GETTING STARTED

Your first FWI in five steps.

From a fresh terminal to running gradient descent on a toy velocity model. By the end you'll be ready to open any of the notebooks.

01

Install

PyTorch is always installed — it is a required dependency. The [jax] extra adds JAX on top, for the JAX backend.

PyTorch backend recommended

$ pip install sweepx

JAX backend

$ pip install "sweep-solver[jax]"

tip The wheel ships a prebuilt CUDA core, so the compiled backend (impl="c") works right after install — nothing compiles. For a dev clone, see Installation.

02

Build a velocity model

A simple 2-layer model: 1500 m/s in the top half, 2000 m/s below. 10 m grid spacing.

quickstart.py

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

# A 2-layer 100×100 model.
vp = np.ones((100, 100), dtype=np.float32) * 1500
vp[50:, :] = 2000

vp · 100 × 100 · 10 m grid

03

Run a forward shot

Construct the equation, wrap it in a propagator, and call forward().

propagator

eq   = Acoustic(spatial_order=8, device="cuda", backend="torch")
prop = PropTorch(eq, shape=(100, 100), device="cuda", dh=10., dt=2e-3,
                 source_type=["h1"], receiver_type=["h1"])

wave      = ricker(np.arange(0, 1.5, 2e-3) - 0.1, f=8)
sources   = np.array([[50, 2]])
receivers = np.array([[[ix, 2] for ix in range(10, 90)]])

obs = prop.forward(wave, sources, receivers, models=[torch.from_numpy(vp).to("cuda")])

shot record · 1 src · 80 rec · 750 samples

04

Get the gradient

Tag vp with requires_grad_, run forward, and backward(). That's it.

vp_t = torch.from_numpy(vp).to("cuda").requires_grad_(True)
obs  = prop.forward(wave, sources, receivers, models=[vp_t])
obs.pow(2).sum().backward()

# vp_t.grad now holds the sensitivity kernel.