Accelerating geostatistical seismic inversion using TensorFlow: A heterogeneous distributed deep learning framework. (March 2019)

• Record Type: Journal Article
• Authors: Liu, Mingliang; Grana, Dario
• Is Part Of: Computers & geosciences. Volume 124(2019)
• Page Start: 37
• Page End: 45
• Publication Date: 2019-03
• Subjects: Seismic inversion -- Geostatistics -- TensorFlow -- GPU; Environmental policy -- Periodicals; 550.5
• DOI: 10.1016/j.cageo.2018.12.007
• Languages: English
• ISSNs: 0098-3004

Abstract

Abstract: Geostatistical seismic inversion is one of emerging technologies in reservoir characterization and reservoir uncertainty quantification. However, the challenge of intensive computation often restricts its application in practical studies. To circumvent the computational limitation, in this work, we present a distributed parallel approach using TensorFlow to accelerate the geostatistical seismic inversion. The approach provides a general parallel scheme to efficiently take advantage of all the available computing resources, i.e. CPUs and GPUs: the computational workflow is expressed and organized as a Data Flow Graph, and the graph can be divided into several sub-graphs which are then mapped to multiple computing devices to concurrently evaluate the operations in them. The high-level interface and the feature of automatic differentiation provided by TensorFlow also makes it much easy for users to implement their algorithms in an efficient parallel manner, and allows employing programs on any computing platform almost without alteration. The proposed method was validated on a 3D seismic dataset consisting of 600 × 600 × 200 grid nodes. The results indicate that it is feasible to the practical application and the computational time can be largely reduced by using multiple GPUs. Highlights: We proposed a parallel approach for geostatistical inversion using TensorFlow. The parallel scheme is user-friendly and highly scalable. Adam is used as the optimization method to efficiently update model parameters. The approach can largely reduce the computing time using GPUs. The presented method can be widely used in other inverse problems.