- The paper introduces Neural-network-based Full-waveform Inversion (NNFWI), which integrates deep neural networks for velocity model reparametrization and uses automatic differentiation for gradient calculation to improve inversion.
- Experimental validation demonstrates NNFWI's superior performance on benchmark models with reduced error and improved geological feature continuity compared to conventional methods, even in noisy scenarios.
- NNFWI enables efficient uncertainty quantification through integrated Monte Carlo dropout, approximating Bayesian inference for improved predictive reliability in the high-dimensional model space of FWI.
Integration of Deep Neural Networks with Full-waveform Inversion
In the paper "Integrating Deep Neural Networks with Full-waveform Inversion: Reparametrization, Regularization, and Uncertainty Quantification," the authors outline a novel approach to tackle the challenges inherent in Full-waveform inversion (FWI), a method extensively utilized in seismic exploration to estimate subsurface velocity structures. Given the limitations of conventional FWI, notably local minima entrapment due to non-linearity and noise sensitivity, the authors propose integrating deep neural networks into the inversion process.
The conventional FWI methodology, constrained by partial differential equations (PDEs), seeks to minimize misfit between observed and predicted waveforms. However, it faces significant hurdles, such as cycle skipping and amplitude discrepancies especially in scenarios involving complex geological features. The proposed Neural-network-based Full-waveform Inversion (NNFWI) seeks to enhance the inversion process through neural network-driven reparametrization of the velocity model and automatic differentiation for gradient calculation.
Key Aspects of NNFWI
- Reparametrization: NNFWI employs generative neural networks to represent the velocity model instead of direct estimation. Convolutional layers introduce spatial correlations as implicit regularization, which assists in overcoming local minima entrapments and enhancing robustness against noise. The generative framework allows for over-parameterization, which can support imaging complex geological structures while avoiding direct pre-training.
- Experimental Validation: NNFWI's efficiency was evaluated against conventional methods using benchmark models such as the Marmousi model and the 2004 BP model. In scenarios with added noise, NNFWI displayed superior performance with reduced error and improved continuity of geological features. With no requirement for additional training datasets, NNFWI offers a straightforward alternative with minimal computational overhead compared to conventional methods.
- Uncertainty Quantification: Added dropout layers facilitate uncertainty analysis through Monte Carlo dropout, approximating Bayesian inference. This approach, integrated within NNFWI, paves the way for efficient uncertainty quantification in the high-dimensionality model space of FWI.
Implications and Future Directions
The NNFWI framework demonstrates integration of neural network characteristics and PDE constraints, thereby achieving high-resolution subsurface imaging with improved noise resilience. Its implications extend beyond mere inversion improvements, suggesting pathways towards integrating more complex deep learning architectures and optimization methods.
Moreover, its capacity for uncertainty quantification establishes a promising avenue for reducing ambiguities in velocity model estimations—an area traditionally challenging due to computational demands of conventional FWI methods. The paper suggests such quantification methods as pivotal for enhancing predictive reliability in real-world seismic exploration scenarios.
Future research might focus on leveraging advanced architectures such as Bayesian Neural Networks or encoder-decoder models, optimizing the generative process to further refine inversion results and uncertainty estimations.
Conclusion
The integration of deep neural networks into FWI illustrates a paradigm shift in seismic data inversion strategy. The NNFWI method proposed by Zhu et al. represents an advance in harnessing the synergistic potential of machine learning and established physical modeling approaches. This paper lays the groundwork for continued innovation in seismic exploration technologies, offering robust solutions to some of the key limitations faced by conventional techniques.