← Back to Sessions

Seismic Geotechnical Imaging Using Full-waveform Inversion and Physics-informed Neural Networks

Description: 

Large uncertainties in natural geomaterials constrain the practical application of theoretical predictive models in geotechnical engineering. Although geophysical techniques have been employed to reduce the epistemic part of these uncertainties, current methods often rely on sparse surface or borehole measurements and selectively utilize limited data such as first-arrival times. Full-waveform inversion (FWI) techniques provide a comprehensive approach by leveraging entire seismic records; however, their high computational cost and complex formulation have limited their widespread adoption. This study presents a novel approach using Physics-Informed Neural Networks (PINNs) to develop a seismic inversion framework for geotechnical applications that is both robust and computationally efficient. By integrating the underlying physics into the loss function of the optimization process, PINNs enable more accurate subsurface characterization with fewer data points and improve the prediction of responses beyond the range of the training dataset. PINNs can also be used to solve seismic inversion problems by defining unknown P- and S-wave velocities as trainable parameters. To demonstrate the efficacy of our approach, we focus on the 1-D problem involving site response analysis and geotechnical subsurface characterization. Geometric and material parameters are consolidated into normalized parameters, such as dimensionless frequency and normalized thickness, to enhance the generality of the results. We generate synthetic training datasets using a Finite Volume forward solver for the Navier-Cauchy equation and apply our FWI-PINNs approach to retrieve unknown material properties and wavefield components. Given the rapid advancements in GPU-based machine learning algorithms, coupled with their increasing accessibility and ease of use, we believe our proposed method can evolve into a fast, robust, and practical site characterization tool for the geotechnical engineering community.

Session: Scientific Machine Learning for Forward and Inverse Wave Equation Problems - I

Type: Oral

Date: 4/15/2025

Presentation Time: 05:30 PM (local time)

Presenting Author: Kami

Student Presenter: No

Invited Presentation: 

Poster Number:

Authors