In recent years, several data-driven and data-free deep-learning approaches for solving full waveform inversion (FWI) problems have been proposed. The backbone of these approaches is the use of (deep) neural networks, which have proven to be capable of learning complex non-linear relationships between the velocity model and the shot records. Current data-driven approaches, where training is performed using acoustic wave equation simulation data or historical datasets, have only been shown to work on 1D/2D problems. Data-free methods or Physics-Informed Neural Networks (PINN), where training does not require wave equation simulation data and the physics is explicitly enforced during training, require elaborate hyper-parameter tuning and only work for a single instance of the problem. We will present a suite of solution methodologies that do not have these limitations and are practical, scalable, and robust for large-scale 3D FWI problems. Our proposed distributed deep-learning approach addresses GPU memory limitations to build deep-learning models and perform near real-time inference for large-scale 3D velocity models of resolutions up to 5123.

This content is only available via PDF.
You can access this article if you purchase or spend a download.