An empirical study of large-scale data-driven full waveform inversion

Area of Science:

• Geophysics
• Machine Learning
• Data Science

Background:

• Deep learning models show promise for solving complex geophysical problems like full waveform inversion (FWI).
• The impact of large-scale, diverse datasets on deep learning for FWI remains under-explored.
• OPENFWI provides a valuable resource for investigating big data's role in FWI.

Purpose of the Study:

• To empirically evaluate the effect of big data on deep learning models applied to the FWI problem.
• To quantify performance improvements in FWI using large, multi-structural synthetic datasets.
• To determine the relationship between model capacity and dataset size for optimal FWI performance.

Main Methods:

• Deep learning models were trained and evaluated on 10 2D subsets of the OPENFWI dataset, totaling 470,000 seismic data and velocity map pairs.
• Performance was assessed using Mean Absolute Error (MAE), Mean Squared Error (MSE), and Structural Similarity Index (SSIM).
• Experiments included comparisons between training on combined datasets versus individual subsets and leave-one-out generalization tests, alongside varying model capacities.

Main Results:

• Training on the combined OPENFWI dataset improved MAE by 13.03%, MSE by 7.19%, and SSIM by 1.87% compared to split datasets.
• Leave-one-out generalization tests showed average improvements of 28.60% in MAE, 21.55% in MSE, and 8.22% in SSIM.
• Increasing model capacity alongside data size yielded significant performance gains, with the largest model outperforming the smallest by 20.06% (MAE), 13.39% (MSE), and 0.72% (SSIM).

Conclusions:

• Big data, particularly large and diverse synthetic datasets like OPENFWI, demonstrably enhances deep learning model performance for full waveform inversion.
• Optimal performance in data-driven FWI requires a synergistic scaling of model capacity with the size and complexity of the training dataset.
• This study validates the effectiveness of big data in improving FWI accuracy and generalization, paving the way for more robust geophysical subsurface imaging.