Introductory presentation for an external Master's thesis. Kai is advised by Prof. Bader, Dr. Ban-Sok Shin and Sebastian Wolf.
Previous talks at the SCCS Colloquium
Kai Nierula: Physics-informed Deep Learning for Wave-based Seismic Imaging
SCCS Colloquium |
Physics-informed deep learning has recently emerged as an alternative to classical solvers for partial differential equations (PDEs). In this thesis, conditional generative adversarial networks (cGANs) will be used for simulating seismic wave propagation and seismic inversion. In the first step, a single cGAN will be trained to learn wave propagation in an acoustic medium. By incorporating the underlying physics in the training loss function of both the generator and discriminator and adding the subsurface velocity distribution as a condition, we obtain a physics-informed cGAN. In the second step, a second cGAN will be used in conjunction with the first one to obtain a subsurface velocity model from measurement data. Using physics-informed cGANs instead of classical methods should allow for a faster solution of the inversion problem while still retaining high accuracy. Additionally, the use of cGANs allows for uncertainty quantification, providing a possibility of measuring the confidence in the network's prediction.