Multi-Task Deep Learning for Multi-Parameter Elastic Inversion

Abstract Elastic waveform inversion plays a vital role in estimating the Earth's subsurface property. The inversion of multiple elastic parameters from observation data has been regarded as a cutting-edge and challenging issue due to its severe non-linearity and ill-posed nature. Deep learning approaches have recently demonstrated incredible potential in simulating non-linear mapping and made remarkable achievements in geophysical inversion. In this work, we consider multi-parameter elastic inversion as a multi-task learning problem and propose to accomplish the three tasks sequentially by a deep neural network with sequential decoders, which we name ElasInvNet. Specifically, we reconstruct the three elastic parameters, P-velocity, S-velocity, and density, one after the other, and use the features from the former task as the prior information to assist the later tasks' reconstruction. We verified the effectiveness of the proposed ElasInvNet through comprehensive comparisons and ablation studies.