Self-Supervised Inversion of Geophysical Data via Physics-Driven Active Learning

Abstract Standard machine learning inversion of geophysical data is difficult to generalize over multiple and diverse datasets. The goal of our research is to enhance machine learning inversion to be robustly applied to field geophysical data. We develop workflows and algorithms that implement a general self-supervised machine learning inversion scheme guided by physics. Self-supervised machine learning inversion is achieved via online training where multiple learning paradigms are combined in an iterative workflow. These consist of active learning, unsupervised learning, reinforcement learning and supervised learning. Physics provides supervisory signals to guide the process. Active learning reduces the data pool while preserving the information content. Unsupervised learning by clustering serves to organize the geophysical data in representative components. Inversion is used as the reinforcement learning agent to generate models proximal to the true models. Supervised learning uses the produced labels for better training. The process converges over a few cycles to enhance and automate geophysical inversion results. The self-supervised machine learning inversion scheme is applied to two field datasets consisting of a helicopter-borne transient electromagnetics (TEM) survey in a desert environment covered by sand dunes, and of a full waveform inversion (FWI) for near surface velocity model building. The helicopter-borne TEM survey is used for mapping the topography of the sand depositional surface and enable accurate statics corrections of the sand body. The self-supervised scheme enhances and de-risk the interpretation of low-relief potential leads. The self-supervised machine learning inversion scheme is further applied for the FWI of a conventional seismic dataset acquired in complex geologic settings. The results provide boosted velocity model building for the near surface while using only 2% of the seismic gathers for training. The developed method shows robustness, accuracy and fast turnaround for multiple geophysical datasets encompassing TEM and seismic FWI. The developed method is a unique self-supervised machine learning inversion method for geophysical applications. The scheme is general such that it can be applied to a large variety of geophysical datasets. The method can be possibly extended to several other disciplines encompassing engineering and science applications where robust, fast and accurate inverse problem solutions are required.