Skip to content

SEM

The spectral-element solver, sweep.sem.PropSEM: PropTorch's call on a Quad9 mesh that follows the free surface. Forward and backward on a two-layer model first, then topography against the finite-difference solvers, then FWI on Marmousi, one model and every shot's window in one batched call. The SEM user guide has the rest of the API.

  • Hello SEM

    Hello · SEM · forward and backward

    PropSEM on a two-layer model: one shot gather, then one .backward() for the vp gradient on the grid, with the full tape and with memory=BoundarySaving().

    Open notebook

  • FD vs SEM topography

    Wavefield · FD vs SEM on topography

    The same Gaussian hill two ways: FD on the grid (the image method acoustic, the boundary-fitted curvilinear grid elastic) and SEM on a Quad9 mesh deformed to the surface, snapshots and gathers side by side.

    Open notebook

  • FWI Acoustic Marmousi SEM

    FWI · Acoustic · Marmousi · SEM

    The Marmousi workflow on the spectral-element solver: the grid is interpolated onto the GLL nodes and the gradient comes back on the grid. impl='eager' or 'c'.

    Open notebook

  • FWI batched local windows SEM

    FWI · Batched per-shot local windows · SEM

    The SEM twin of the streamer example: every shot's window in one batched PropSEM call (impl='c', memory=BoundarySaving()), frequency continuation from a smooth start.

    Open notebook