DiffGRF: Differentiable Gaussian random field generation

Gaussian random fields (GRFs) are central to geostatistics, uncertainty quantification and Monte Carlo methods. Established implementations of the spectral randomization method are non-differentiable, so inverse parameter inference and gradient-based design rely on either finite differences or neural surrogates. DiffGRF is a PyTorch -native implementation that reparameterises the spectral summation such that analytical gradients flow through variance, correlation length, anisotropy and rotation. The package supports Gaussian and Matérn kernels in 1D, 2D and 3D, structured grids and unstructured meshes (via meshio ), and CPU, CUDA and Apple MPS execution. We report forward agreement with analytical kernels, autograd–finite-difference agreement at the O ( h 2 ) truncation floor of central differences, gradient-based parameter recovery, and wall-clock scaling.