A novel full waveform inversion method using machine learning and compressive sensing techniques

What is it about?

Full waveform inversion (FWI) can provide quantitative and high-quality images and models of the subsurface that worth exploration. However, current FWI methods are too computationally expensive because they solve very huge optimization problems. We develop a new FWI method that can significantly accelerate the computational speed of FWI with the help of machine learning and compressive sensing techniques. The key to solve this problem is to compress the subsurface model image from, say, millions of variables to much fewer. In order to maximally compress the subsurface model images, we implement adaptive transforms specifically designed for them by machine learning techniques. The kernel of these transforms is a dictionary learned from previously obtained model images by efficient machine learning algorithms. By this means, we are able to accordingly reduce the amount of data used for computation. We design a new optimization problem to respect the new method and solve it with efficient stochastic optimization algorithms. The tests indicate that our method can significantly reduce the amount of data as well as the computational cost needed for FWI.

Featured Image

Why is it important?

Our method significantly reduces the computational cost of the recently emerging full waveform inversion (FWI) problem so that even one without computational capacity (like Supercomputers, Computer clusters) can still afford that. Meanwhile, our method is the first attempt of introducing machine learning to solve the complex FWI problem. Our method proves that the machine learning technique can effectively compress the model images and hence reduce the amount of input data for calculation. Besides, we have also developed our FWI platform as an open-source software and the project can forked from: https://github.com/zhulingchen/SSSI

Perspectives