Data-Driven Inversion for Ultrasonic Wave Propagation: From Full-Waveform Inversion to Gaussian Process Emulation

Abstract: Full-waveform inversion (FWI) is a powerful technique for reconstructing subsurface properties from wavefield data, typically employing adjoint-state methods to compute sensitivity kernels and optimize model parameters. However, in highly attenuative media, strong energy dissipation complicates backpropagation, posing challenges for gradient-based optimization methods. In this talk, I will discuss two alternative approaches to wavefield-based inversion, using ultrasonic laboratory measurements as an example application. First, I will outline a standard FWI approach utilizing conjugate-gradient optimization to iteratively refine P-wave velocity and S-wave velocity. Second, I will present a data-driven alternative: a Gaussian process emulator that approximates the misfit function over the parameter space, enabling efficient inversion without the need for explicit backpropagation. This approach offers advantages in scenarios where traditional FWI struggles due to high attenuation or limited data coverage. However, so far, it has only been tested for determining P- and S-wave velocities as well as attenuation parameters in homogeneous rock samples. I will discuss potential extensions and applications beyond ultrasonic experiments, highlighting broader implications for geophysical inversion problems.

numerical analysisoptimization and control

Audience: researchers in the topic

CAM seminar

Series comments: Online streaming via zoom on exceptional cases if requested. Please contact the organizers at the latest Monday 11:45.