Tsunami inversion using deep neural representations.

Area of Science:

• Oceanography
• Geophysics
• Computational Science

Background:

• Accurate tsunami forecasting is crucial for rapid response to ocean disturbances.
• Traditional seismic source assumptions limit accuracy for non-seismic events.
• Existing sensor-based methods often rely on fixed offshore sensor sets and simulations.

Purpose of the Study:

• To develop a new approach for tsunami forecasting using offshore sensor data.
• To improve the accuracy and robustness of tsunami prediction, especially for non-seismic events.
• To address computational challenges in modeling Green's functions for tsunami forecasting.

Main Methods:

• Modeling Green's functions to represent impulse responses from ocean disturbances.
• Utilizing neural networks for compressed representation of numerous Green's functions.
• Employing iterative constrained optimization for inversion of sensor data.
• Developing a method robust to changes in offshore sensor configurations.

Main Results:

• The approach successfully models tsunami forecasting from offshore sensor data.
• Neural networks effectively reduce storage requirements for Green's functions.
• The inversion method is robust to sensor network variations and accommodates non-seismic events.
• Simulations of historical and hypothetical tsunamis near Japan validate the method's effectiveness.

Conclusions:

• This Green's function-based inversion with neural networks offers a promising advancement in tsunami forecasting.
• The method enhances accuracy by constraining initial conditions and reducing forecasting uncertainty.
• It provides a flexible and robust alternative to traditional tsunami prediction techniques, particularly for non-seismic events.