Advanced seismic inversion using physics-constrained data augmentation and deep learning architectures: A case study in clastic sequences, southwest Iran

Accurate ۳D acoustic impedance (AI) estimation remains a significant challenge in subsurface modeling. This challenge becomes critical in exploration and development scenarios, especially in areas characterized by sparse wells control.This study introduces a novel hybrid methodology designed to address these limitations. Our approach synergistically integrates physics-based pseudo-well generation with a deep feedforward neural network (DFNN), enabling reliable impedance modeling even with restricted subsurface information. The comprehensive workflow meticulously leverages calibrated rock physics principles and advanced AI geostatistical techniques to synthesize realistic elastic log data that faithfully preserves critical lithological properties. We utilize a carefully selected suite of seismic attributes—encompassing amplitude, phase, and frequency characteristics—as primary inputs for the DFNN. The network is then rigorously trained and validated using a leave-one-well-out cross-validation protocol, ensuring the model's generalizability. Key stages of this methodology are elucidated through detailed flowcharts, illustrative rock physics cross-plots, and compelling ۳D visualizations of the resultant AI model. The core contribution of this study is a stable workflow that emphasizes methodological rigor, model design, and validation, highlighting its practical utility and scalability for modern reservoir characterization.

کلیدواژه ها:

Seismic inversion ، acoustic impedance ، deep learning ، rock physics modeling ، data augmentation