Simplifying Full Waveform Inversion via Domain-Independent Self-Supervised Learning

Geophysics has witnessed success in applying deep learning to one of its core problems: full waveform inversion (FWI) to predict subsurface velocity maps from seismic data. It is treated as an image-to-image translation problem, jointly training an encoder for seismic data and a decoder for the velocity map from seismic-velocity pairs. In this paper, we report a surprising phenomenon: when training an encoder and decoder separately in their own domains via self-supervised learning, a linear relationship is observed across domains in the latent spaces. Moreover, this phenomenon connects multiple FWI datasets in an elegant manner: these datasets can share the self-learned encoder and decoder with different linear mappings. Based on these findings, we develop SimFWI, a new paradigm that includes two steps: (a) learning a seismic encoder and a velocity decoder separately by masked image modeling over multiple datasets; (b) learning a linear mapping per dataset. Experimental results show that SimFWI can achieve comparable results to a jointly trained model from the supervision of paired seismic data and velocity maps.