Physics-based and machine learning inversions for geophysical characterization of subsurface

Abstract

Recovering the subsurface physical properties for near-surface characterization and reservoir monitoring using geophysical inversion remains challenging. For near-surface characterization, we need to integrate the interpretation of different geophysical data to reduce the uncertainty and improve the robustness of the recovered earth model. Therefore, in the presence of many geophysical data sets, we need to develop an inversion framework for joint interpretation to generate a reliable earth model that can explain all geophysical data and prior information. For reservoir monitoring, we need to use time-lapse data in conjunction with a fluid flow simulator to monitor complex reservoir dynamics and image the fluid migration. This entails a number of challenging aspects, such as sensitivity to parameter changes, resolution of the produced data, and the expense of long-term deployment. To resolve those challenges, I investigated three classes of inversion techniques to invert geophysical data, including (i) physics-based deterministic inversion, (ii) machine learning inversion, and (iii) a physics-based machine learning inversion.

The physics-based deterministic inversion is an efficient local optimization technique utilizing the physics embedded in the forward problem. I used the physics-based inversion to develop a joint inversion framework to integrate data from different geophysical methods, particularly seismic full-waveform data and MT data, to generate a P-wave velocity model using multiple coupling constraints. Combining the interpretation of both types of data collected over the same area using coupling strategies is one way forward to reduce non-uniqueness in inversion results and obtain more reliable models for subsurface characterization. In this work, I solved the joint inversion problem by minimizing an objective function with three coupling constraints; structural, rock physics, and clustering constraints.

The Machine Learning (ML) inversion applied to geophysical problems reflects a new field of ongoing research. I used a machine learning method, namely the feed-forward neural network, to invert time-lapse three-axis borehole gravity data to predict carbon dioxide plume locations. The neural network is trained on the input data, including changes in reservoir density and the corresponding gravity data. The network inversion using three-axis time-lapse borehole gravity applications have shown the feasibility of efficiently monitoring changes in density for the Johansen formation offshore Norway.

The physics-based machine learning inversion is a new research field that aims to integrate the advantages of both physics-based deterministic inversion and machine learning inversion in a complementary manner. I developed an inversion-based neural network procedure that combines the benefit of deterministic and neural network inversions in a coupled inversion scheme. The new inversion algorithm is formulated as a constrained problem solved by minimizing an objective function composed of data misfit, neural network misfit, and a coupling model objective function that links the two inversion schemes. Our synthetic 2D magnetotelluric data demonstrations show that the recovered models from the physics-based neural network inversion are superior.