Machine Learning and Full Waveform Inversion

Project description
Machine learning has been used successfully in both the natural and social sciences to process large amount of data and, for example, to determine and analyze complex patterns in these data. Since geophysics, and especially seismic exploration, also involves the processing of large amount of data it is natural to apply machine learning algorithms to seismics and, in particular, to the most challenging part of seismic processing: the generation of accurate velocity models of the subsurface.

Determining Earth’s subsurface properties is one of the most challenging tasks in geophysics, both theoretically and computationally. Typically, this is done using either imaging or full waveform inversion (FWI). The latter is more expensive but also more accurate. In practice imaging/FWI consists of optimizing the misfit between the observed and computed waveforms using gradient based methods. One particular type of machine learning method, stochastic gradient descent (SGD), consists of a variation of gradient based methods, and this method has already successfully been used in FWI. One of the goals of this thesis is to further develop existing FWI methods that use SGD. A second goal is to apply neural network algorithms to FWI and to compare these to the SGD inversion results.

The default modeling method in imaging and FWI, both for acoustic and elastic media, has been finite difference modeling. An alternative modeling method, which is faster and specifically valid at higher frequencies, is based on the ray-Born approximation. FWI using ray-Born modeling method has successfully been done for a number of models and codes are available. It has not been used in conjunction with SGD however. Therefore, a first task in this project would be to implement SGD for ray-Born based FWI. Various implementations of this already exist and so it is likely that this part of the project can be finished relatively soon. The second part of the work then involves using a neural network based optimization algorithm for the inversion. Special attention will be given to speeding up the neural network algorithm, using shot encoding and/or reparameterization. which will then be used to compare this new algorithm with SGD as well as conventional inversions.

Proposed course plan during the master's degree (60 ECTS)
Inf264, Mat160, Geov272, Geov302, Geov375, AG335