GeoML: Integrated geophysical imaging driven by machine learning

Project: Research

Project Details

Description

Minimally invasive geophysical methods allow an effective way to understand subsurface structures and their physical properties. The inherent ambiguity associated with any single geophysical technique can be alleviated by considering one or more of the different data types. By using complementary information from different data modalities, integrated imaging can help improve the characterization of subsurface structures and establish rock physics models for quantitative analysis. Numerous case studies have established the case for joint inversion of different geophysical data types. For example, integrating gravity and magnetic fields with other geophysical data has been shown to add value throughout the exploration life cycle. Despite its usefulness, integrated imaging largely remains an undervalued tool in the geophysical community. This is mainly caused by the challenge of handling the mixed resolution of different data types that exhibit sensitivity to different geophysical properties for varying geologic conditions. While different approaches have been proposed to address the challenge, they are typically hand-crafted and leave ample room for personal judgment. Therefore, to enable robust and accurate integrated imaging, we propose a joint inversion algorithm based on the paradigm of physics-informed neural networks (PINNs). In addition to an increased level of automation, the proposed approach offers a natural way of addressing the underlying challenges hampering the widespread acceptance of integrated imaging in the geophysical community. To overcome the non-uniqueness of the inverse problem, conventional techniques typically use smoothing or damping regularizers that result in limiting the resolution of the inverted models. On the contrary, we propose to use a physics-informed regularizer that honors the wave propagation physics by ensuring the obtained geophysical models and quantities satisfy the governing partial differential equations for each of the different data types. Moreover, it is well-known that the choice of initial models can adversely affect the performance of conventional techniques; therefore, an added advantage here is that the performance of the proposed formulation is largely independent of the initial model distribution. Furthermore, since PINN is a mesh-free method, there is no limitation in modeling irregular topography or undulating geological interfaces a flexibility not afforded by most joint inversion algorithms due to the use of a regular mesh. It is worth noting that currently the PINN framework suffers from certain limitations that hamper its capabilities for high-resolution imaging. These challenges include its learning capability for high frequency features in the solution and instability in convergence due to multiscale interactions between different terms in the neural networks loss functions. These limitations will also be addressed during the project to ensure that the method yields a high-fidelity earth model. The outlined challenges with PINNs will be addressed through Fourier feature mapping and an adaptive training strategy. These advances will allow the geophysical community to make full use of the different data types for accurate characterization and monitoring of complex subsurface environments and allow operators to reduce prospecting and drilling risks. Moreover, an integrated imaging framework will also be crucial in the discovery and exploitation of other forms of subsurface energy, such as geothermal, and mineral resources.
StatusFinished
Effective start/end date1/11/211/11/22

Fingerprint

Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.