A new way to generate wavefields in the frequency domain

A conditional mean flows model produces scattered wavefields more efficiently than diffusion models.

A conditional mean flows model was introduced for generating complex scattered-wavefields. This model works in the frequency domain.

The model uses an analytical background wavefield as a condition. This condition works alongside the velocity model to provide information about source location and frequency.

The model estimates the average transport velocity over a finite interval. This interval connects Gaussian noise to the scattered field.

The model was trained on OpenFWI. It was evaluated on frequency generalization, inference efficiency, and velocity structures that are in-distribution or out-of-distribution.

One-step mean flows sampling achieved lower wavefield error than multi-step diffusion in most settings. This one-step generation also reduces inference time.

Results indicate that mean flows can provide accurate and efficient frequency-domain scattered-wavefield generation.

Key facts

  • The model estimates average transport velocity over a finite interval.
  • One-step mean flows sampling can achieve lower wavefield error than multi-step diffusion.
  • The model was trained on OpenFWI.

All sections

Edition