Machine learning inversions incorporating geologic information through variational autoencoder and physics-based neural network

Geophysical inversions require the inclusion of prior geological information. Many types of prior information are implicit or conceptual in nature. For instance, the information about the features in the subsurface geology is embedded in the multiple geophysical data sets, but not directly accessible; or we may know the types of structural elements present in the subsurface but not their specifics. A long-standing challenge is in how to integrate this type of prior information. The ability of machine learning (ML) in capturing such conceptual and implicit information provides a new avenue to address this challenge. In this presentation, we discuss two approaches which we have developed. The first is a physics-based neural network that captures the implicit geological information embedded in the multiple data sets to be inverted. The second is a conditional variational autoencoder that captures the prior geological information embedded in the training models with a set of known features and associated geophysical data. Both approaches highlight the unique advantage of ML in capturing general geology information.