LARGE SCALE RANDOMIZED LEARNING GUIDED BY PHYSICAL LAWS WITH APPLICATIONS IN FULL WAVEFORM INVERSION.

Area of Science:

• Geophysics
• Computational Seismology
• Inverse Problems

Background:

• Full waveform inversion (FWI) is crucial for subsurface imaging but computationally intensive.
• Optimizing media velocity models in large-scale non-linear problems requires efficient learning strategies.
• Existing methods often face challenges with convergence rate and computational cost.

Purpose of the Study:

• To develop a novel method for learning velocity models in FWI.
• To improve the convergence rate and reduce computational cost of FWI.
• To enhance the accuracy of subsurface imaging through velocity model optimization.

Main Methods:

• The proposed Sub-Sampled Newton (SSN) method combines randomized subsampling with a second-order optimization algorithm.
• SSN approximates the Hessian matrix using a non-uniform subsampling scheme.
• The method incorporates curvature information for efficient optimization.

Main Results:

• SSN demonstrates a faster convergence rate compared to commonly used methods.
• The method achieves a more accurate velocity model, indicated by lower mean squared error.
• SSN preserves a comparable convergence rate to Newton's method with significantly reduced iteration costs.

Conclusions:

• The Sub-Sampled Newton method offers an efficient and accurate approach for velocity model learning in FWI.
• SSN provides a viable solution for complex physical system learning problems.
• This method advances the field of geophysical imaging and subsurface characterization.