DiffGRF

Differentiable Gaussian random field generation in PyTorch.

DiffGRF implements the spectral randomization method for Gaussian random field (GRF) synthesis (Kraichnan 1970; Hesse et al. 2014) with full PyTorch autograd support. Gradients flow analytically through every physical parameter — variance, correlation length(s), anisotropy, rotation angles and the Matérn smoothness — so the generator can be used as a differentiable building block in inverse problems, design optimisation, and torch-native training pipelines.

Features

• Gaussian and Matérn kernels
• 1D, 2D, and 3D spatial domains
• Isotropic or per-axis anisotropic correlation lengths
• Differentiable rotation (2D angle, 3D Tait–Bryan triplet)
• Structured (regular grid), unstructured (arbitrary points), and mesh-native (meshio / gmsh .msh) sampling
• CPU, CUDA, and Apple MPS device support
• End-to-end autograd through all physical parameters

Installation

pip install diffgrf

Or from source:

git clone https://github.com/EJRicketts/diffgrf.git
cd diffgrf
pip install -e .[dev]

Requires Python ≥ 3.10, PyTorch ≥ 2.0, NumPy ≥ 1.22.

GPU acceleration is available via device="cuda" (NVIDIA) or device="mps" (Apple Silicon). MPS does not support float64, so pass dtype=torch.float32 when using MPS:

import torch
from diffgrf import DifferentiableGRF

grf = DifferentiableGRF(
    kernel="gau", dim=2, variance=1.0, len_scale=10.0, seed=0,
    device="mps", dtype=torch.float32,
)
field = grf.structured([256, 256])

Observed speedups on Apple M-series over CPU: 4–5× at 128²–256² 2D, ~10× at 96³ 3D.

Quickstart

Generate a 2D Gaussian random field on a regular grid:

from diffgrf import DifferentiableGRF

grf = DifferentiableGRF(
    kernel="gau", dim=2, variance=1.0, len_scale=10.0, seed=0,
)
field = grf.structured([128, 128])
# field is a torch.Tensor of shape (128, 128)

Flow gradients through the correlation length:

import torch
from diffgrf import DifferentiableGRF

len_scale = torch.tensor(10.0, requires_grad=True)
grf = DifferentiableGRF(
    kernel="gau", dim=2, variance=1.0, len_scale=len_scale, seed=0,
)
field = grf.structured([128, 128])
loss = field.var()
loss.backward()
print(len_scale.grad)   # analytical gradient via spectral reparameterisation

Evaluate a field directly on an unstructured mesh (requires meshio, pip install diffgrf[mesh]):

import meshio
from diffgrf import DifferentiableGRF

grf = DifferentiableGRF(
    kernel="mat", dim=2, variance=1.0, len_scale=5.0, nu=1.5, seed=0,
)
# Accepts a meshio.Mesh OR any path readable by meshio.read.
# Evaluate at vertices (default) or cell centroids.
field_at_nodes = grf.on_mesh(meshio.read("domain.msh"))
field_at_cells = grf.on_mesh("domain.msh", location="cells")

Recover a correlation length from a target sample by gradient descent:

import torch
from diffgrf import DifferentiableGRF

target = DifferentiableGRF(
    "gau", 2, 1.0, 12.0, seed=0,
).structured([128, 128])

ell = torch.tensor(5.0, requires_grad=True)
opt = torch.optim.Adam([ell], lr=0.2)
for _ in range(200):
    opt.zero_grad()
    pred = DifferentiableGRF("gau", 2, 1.0, ell, seed=0).structured([128, 128])
    loss = (pred - target).pow(2).mean()
    loss.backward()
    opt.step()
print(float(ell))   # -> ~12

API

DifferentiableGRF(kernel, dim, variance, len_scale, ...)

The single entry point. See diffgrf/core.py for full signature and docstrings.

Methods:

• .structured(grid_dim) — generate on a regular grid.
• .unstructured(points) — generate at arbitrary spatial points.
• .on_mesh(mesh, location='points') — generate on a meshio.Mesh or a path to any meshio-compatible file (.msh, .vtk, .xdmf, ...); evaluate at vertices (location='points') or cell centroids (location='cells'). Requires the optional meshio dependency (pip install diffgrf[mesh]).

Citation

If you use DiffGRF in published work, please cite:

@software{diffgrf,
  author = {Ricketts, Evan John},
  title  = {DiffGRF: Differentiable Gaussian random field generation in PyTorch},
  year   = {2026},
  url    = {https://github.com/EJRicketts/diffgrf},
}

A peer-reviewed SoftwareX paper describing DiffGRF is in preparation; the citation will be updated once published.

License

MIT. See LICENSE.