Neural network joint implicit inversion for seismic and gravity data

Tianze Zhang, Kristopher A. Innanen

In this study, we introduce a neural network framework for multi-physics joint inversion of geophysical data. Typically, multi-physics models, such as velocity, conductivity, and permeability, exist in separate domains and vary significantly in scale, making it challenging to represent them as a unified entity. To address this issue, we employ neural networks to generate features for these geophysical models. These features are unitless, typically ranging from -1 to 1. We then use prior information to scale these features, mapping them to their respective physical domains, enabling forward modeling. The forward modeling for each type of geophysical model produces corresponding synthetic data, allowing for the evaluation of loss values specific to each data type. The total loss for the joint inversion is calculated by summing these individual losses, and the neural network is updated to minimize this total loss. By using this inversion strategy, we use the neural network to represent the geophysical model of a particular area as an entity, and the joint inversion of the multi-physics model can be evaluated. In this paper, we examine the joint inversion of gravity and seismic data. The results indicate that incorporating gravity data improves the convergence rate of the density model inversion compared to using seismic data alone.