Creating geologically informed training models for deep learning inversion: Applications to potential-field data

ABSTRACT Recent advancements in deep learning-based geophysical inversion have drawn considerable attention. Most of these inversions are supervised, which requires the creation of training models that capture as much prior geological information as is available in an area of interest. However, creating such geologically informed training models is challenging because some geological knowledge is difficult to be expressed in mathematical terms. Moreover, geological prior information is not always tied to specific spatial locations. To address these challenges, a novel method based on alpha shapes, a concept from computational geometry, was proposed to generate training models that can easily integrate five key types of geological prior information, namely, (1) top boundaries, (2) dip angles, (3) surface outcrop contacts, (4) mineralization zones intersected by drillholes, and (5) measured physical property values on rock samples. Three distinct scenarios were presented to demonstrate how the proposed method can be used to systematically generate geologically informed training models. It was also shown that deep generative models, such as the conditional variational autoencoder, trained on these geologically informed models, can not only output inversion results that align with prior geological knowledge but also quantify the associated uncertainties. To validate our approach, it was applied to a set of magnetic measurements collected in Qinghai Province, China, for the exploration of Cu-Mo critical mineral deposits. The resulting susceptibility models reveal a major dipping structure that is consistent with surface geology and the magnetic data. The proposed method offers a flexible and unified framework for generating geologically informed training models for deep learning-based geophysical inversions.