Encoder-Decoder Architecture for 3D Seismic Inversion.

Area of Science:

• Geophysics
• Seismic Data Analysis
• Machine Learning in Earth Sciences

Background:

• Seismic data inversion for 3D geological structures is computationally intensive and data-heavy.
• Industry-standard Full Waveform Inversion (FWI) requires significant resources for processing large seismic datasets.
• Reconstructing accurate 3D models from noisy field data presents a major challenge in geophysics.

Purpose of the Study:

• To develop an efficient deep learning solution for 3D geological model reconstruction from seismic data.
• To address the computational burden and data volume issues associated with traditional seismic inversion methods.
• To evaluate the performance of a deep learning model in reconstructing realistic 3D geological structures in the presence of field noise.

Main Methods:

• Implementation and analysis of a convolutional encoder-decoder deep learning architecture.
• Processing of hundreds of seismic shot-gather cubes for 3D model reconstruction.
• Testing the model's robustness against field noise at a 10 dB signal-to-noise ratio.

Main Results:

• The deep learning model successfully reconstructed realistic 3D geological models.
• Achieved a Structural Similarity Index Measure (SSIM) of 0.9143, indicating high fidelity.
• Demonstrated efficient processing of large seismic datasets, overcoming computational limitations.

Conclusions:

• Deep learning offers a viable and efficient alternative to traditional methods for seismic data inversion.
• The proposed convolutional encoder-decoder architecture effectively handles large datasets and field noise.
• Accurate 3D geological models can be reconstructed using AI, even under challenging field conditions.