SWEEPX

Seismic Wave Equation Exploration Platform — a differentiable framework for seismic wave-equation modeling, migration, and full-waveform inversion. One API, 20+ equations (acoustic / elastic / VTI / TTI / DAS), PyTorch and JAX backends, eager and compiled CUDA paths.

📖 Documentation: https://sweepx.deepwave.group/solver/

Install

pip install torch --index-url https://download.pytorch.org/whl/cu126   # match your driver, see below
pip install sweepx

Install a torch built for your driver first: PyPI's default torch is CUDA 13 and needs driver >= 580, so on an older driver (or a V100) use cu126 as above. The install selector gives the exact command for your GPU and driver.

The wheel ships prebuilt CUDA cores, so the compiled backend (impl='c') works right after install with any PyTorch version: no nvcc, no compiler, no build step. It needs an NVIDIA GPU and driver; the core for your torch's CUDA version is picked automatically (CUDA 12: V100 and newer; CUDA 13: T4 and newer, driver >= 580). The eager PyTorch and JAX backends are pure Python.

From source: a clone has no prebuilt core, so the CUDA core is compiled locally once (with an nvcc matching your torch's CUDA version) and cached:

pip install .
python -m sweep.build   # optional: compile it now instead of on the first impl='c' call

sweepx (Python >= 3.10) installs the solver together with the sweepx toolkit (task runner, agent, losses, I/O, networks, optimizers); you import sweep. pip install sweep-solver installs the solver alone (Python >= 3.9). GPU coverage, custom cores and developer builds are covered in Building the CUDA core.

Hello SWEEP

One shot, one receiver, one .backward() — read off the velocity-model gradient for a single trace:

import numpy as np
import torch
from sweep.equations import Acoustic
from sweep.propagator.torch import PropTorch
from sweep.signal import ricker

shape = (96, 128)
dh, dt, nt = 10.0, 0.002, 800
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

vp_true = np.full(shape, 1500.0, dtype=np.float32)
vp_true[shape[0] // 2:, :] = 2500.0
vp_init = np.full(shape, 1500.0, dtype=np.float32)

solver = PropTorch(Acoustic(device=device), shape=shape, dh=dh, dt=dt,
                   device=device, use_ckpt=False)

t = np.arange(nt) * dt
wavelet = ricker(t - 0.14, f=10.0).astype(np.float32)
sources   = np.array([[shape[1] // 4, shape[0] // 2]], dtype=np.int64)
receivers = np.array([[[3 * shape[1] // 4, shape[0] // 2]]], dtype=np.int64)

with torch.no_grad():
    obs = solver(wavelet, sources, receivers, models=[torch.tensor(vp_true, device=device)])

vp_t = torch.tensor(vp_init, device=device, requires_grad=True)
pred = solver(wavelet, sources, receivers, models=[vp_t])
(0.5 * (pred - obs).pow(2).sum()).backward()

print("vp gradient shape:", tuple(vp_t.grad.shape))

Swap Acoustic for Elastic, AcousticVTI, ElasticTTI, ... — the surrounding code is unchanged.

Notebooks & examples

• Hello SWEEP — forward / backward / 5-line FWI loop: docs/notebooks/00_hello_fwi.ipynb
• FWI on Marmousi (acoustic / elastic / multiscale): see docs/notebooks/01_*–03_*
• Wavefields, DAS, anisotropic, RTM: docs/notebooks/04_*–08_*
• Production scripts (multi-GPU shot parallelism and domain decomposition, topography demos): under examples/

Citing

@misc{wang2026sweep,
  title  = {{SWEEP} ({S}eismic {W}ave {E}quation {E}xploration {P}latform):
            A Unified Solver Framework for Differentiable Wave Physics},
  author = {Wang, Shaowen and Alkhalifah, Tariq},
  year   = {2026},
  eprint = {2604.14189},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/2604.14189},
}

License

MIT — see LICENSE.