The development of machine learning techniques and the digitization of the subsurface geophysical/petrophysical measurements provides a new opportunity for the industries focusing on exploration and extraction of subsurface earth resources, such as oil, gas, coal, geothermal energy, mining, and sequestration. With more data and more computation power, the traditional methods for subsurface characterization and engineering that are adopted by these industries can be automized and improved. New phenomenon can be discovered, and new understandings may be acquired from the analysis of big data. The studies conducted in this dissertation explore the possibility of applying machine learning to improve the characterization of geological materials and geomaterials. Accurate characterization of subsurface hydrocarbon reservoirs is essential for economical oil and gas reservoir development. The characterization of reservoir formation requires the integration interpretation of data from different sources. Large-scale seismic measurements, intermediate-scale well logging measurements, and small-scale core sample measurements help engineers understand the characteristics of the hydrocarbon reservoirs. Seismic data acquisition is expensive and core samples are sparse and have limited volume. Consequently, well log acquisition provides essential information that improves seismic analysis and core analysis. However, the well logging data may be missing due to financial or operational challenges or may be contaminated due to complex downhole environment. At the near-wellbore scale, I solve the data constraint problem in the reservoir characterization by applying machine learning models to generate synthetic sonic traveltime and NMR logs that are crucial for geomechanical and pore-scale characterization, respectively. At the core scale, I solve the problems in fracture characterization by processing the multipoint sonic wave propagation measurements using machine learning to characterize the dispersion, orientation, and distribution of cracks embedded in material. At reservoir scale, I utilize reinforcement learning models to achieve automatic history matching by using a fast-marching-based reservoir simulator to estimate reservoir permeability that controls pressure transient response of the well. The application of machine learning provides new insights into traditional subsurface characterization techniques. First, by applying shallow and deep machine learning models, sonic logs and NMR T2 logs can be acquired from other easy-to-acquire well logs with high accuracy. Second, the development of the sonic wave propagation simulator enables the characterization of crack-bearing materials with the simple wavefront arrival times. Third, the combination of reinforcement learning algorithms and encapsulated reservoir simulation provides a possible solution for automatic history matching.