Improving Seismic Wave Simulation and Inversion Using Deep Learning

Abstract

Accurate simulation of wave motion for the modeling and inversion of seismic wave propagation is a classical high-performance computing (HPC) application using the finite difference, the finite element methods and spectral element methods to solve the wave equations numerically. The paper presents a new method to improve the performance of the seismic wave simulation and inversion by integrating the deep learning software platform and deep learning models with the HPC application. The paper has three contributions: 1) Instead of using traditional HPC software, the authors implement the numerical solutions for the wave equation employing recently developed tensor processing capabilities widely used in the deep learning software platform of PyTorch. By using PyTorch, the classical HPC application is reformulated as a deep learning recurrent neural network (RNN) framework; 2) The authors customize the automatic differentiation of PyTorch to integrate the adjoint state method for an efficient gradient calculation; 3) The authors build a deep learning model to reduce the physical model dimensions to improve the accuracy and performance of seismic inversion. The authors use the automatic differentiation functionality and a variety of optimizers provided by PyTorch to enhance the performance of the classical HPC application. Additionally, methods developed in the paper can be extended into other physics-based scientific computing applications such as computational fluid dynamics, medical imaging, nondestructive testing, as well as the propagation of electromagnetic waves in the earth.