Abstract:This paper investigates the impact of big data on deep learning models to help solve the full waveform inversion (FWI) problem. While it is well known that big data can boost the performance of deep learning models in many tasks, its effectiveness has not been validated for FWI. To address this gap, we present an empirical study that investigates how deep learning models in FWI behave when trained on OpenFWI, a collection of large-scale, multi-structural, synthetic datasets published recently. In particular, we train and evaluate the FWI models on a combination of 10 2D subsets in OpenFWI that contain 470K pairs of seismic data and velocity maps in total. Our experiments demonstrate that training on the combined dataset yields an average improvement of 13.03% in MAE, 7.19% in MSE and 1.87% in SSIM compared to each split dataset, and an average improvement of 28.60%, 21.55% and 8.22% in the leave-one-out generalization test. We further demonstrate that model capacity needs to scale in accordance with data size for optimal improvement, where our largest model yields an average improvement of 20.06%, 13.39% and 0.72% compared to the smallest one.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the performance improvement of Full Waveform Inversion (FWI) driven by big data. Specifically, the authors explored the impact of large - scale datasets on deep - learning models in FWI tasks, as well as the performance and generalization ability of these models under different data scales. The following are the main research directions of the paper: 1. **Model performance**: By using large - scale datasets to train deep - learning models, evaluate whether the performance of these models in FWI tasks has been improved. The study found that, compared with models trained on small - scale datasets, models trained on large - scale datasets showed significant performance improvements on multiple evaluation metrics (such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Structural Similarity Index (SSIM)). 2. **Relationship between model size and data scale**: Studied how to adjust the capacity (i.e., model size) of the model to obtain the best performance as the amount of training data increases. The experimental results showed that when more training samples are introduced, a larger model architecture is required to achieve further performance improvements. 3. **Model generalization ability**: Explored the generalization ability of models trained on large - scale datasets on unseen data. The experimental results showed that models trained on large - scale datasets showed better performance in generalization tests, especially on unseen datasets. To verify the above problems, the authors used a large - scale synthetic dataset named OPENFWI, which contains multiple geological structures, such as interfaces, geological faults and field data. By training and evaluating different deep - learning models on these datasets, the authors reached the above conclusions. ### Main findings: - **Big data improves performance**: Models trained on large - scale datasets (BigFWI) showed better performance than the baseline model (InversionNet) on almost all datasets. - **Larger data requires larger models**: As the amount of training data increases, a larger model architecture is required to achieve further performance improvements. - **Big data improves generalization ability**: Models trained on large - scale datasets showed better generalization ability on unseen data. ### Method overview: - **Dataset**: Use the OPENFWI dataset, which contains 10 2D synthetic data subsets, with a total of 470,000 pairs of seismic data and velocity maps. - **Model architecture**: Proposed three BigFWI models of different sizes (Base, Middle, Large), which share an encoder - decoder architecture. - **Loss function**: Combined pixel - level ℓ1 loss and ℓ2 loss to take full advantage of both. ### Experimental results: - **Performance improvement**: Models trained on large - scale datasets showed significant performance improvements on metrics such as MAE, RMSE and SSIM. - **Generalization ability**: Models trained on large - scale datasets showed better generalization ability on unseen data, especially on complex geological structures. Through these studies, the authors provided empirical support for the use of large - scale datasets in FWI tasks and provided an important reference for further research.

An Empirical Study of Large-Scale Data-Driven Full Waveform Inversion

Multiscale Data-driven Seismic Full-waveform Inversion with Field Data Study

OpenFWI: Large-Scale Multi-Structural Benchmark Datasets for Seismic Full Waveform Inversion

Deep-learning seismic full-waveform inversion for realistic structural models

On the Robustness and Generalization of Deep Learning Driven Full Waveform Inversion

Physics-Consistent Data-driven Waveform Inversion with Adaptive Data Augmentation

Inversion-DeepONet: A Novel DeepONet-Based Network with Encoder-Decoder for Full Waveform Inversion

Elastic Full Waveform Inversion with Angle Decomposition and Wavefield Decoupling.

Learning with real data without real labels: A strategy for extrapolated full-waveform inversion with field data

Full-waveform inversion with adversarial losses via deep learning

$\mathbf{\mathbb{E}^{FWI}}$: Multi-parameter Benchmark Datasets for Elastic Full Waveform Inversion of Geophysical Properties

Self-Supervised Deep Learning for Nonlinear Seismic Full Waveform Inversion

Fourier-DeepONet: Fourier-enhanced deep operator networks for full waveform inversion with improved accuracy, generalizability, and robustness

WaveDiffusion: Exploring Full Waveform Inversion via Joint Diffusion in the Latent Space

Integrating Physics of the Problem into Data-Driven Methods to Enhance Elastic Full-Waveform Inversion with Uncertainty Quantification

Deep-learning assisted regularized elastic full waveform inversion using the velocity distribution information from wells

Implicit seismic full waveform inversion with deep neural representation

Reparameterized full-waveform inversion using deep neural networks

An Augmented Lagrangian Method-Based Deep Iterative Unrolling Network for Seismic Full-Waveform Inversion

Learning to Invert Pseudo-Spectral Data for Seismic Waveform Inversion

Enhancing Seismic Waveform Inversion Using a Three-Step Strategy With Adversarial Neural Networks and Seismic Envelope