Machine-Learning Method Determines Salt Structures From Gravity Data

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 201424, “Machine-Learning Method To Determine Salt Structures From Gravity Data,” by Jie Chen, Cara Schiek-Stewart, and Ligang Lu, Shell, et al., prepared for the 2020 SPE Annual Technical Conference and Exhibition, originally scheduled to be held in Denver, 5-7 October. The paper has not been peer reviewed. In the complete paper, the authors develop a machine-learning (ML) method to determine salt structures directly from gravity data. Based on a U-net deep neural network, the method maps the gravity downward continuation volume directly to a salt body mask volume, which is easily interpretable for an exploration geophysicist. The authors conclude that the ML-based method from gravity data complements seismic data processing and interpretation for subsurface exploration. Introduction In subsurface exploration, seismic is the dominant method used to reconstruct the underground image for geophysicists and geologists to locate possible hydrocarbon reservoirs. Seismic acquisition is carried out by human-induced sound waves (by airgun or vibrators) that are recorded, once reflected, on the surface. Through the iterative waveform inversion process, a subsurface image can be reconstructed for reservoir location and property determination. Nonseismic (gravity and magnetic-measurement) methods, on the other hand, are passive measurements and not intrusive to the environment. In gravity data acquisition, gravimeters measure the change in the gravitation-al field, which can be used to determine the density variation on the subsurface. Compared with seismic acquisition, gravity acquisition is cheaper and introduces a much smaller carbon footprint. Gravity data resolution is, in principle, worse than that of seismic. However, especially in areas of salt structures, gravity data provide a unique addition because the density contrast between salt and the surrounding sediments in-creases with depth, while the velocity contrast decreases with depth. Therefore, gravity data provide valuable additional constraints in salt delineation for interpretation and seismic processing. Recently, ML and deep-learning (DL) applications in hydrocarbon exploration have been studied extensively. The authors note developments such as use of ML/DL on seismic data noise attenuation, salt interpretation from seismic stack, least-square inversion, rock-facies classification, and 4D seismic in reservoir management. To the authors’ knowledge, no literature exists that explores use of ML on nonseismic data. The authors’ method can map the gravity downward continuation volume directly to a salt body mask (0/1 for nonsalt/salt) volume, which saves iterative effort of the conventional gravity inversion process and is easily interpretable for explorational geophysicists and geologists. Gravity Data Processing Raw gravity data are measured as a 2D Bouguer anomaly (the difference between measured gravity and theoretical gravity value) grid. The first step of gravity inversion is to perform a downward continuation calculation to generate a 3D volume so that the depth of the density anomaly can be estimated. The equivalent source technique is one of the more-stable downward continuation calculations and is a preferred method for making downward continued volumes used in in-field reference drilling.