Self-Supervised, Active Learning Seismic Full Waveform Inversion

We develop a recursive, self-supervised machine learning inversion for fast and accurate full waveform inversion of land seismic data. Machine learning generalization is enhanced by using virtual super gathers of field data for training. These are obtained from midpoint-offset sorting and stacking after applying surface-consistent corrections from the decomposition of the transmitted wavefield. The procedure implements reinforcement learning concepts by adopting an inversion agent to interact with the environment and explore the model space under a data misfit optimization policy. The generated parameter distributions and related forward responses are used as new training samples for supervised learning. The active learning paradigm is further embedded in the procedure where queries on data diversity and uncertainty are used to generate fully informative reduced sets for training. The procedure is recursive. At each cycle, the physics-based inversion is coupled to the machine learning predictions via penalty terms that promote a long-term data misfit reduction. The resulting self-supervised, active learning, physics-driven deep learning inversion generalizes well with field data. The method is applied to perform full waveform inversion of a complex land seismic dataset characterized by transcurrent faulting and related structures. High signal-to-noise virtual super gathers are inverted with a 1.5D Laplace-Fourier full waveform inversion scheme. The active learning inversion procedure utilizes a small fraction of data for training while achieving sharper velocity reconstructions and a lower data misfit when compared to previous results. Active learning full waveform inversion is highly generalizable and effective for land seismic velocity model building and for other inversion scenarios.