Instances

An instance is a complete, reproducible inverse problem: a config dataclass in which every paper number is a field, scene generators with difficulty classes, an independent high-fidelity data generator, the problem builder, a default curriculum, metrics and baselines. The CLI, the benchmark harness, the diagnostics, the auto-tuner and the gallery all work on any instance, including your own (A new instance).

nefi list instances -v               # the registry: scene classes, baselines, metrics, configs
nefi run <name> --smoke              # seconds on a CPU
nefi run configs/<name>_full.yaml --device cuda
nefi bench <name> --methods neural,grid --classes all --n 8 --seeds 0,1,2

NV relaxometrypaper · NeTMY Sparse-view CTclassic imaging EITelliptic Holographyoptics Full-waveform inversionwaves 3-D sparse-view CTvolumetric

The papers

instance	problem	unknown	physics	measurement
nv_relaxometry	NeTMY — NV-center noise sensing	spin density ρ ≥ 0 and Larmor field ωL (2-D)	dipolar tensor power kernel × Lorentzian (F2)	noise spectra, 50 frequencies × 64²
thermal_tomography	NeFTY — pulsed-thermography tomography	diffusivity α ∈ [αmin, αmax] (3-D)	implicit-Euler heat equation, discrete adjoint	100 surface frames × 64²

Classic imaging

instance	problem	unknown	physics	measurement
toy1d	1-D deblurring (tutorial, CI)	signal x ≥ 0	FFT Gaussian blur	128 samples
deconvolution	image deblurring, Gaussian or Poisson noise	image x ≥ 0	FFT convolution (Gaussian / motion / disk PSF)	blurred image
sparse_view_ct	sparse-view and limited-angle CT	attenuation μ ∈ [0, 1]	differentiable Radon transform	sinogram, 20 views

Elliptic PDEs

instance	problem	unknown	physics	measurement
poisson_source	source recovery	source f ≥ 0	−κΔu = f, spectral solver	potential on 10 % of pixels, a boundary strip or all
eit	electrical impedance tomography	conductivity σ	−∇·(σ∇u) = 0, implicit-function adjoint	boundary voltages of current patterns
darcy_flow	hydraulic tomography	log-permeability	−∇·(k∇p) = q, implicit-function adjoint	well-gauge pressures, 6 configurations
current_density	NV current imaging	stream function of a sheet current	planar magnetostatics (FFT)	stray field Bz

Wave, optics and reaction

instance	problem	unknown	physics	measurement
wave_fwi	full-waveform inversion	sound speed c	acoustic wave equation, leapfrog + PML	shot gathers
diffraction_tomography	optical diffraction tomography	scattering contrast χ	Born / Rytov scattering	complex fields, 16 angles × 64 receivers
holography	inline phase retrieval	phase φ	angular-spectrum propagation	intensities at 3 distances
reaction_diffusion	Gray–Scott parameter inversion	feed-rate field F	reaction–diffusion time stepping	4 snapshots of both species

Volumetric (3-D)

instance	problem	unknown	physics	measurement
ct3d	3-D sparse-view CT	attenuation μ(x, y, z)	slice-stacked Radon transform	sinogram stack, 30 views
deconvolution3d	fluorescence z-stack deconvolution	fluorophore density	3-D FFT convolution, elongated PSF	blurred z-stack
dot3d	diffuse optical tomography	absorption μa	diffusion equation, Robin boundaries, implicit-function adjoint	top-face reflectance, 3 × 3 sources
photoacoustic3d	photoacoustic tomography	initial pressure p0	3-D wave equation, sponge layer	traces of a 16 × 16 planar array

thermal_tomography is volumetric as well; with the four systems above it forms the 3-D showcase.

Every instance page has

• At a glance — the problem, the unknown, the physics, the measurement, the difficulty classes, the baselines, the metrics and the run command, with the gallery figure;
• the forward model and its discretization, and the inverse-crime guard: how the data are simulated independently of the inversion operator;
• the prior, the objective, the curriculum and the full configuration;
• measured results — validation against analytic cases, smoke and paper-scale runs, baselines — and the known limitations, where they have been measured.