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

Overview

• This paper investigates how using large, diverse datasets can improve the performance of deep learning models for solving the full waveform inversion (FWI) problem.
• FWI is a technique used in geophysics to generate detailed 3D models of the Earth's subsurface structure from seismic data.
• The researchers trained and evaluated deep learning FWI models on a large, synthetic dataset called OpenFWI, and compared the results to models trained on smaller subsets of the data.

Plain English Explanation

Deep learning models have shown great potential for solving complex problems in fields like geophysics. One such problem is full waveform inversion (FWI), which aims to create detailed 3D maps of the Earth's subsurface structure using seismic data.

In this study, the researchers wanted to see if providing deep learning models with larger, more diverse datasets could further boost their performance on FWI tasks. They trained and evaluated FWI models on a new, large-scale synthetic dataset called OpenFWI, and compared the results to models trained on smaller subsets of the data.

The key finding was that training on the full OpenFWI dataset led to significant improvements in the models' accuracy, as measured by metrics like mean absolute error (MAE), mean squared error (MSE), and structural similarity index (SSIM). On average, the models trained on the combined dataset showed 13-28% better performance compared to those trained on individual subsets.

The researchers also found that the model's capacity (size and complexity) needed to scale with the dataset size to achieve optimal improvements. Larger, more complex models were able to extract more value from the additional data.

Overall, this study provides empirical evidence that using large, high-quality datasets can meaningfully enhance the performance of deep learning models - even for specialized geophysical problems like FWI. This has important implications for practitioners looking to apply deep learning to challenging real-world tasks.

Technical Explanation

The researchers conducted an empirical study to investigate the impact of using a large, diverse dataset on the performance of deep learning models for the full waveform inversion (FWI) problem.

They trained and evaluated multiple FWI models on a combination of 10 2D subsets from the recently published OpenFWI dataset, which contains a total of 470,000 pairs of seismic data and velocity maps. This allowed them to compare model performance when trained on the full dataset versus individual subsets.

The experiments demonstrated that training on the combined OpenFWI dataset yielded significant improvements compared to the individual subsets. On average, the models showed a 13.03% improvement in MAE, a 7.19% improvement in MSE, and a 1.87% improvement in SSIM. In a leave-one-out generalization test, the average improvements were even larger at 28.60%, 21.55%, and 8.22%, respectively.

The researchers also found that scaling the model capacity (size and complexity) in proportion to the dataset size was important for achieving optimal performance gains. Their largest model outperformed the smallest model by an average of 20.06% in MAE, 13.39% in MSE, and 0.72% in SSIM.

Critical Analysis

The paper provides a thorough empirical validation of the impact of large, diverse datasets on deep learning models for the FWI problem. However, a few potential limitations and areas for further research are worth noting:

1. Synthetic Data Limitations: The OpenFWI dataset used in this study is entirely synthetic, generated using numerical simulations. While this allowed for the creation of a large, diverse dataset, it may not fully capture the complexity and variability of real-world seismic data. Further research is needed to validate these findings on real-world seismic data.

2. Generalization to Other Domains: The paper focuses solely on the FWI problem in geophysics. It would be valuable to investigate whether these findings around dataset size and model capacity scaling generalize to other domains where deep learning is applied, such as underwater image enhancement or reverse gradient matching.

3. Computational and Training Considerations: The paper does not explore the computational and training time requirements for the larger models used with the full OpenFWI dataset. In real-world applications, these factors may be important considerations alongside model performance.

Overall, this study provides compelling evidence that leveraging large, diverse datasets can significantly boost the performance of deep learning models, even for specialized geophysical problems like FWI. The insights on the importance of model capacity scaling are particularly noteworthy and warrant further investigation.

Conclusion

This paper demonstrates the substantial benefits that large, diverse datasets can provide for training deep learning models to solve the full waveform inversion (FWI) problem in geophysics. By evaluating models on the comprehensive OpenFWI dataset, the researchers were able to show average improvements of 13-28% across key performance metrics compared to models trained on smaller subsets of the data.

The findings highlight the importance of scaling model capacity (size and complexity) in proportion to dataset size to extract the maximum value from large, high-quality datasets. These insights have important implications for practitioners looking to apply deep learning to challenging real-world problems, where access to diverse, high-quality data can be a key driver of model performance.

While this study was focused on the FWI domain, the principles around dataset size and model capacity scaling likely extend to other fields as well. Further research is needed to validate these findings on real-world seismic data and across a broader range of applications.