A Primer on Machine Learning in Subsurface Geosciences

In the first chapter, we will learn about big data, data analytics, and machine learning, as well as their utilities in different disciplines, including geosciences. A thorough understanding of data analytics problems, algorithms, and geoscience-specific data is critical before applying these sophisticated tools. We will also go into the brief history of the advent of different machine learning algorithms, the different types of geoscience databases, and their fitness to machine learning applications.

As geoscientists, we use a variety of data types (see Chapter 1). To analyze different varieties of geodata, we must use different statistical measures. A thorough understanding of various statistical measures and their applications to data analysis methods is important. In this chapter, we will learn about fundamental concepts in statistics and different data analysis measures. This chapter begins with the basic concept of random variables, which are widely used in statistics. We then move on to univariate, bivariate, time series, spatial, and multivariate graphing and analysis techniques.

Data-driven machine learning (ML) approaches are becoming very popular in analyzing facies, fractures, faults, rock properties, and fluid flow in subsurface characterization and modeling. Our reservoirs are becoming more data-rich due to the advent of new drilling, completion, and sensor technologies. Modeling of these variables from such data-rich reservoirs is a complex multivariate, multiscale, and multidisciplinary problem that we can handle with ML algorithms. In this chapter, we will learn about the fundamental steps in deploying ML models to solve our problems. Although there are several ML models available, including both traditional algorithms and deep learning algorithms, some steps are very similar. The selection of a particular algorithm over other depends on the data and the problem itself, complexity, interpretability, time, and cost. In this chapter, we focus on the nature of the problems and provide a systematic guide to building, evaluating, and explaining these data-driven models, irrespective of the algorithms.

In the last several decades, computer scientists and statisticians have developed and implemented a plethora of machine learning (ML) algorithms. Although the application of data-driven modeling is relatively new to geoscience, we can trace back some of its early applications to the 1980’s and 1990’s. This chapter will discuss the fundamental theory and analytic framework of many popular ML algorithms. Understanding the fundamentals of these algorithms, network-specific hyperparameters, and their meaning is essential to better implement these algorithms in our datasets and enhance the success rate of data-driven modeling. These algorithms are based on solid mathematical and statistical theories. Indeed, some algorithms are better than others for certain types of applications; however, sometimes, our lack of understanding of algorithms and the nuances of their applications to specific datasets cause them to underperform compared to others. Once we understand the fundamentals of algorithms and our datasets, ML will be more fun and provoking, which will facilitate further progress of geo-data science.

Geoscientists have been implementing machine learning (ML) algorithms for several classifications and regression related problems in the last few decades. ML’s implementation in geosciences came in different phases, and often these broadly followed or lagged after certain advances in computer sciences. We can trace back some of the early applications of modern ML techniques to 1980–1990. Geoscientists were mostly dealing with deterministic analytical solutions at that time, and they were encouraged to do so at their organizations. This is also the time when geostatistics started flourishing in reservoir characterization and modeling efforts. Then, the early 2000’s saw a slight uptick in ML applications, mostly neural networks and decision trees. Since 2014–2015, a lot of ML-related work was published. In addition to open-source languages, this also has to do with access to the massive volume of data from unconventional reservoirs. And then, since 2017, there has been an explosion of deep learning related work. This again corresponds to the convolutional neural network architecture published by Goodfellow in 2014. Initially, the ML work in geosciences focused on petrophysics, seismic, and now core and thin section images. Another growing trend is the application of ML in passive geophysical data analysis (seismology, gravity, and magnetic, etc.). As of now, most of the published studies on ML are confined to outlier detection, facies, fracture, and fault classification, rock property (e.g., poro-perm-fluid saturation-total organic carbon-geomechanics) prediction, predicting missing logs/variables, and well log correlation. In this chapter, we will review some of these popular research problems tackled by ML.

In the last chapter, I discuss the future of data analytics (DA) and machine learning (ML) in geosciences research, instruction, community, and business, as a whole. It sets an agenda of ML-focused studies that need to be conducted to solve critical problems in geosciences. This endeavor will not only help understand the fundamental geologic processes and better analyze rocks but also assist the businesses to make better decisions and grow as needed.