Self-supervised active learning inversion: Transient electromagnetics and full-waveform seismic

Supervised learning requires training samples extracted from a probability distribution compatible with the data to be tested in the inference step. We develop a self-supervised machine learning inversion scheme able to learn directly from field data by using physics as the auxiliary task for generating training labels. The self-supervised online learning process is achieved through the combination of multiple learning paradigms. Active learning selects the most informative samples of a data pool. Unsupervised learning quantifies diversity in the data. Reinforcement learning explores the parameter space through a physics-based agent and generates labels. The labeled data pool is then used for supervised learning. The iterative process involves data augmentation at each cycle and results in a gradual steering of the learning toward the field data distribution. Several benefits are provided to the inversion of geophysical data by introducing nonlinearity and stochastic processes while reducing, at the same time, the computer resource utilization. Offline training involving expensive forward modeling operations is eliminated where field data become the source of online learning. Self-supervised active learning inversion is of immense impact for field data applications. Exploration-scale implementations include the inversion of helicopter-borne transient electromagnetics and land seismic full-waveform inversion in complex geologic settings. The developed self-supervised active learning inversion scheme is likely generalized to a tutorial for a variety of scientific applications.