Title: An automated approach to incorporate structural information into the inversion using image segmentation
Geological structure is distinctive and boundaries are sharp contacts between units. On the other hand, geophysical models are often smooth and geological meaning interpreted. These interpretations are often subjective and the model can be unconstrained and not always influenced by prior information. Presenting these results to a non-geophysicist can be challenging. Typically, delineations are sketched or overlain on the geophysical model to communicate the results thoroughly. Information on structure orientation is often ambiguous. With prior information, we can select more geologically relevant models using regularization. However, it is more common that prior information is not available. This work proposes an automated approach to infer structure in geophysical models. This is done by applying image segmentation methods borrowed from computer vision. This work uses a Transformer network (Vaswani, 2017) to perform the segmentation. The segmentation then becomes the prior information that we update the regularization with as we iterate an inversion. Within the regularization, we can indicate spaces in the model and the directions to regularize. We also explore additional methods to incorporate segmentation to the inversion via proximal operators and the Alternating Method of Multipliers (ADMM). A more clearly defined targeting system for interfaces like unconformities or structural orientations of intrusive units, which are often associated with mineralization zones, is provided by incorporating segmentation methods. The results provide an outlook on using machine learning and foundational models for structural interpretation of geophysical models.