Automatic Differentiation for Geophysical Inversion

朱尉强 博士

加州理工学院

2021.11.05（星期五）15:00，腾讯会议号：572 931 127

报告摘要：

Geophysical inversion is commonly used to constrain physical parameters of the Earth, such as its internal structure, that cannot be measured directly but instead must be inferred from observations. In contrast to complex geophysical inversion procedures (e.g., the adjoint-state method), modern deep learning frameworks (e.g., Tensorflow) automatically optimize neural networks without case-by-case derivation and implementation. To fill the gap between these two fields, we developed a general seismic inversion framework using automatic differentiation called ADSeismic. We demonstrated its capability to solve problems of velocity model estimation, rupture imaging, earthquake location, and source time function retrieval within a unified framework.

This inversion framework further enables the optimizing of both neural networks and PDEs using automatic differentiation. To this end, we developed a new inversion method, NNFWI, to integrate neural networks (NN) into full-waveform inversion (FWI) by reparametrizing the velocity model with a generative neural network. NNFWI demonstrates the successful combination of the feature learning capability of deep neural networks and the high accuracy of PDE solvers to improve geophysical inversions. The generative neural network provides a new regularization approach based on Deep Image Prior and a new uncertainty quantification approach using the Monte Carlo dropout technique to make inversion results more robust to local minima (cycle-skipping) and more stable to noisy seismic data.

报告人简介：

朱尉强，现加州理工博士后。在北京大学取得学士(2013)和硕士(2016)学位。在斯坦福大学取得博士学位(2021)。博士期间主要研究方向: 深度学习在地震观测中的应用, 深度学习与反演问题的结合, 地震数值模拟。