Deep learning for well-to-seismic data mapping and reconstruction

Abstract

Seismic inversion, a critical geophysical technique for quantitative subsurface characterization, faces persistent challenges including the scarcity of labeled well-log data, the non-stationarity of seismic signals, the ill-posedness of multi-parameter estimation from limited-offset data, and practical difficulties with real-field data imperfections. This thesis introduces innovative deep semi-supervised learning (DSSL) frameworks to enhance inversion accuracy, robustness, and data efficiency. The methodology employs 1-D U-Net architectures in DSSL paradigms, learning from limited labeled and abundant unlabeled seismic data. It integrates physical principles via neural network emulators of complex wave phenomena and explicit non-stationary forward models. Specialized training, including two-phase/stage learning, on-the-fly synthetic data generation, consistency regularization, Total Variation loss, and adaptive masking, are employed to optimize performance and generalization. This research demonstrates DSSL’s efficacy across key inversion problems: first, for reflectivity inversion, effectively handling sparsity and non-stationary wavelets with limited data; second, for simultaneous acoustic impedance and density estimation from zero-offset data using a physics-based reconstructor, achieving accuracy even with extreme data scarcity; third, by explicitly modeling wavelet non-stationarity for improved accuracy and spectral fidelity; and finally, by robustly tackling real-field geometric well-seismic misalignments and incomplete logs through novel alignment and masking. The proposed DSSL frameworks consistently surpass traditional and supervised learning benchmarks, offering superior accuracy, enhanced geological plausibility, and notable computational efficiency. This thesis delivers novel, practical, data-efficient solutions to enduring seismic inversion challenges by intelligently integrating deep learning with geophysical domain knowledge. The developed techniques hold significant potential for improving subsurface exploration, characterization, and management in diverse energy and environmental applications.