Physics-Guided Neural Network for Full Waveform Inversion With Structural Enhancement
Description:
Seismic velocity modeling plays a critical role in understanding subsurface structures. Full-waveform inversion (FWI) is a powerful technique that enables high-resolution velocity modeling by iteratively minimizing the discrepancies between observed and predicted seismic waveforms. However, the inherent strong nonlinearity of FWI can often lead to convergence in local minima, especially when fitting oscillatory waveforms. This challenge becomes even more pronounced when dealing with insufficient starting models and noisy data characterized by low signal-to-noise ratios.
Our study introduces an innovative learning-based approach to FWI, which integrates deep learning with physical equations. In this method, we parameterize the velocity model using a convolutional neural network (CNN). CNN inherently introduces spatial correlations, serving as regularization for the velocity model. This renders deep learning well-suited for FWI, effectively mitigating noise in model gradients and alleviating issues related to local minima. The velocity model generated by the CNN is seamlessly incorporated into the acoustic partial differential equation (PDE) solvers used in traditional FWI. Automatic differentiation is harnessed to compute PDE gradients, facilitating the updating of network parameters through backpropagation. Additionally, inspired by recent developments in structure-tensor coherence techniques, we further improve FWI through anisotropic diffusion smoothing. We compute oriented derivatives along directions both perpendicular and parallel to seismic reflectors, giving rise to oriented structure tensors. These oriented tensors enhance the representation of lateral velocity discontinuities, particularly those associated with features that align with dipping structures in the velocity model. The numerical experiments conducted using the well-established Marmousi model demonstrate better representations of discontinuous and stratigraphic velocity features compared to conventional FWI.
Session: Machine Learning for Full Waveform Inversion: From Hybrid to End-to-End Approaches [Poster Session]
Type: Poster
Date: 5/3/2024
Presentation Time: 08:00 AM (local time)
Presenting Author: Zhengfa
Student Presenter: No
Invited Presentation:
Authors
Zhengfa Bi Presenting Author Corresponding Author zfbi@lbl.gov Lawrence Berkeley National Laboratory |
Nori Nakata nnakata@lbl.gov Lawrence Berkeley National Laboratory |
|
|
|
|
|
|
|
Physics-Guided Neural Network for Full Waveform Inversion With Structural Enhancement
Category
Machine Learning for Full Waveform Inversion: From Hybrid to End-to-End Approaches