| Abstrak/Abstract |
Seismic inversion is used to evaluate hydrocarbon reservoirs by inferring subsurface physical properties from seismic data. The most prevalent seismic inversion method is post-stack seismic inversion, as most seismic data is available in post-stack format. However, this method only transforms impedance parameters from seismic data. Therefore, to extract additional parameters such as density and P wave velocity/Vp from post-stack seismic data, we developed Cycle Loop – Semi-Supervised Learning for multi-parameter post-stack seismic inversion (Cycle-MPInv). This method combines a deep learning-based inversion network and forward modeling to perform seismic inversion and forward modeling imultaneously (semi-supervised learning). The forward model provides geophysical constraint, while deep learning employs convolutional neural networks (CNNs) and bidirectional gated recurrent unit (Bi-GRU) to extract both high- and low-frequency features. In Cycle-MPInv, we also process labeled parameter data and calculate the loss to enhance the learning process and improve accuracy, so there are two loops that we call the cycle loop process. The proposed method was tested on synthetic data (Marmoussi II model) and data from the Netherlands offshore F3 block. Results from both datasets demonstrate that Cycle-MPInv effectively leverages both low- and high-frequency features of multi-parameter properties (density, Vp, and impedance) with limited labeled data. Furthermore, Cycle-MPInv achieves superior accuracy and robustness compared to other deep learning methods, even when handling noisy seismic data. These findings suggest that this method can offer valuable additional parameter insights for hydrocarbon reservoir evaluation using post-stack seismic data. |
