sweep-nn

Neural reparameterizations and priors for full-waveform inversion, in plain PyTorch. Instead of updating the velocity grid directly, implicit FWI updates the weights of a network that renders the grid:

vp  = net()                       # e.g. vp_init + vp_std · SIREN(hash(z, x))
pred = solver(wavelet, sources, receivers, models=[vp])
loss = misfit(pred, obs)
loss.backward()                   # the gradient flows through vp into the weights

The network's own smoothness regularizes the inversion, and the model has far fewer free parameters than grid cells. sweep-nn has no dependency on the wave solver: any differentiable PyTorch solver works, and the examples use sweep.

Installed with pip install sweep-nn (also bundled by pip install sweepx) → import sweep_nn.

• Getting started

  ---

  Install, then fit a SIREN to a velocity model in a few lines.

• Examples

  ---

  Implicit FWI on Overthrust and Marmousi, reproducing the Geophysics paper.

• API reference

  ---

  VelocityINR, the hash grid, SIREN, the priors.

What is in it

Module	What it gives you
sweep_nn.velocity_inr	VelocityINR: a hash-encoded SIREN rendering vp_init + vp_std · net(coords), 2-D or 3-D. The recommended reparameterization for FWI.
sweep_nn.hash_encoding	MultiResHashGrid, the Instant-NGP multiresolution hash grid (Müller et al. 2022), pure PyTorch with an optional Triton kernel
sweep_nn.siren	SIREN, SirenMLP, SineLayer (Sitzmann et al. 2020)
sweep_nn.multi_param_inr	MultiParamINR: one network for several models at once (vp, vs, rho, …)
sweep_nn.priors	TVPrior, SeabedFreezeMask, learned-prior wrappers
sweep_nn.diffusion	A DDPM/DDIM velocity prior: UNet2D/UNet3D, GaussianDiffusion, and DiffusionVelocityPrior, which turns a trained checkpoint into a plug-and-play (RED) regularizer
sweep_nn.wavelet	SirenWavelet: a 1-D SIREN for a source wavelet
sweep_nn.dip	DIPReparam: a deep image prior, a small U-Net fed by a fixed latent