Seismic reflectivity inversion using deep learning and model-based methods

Abstract

This thesis advances seismic imaging and inversion, crucial for identifying hydrocarbon prospects and understanding the Earth's internal structure by retrieving rock parameters from seismic data. Facing challenges such as solution non-uniqueness, slow convergence, and high computational demands, this work integrates deep learning frameworks, classical regularization theory, and two-way wave equation propagators. We develop iterative and non-iterative deep learning methods, such as interlacing convolutional neural networks (CNNs) within traditional Least-Squares Reverse Time Migration (LSRTM) schemes, and employing deep autoencoders to refine inverse problem spaces, enhancing resolution and speeding up convergence. Additionally, Sparse Full Waveform Least-Squares Reverse Time Migration (Sparse FWLSRTM) is introduced, combining sparse regularization with the full wavefield vector reflectivity modeling engine to substantially improve imaging quality under complex geological settings. Furthermore, the thesis explores seismic broadband deconvolution using deep learning to derive full-band reflectivity from band-limited data, integrating learned null space components for better data consistency and resolution. Collectively, these methodologies significantly enhance the fidelity and efficiency of seismic imaging, merging advanced machine learning techniques with traditional approaches to offer a robust toolset for geophysical prospecting in challenging subsurface environments.