Physics-Guided Machine Learning Approach to Characterizing Small-Scale Fractures in Geothermal Fields

Abstract

Characterizing fracture zones is crucial in geothermal exploration, drilling, and development. We aim to characterize small-scale fractures with scales less than the seismic wavelength. Recently, machine learning (ML) methods have been popular in interpreting large-scale faults by finding offsets in seismic images. However, such offsets may not be associated with small-scale fractures. By shooting a seismic beam from the Earth’s surface to subsurface fracture zones, we can extract a receiver-beam interference pattern created by fracture-generated multiple-scattering waves in observed seismic data. The double-beam interference pattern is a two-dimensional image that carries information about the discrete fracture network pattern. Under the ideal situation (e.g., perfect data acquisition and homogeneous medium), the beam inference pattern and the discrete fracture pattern are Fourier transform pairs. However, in real-world cases, such a Fourier transform relation is perturbed. We need to train the machine learning algorithm to be able to handle such a physical constraint. To demonstrate the capability of the method for small-scale fracture characterization, we construct a subsurface model containing small-scale fractures based on the Soda Lake geothermal field. We perform seismic modeling to generate 3D seismic data and apply our method to the data to characterize the discrete fractures, which are almost invisible on conventional seismic migration images.