SVInvNet: A Densely Connected Encoder-Decoder Architecture for Seismic Velocity Inversion

TL;DR

SVInvNet introduces a densely connected encoder–decoder network for seismic velocity inversion, framing the task as learning the nonlinear mapping from seismic data to velocity models while addressing ill-posedness. The architecture uses dense blocks and cross-scale skip connections to maintain gradient flow with about $4\times10^6$ parameters, and it trains on a large, diverse synthetic dataset with both noisy and noiseless data. Across benchmarks including OpenFWI subsets and Marmousi-based models, SVInvNet consistently outperforms a modified InversionNet baseline, with performance improving as training data size increases and demonstrating robustness to coherent and stochastic noise. These results highlight the method’s potential for efficient, accurate velocity model estimation in complex geological settings, and the authors provide extensive dataset design and benchmarking strategies to support generalization assessments.

Abstract

This study presents a deep learning-based approach to seismic velocity inversion problem, focusing on both noisy and noiseless training datasets of varying sizes. Our Seismic Velocity Inversion Network (SVInvNet) introduces a novel architecture that contains a multi-connection encoder-decoder structure enhanced with dense blocks. This design is specifically tuned to effectively process time series data, which is essential for addressing the challenges of non-linear seismic velocity inversion. For training and testing, we created diverse seismic velocity models, including multi-layered, faulty, and salt dome categories. We also investigated how different kinds of ambient noise, both coherent and stochastic, and the size of the training dataset affect learning outcomes. SVInvNet is trained on datasets ranging from 750 to 6,000 samples and is tested using a large benchmark dataset of 12,000 samples. Despite its fewer parameters compared to the baseline model, SVInvNet achieves superior performance with this dataset. The performance of SVInvNet was further evaluated using the OpenFWI dataset and Marmousi-derived velocity models. The comparative analysis clearly reveals the effectiveness of the proposed model.

Figures (11)

• Figure 1: Diagram illustrating (a) forward modeling and (b) inversion processes in seismic velocity studies utilizing DL-based models. The seismic propagation equation, $\mathcal{L}$, facilitates the computation of seismic data; however, no explicit equation is available for deriving velocity based on existing seismic data.
• Figure 2: Samples from our dataset; top: layered, middle: fault, bottom: salt dome velocity model samples. The samples are selected from 4, 6, and 8 layered models. The model's spatial resolution is $700\times700$ m.
• Figure 3: The velocity model and its associated synthetic shot gathers at the first, tenth, and twentieth source locations respectively, from left to right.
• Figure 4: The samples of stochastic, coherent, noiseless and noisy gather are arranged from left to right, respectively. The gathers are plotted on the same scale.
• Figure 5: A dense block with 3 layers.