Physics-informed neural network for elastic wave-mode separation

Area of Science:

• Geophysics
• Computational Seismology
• Machine Learning in Geosciences

Background:

• Mode conversion in nonhomogeneous elastic media complicates accurate interpretation of physical properties.
• Correct decomposition of seismic wave modes (P and S) is essential for various scientific applications.
• Existing machine learning methods often rely on Helmholtz decomposition for mode separation.

Purpose of the Study:

• To investigate the efficacy of a physics-informed neural network (PINN) for separating P and S wave modes.
• To evaluate a novel scalar Poisson equation formulation for computational efficiency and scalability.
• To demonstrate the method's performance in both homogeneous and nonhomogeneous elastic media.

Main Methods:

• Implementation of a physics-informed neural network (PINN) to solve a scalar Poisson equation.
• Application of the scalar formulation for dimensionally scalable reduction in computational cost compared to vector formulations.
• Verification of the method using homogeneous and realistic nonhomogeneous elastic models.

Main Results:

• The PINN successfully separated P and S wave modes with high accuracy.
• The separated modes closely align with results from conventional numerical techniques.
• The proposed scalar formulation demonstrated reduced transverse wave leakage.

Conclusions:

• PINNs offer a computationally efficient and accurate approach for seismic wave mode separation.
• The scalar Poisson equation formulation provides a scalable alternative to traditional vector methods.
• This technique enhances the interpretation of physical properties in nonhomogeneous elastic media.