Active learning full-waveform inversion

We developed a self-supervised machine learning method for performing seismic full waveform inversion and we apply it to a complex land dataset. The system involves the utilization of multiple machine learning paradigms in a workflow to enable the machine learning algorithm to learn directly from representations of field data to ensure optimal generalization. The scheme further utilizes the active learning paradigm based on query heuristics on data diversity and uncertainty to select reduced sets of the data pool for training. The self-supervised, active learning full waveform inversion is successfully implemented on synthetic and field data where it enhances the parameter reconstruction resolution, reduce the overall data misfit while enabling a broad exploration of the solution space.