutional GAN (DCGAN) and Wasserstein GAN with gradient penalty (WGAN-GP). The results show that the proposed architecture significantly enhances variability and reduces the spatial bias induced by mode collapse, outperforming both DCGAN and WGAN-GP in the application of generating subsurface property distributions. Finally, an advanced analytics technique for efficient history matching is proposed in the appendix. In this part of the thesis, an ensemble of surrogates (proxies) with generation-based model-management embedded in CMA-ES is proposed to reduce the number of simulation calls efficiently, while maintaining the history marching accuracy. History matching for a real field problem with 59 variables and PUNQ-S3 with eight variables was conducted via a standard CMA-ES and the proposed surrogate-assisted CMA-ES. The results showed that up to 65% and 50% less simulation calls for case#1 and case#2 were required.