Decentralized Full-Waveform Inversion

With the advent of efficient seismic data acquisition, we are having a surplus of seismic data, which is improving the imaging of the earth using full-waveform inversion. However, such inversion suffers from many issues, including (i) substantial network waiting time due to repeated communications of function and gradient values in the distributed environment, and (ii) requirement of the sophisticated optimizer to solve an optimization problem involving non-smooth regularizers. To circumvent these issues, we propose a decentralized full-waveform inversion, a scheme where connected agents in a network optimize their objectives locally while being in consensus. The proposed formulation can be solved using the ADMM method efficiently. We demonstrate using the standard Marmousi model that such scheme can decouple the regularization from data fitting and reduce the network waiting time.