← Back to Sessions

Physics-informed Neural Networks for P–SV Wave Propagation and Diffraction in Wedge-shaped Media

Description: 

Wedge-shaped discontinuities provide canonical models for continental margins, mountain roots, crustal offsets, and surface topography. Predicting scattered in-plane (P–SV) wavefields in wedges remains challenging because traction-free boundaries enforce P–SV coupling and mode conversion, while the wedge tip introduces a geometric singularity that generates diffraction and highly complex post-tip wavefield patterns. In contrast to the scalar out-of-plane (SH) wedge problems, closed-form solutions for the in-plane vector wedge problem exist only for a few special cases (e.g., diffraction-free 90° wedges and 120° wedges for specific Poisson ratios). Numerical methods can extend beyond these cases, but typically demand aggressive mesh refinement near the tip, rigorous stability constraints, and carefully designed radiation/absorbing boundaries to represent an effectively unbounded domain. We use Physics-Informed Neural Networks (PINNs) to investigate the response of a wedge subjected to a vertically propagating SV wave. The proposed PINNs is trained by minimizing residuals of the elastodynamic equations, together with traction-free boundary conditions, radiation conditions, and a small set of early-time incident-wave constraints. These incident constraints are imposed through a data-misfit term built from two snapshots of the prescribed incoming field inside the wedge, constructed by superposing analytical inclined-plane-wave half-space solutions associated with the wedge faces. Despite this minimal labeled data, embedding the governing physics in the loss function enables accurate extrapolation beyond the training window. Before the incident wave reaches the tip, predicted displacements match the analytical solution to within 1% error. More importantly, after tip interaction, the framework resolves the full scattered response, capturing the post-tip wavefield structure, including mode-converted reflections from the traction-free faces and tip-diffracted waves, without requiring additional labeled data. The resulting mesh-free surrogate yields the full spatiotemporal wavefield for various wedge angles.

Session: New Frontiers in Seismic Observations and Modeling with Innovative Methods and Emerging Data on Earth and Other Planets [Poster]

Type: Poster

Date: 4/17/2026

Presentation Time: 08:00 AM (local time)

Presenting Author: Kami Mohammadi

Student Presenter: No

Invited Presentation: 

Poster Number: 117

Authors