odmr-spectra

50,000 simulated NV-center ODMR spectra with exact physical labels, for training and benchmarking spectral estimators.

split	rows	regime
train	40,000	0.8–5 mT, 2–8% contrast, 4–12 MHz linewidth
validation	5,000	same distribution as train
test_hard	5,000	0.05–1.2 mT, 0.3–1.5% contrast, 8–20 MHz linewidth, 280–330 K

test_hard is deliberately out of distribution: weak fields, low contrast, broad overlapping lines, wide temperature range. It exists so that estimators are reported in the regime where they actually break, not only where they look good.

Schema

column	type	description
spectrum	list[float32] × 512	normalised fluorescence, 2700–3040 MHz
p1_mT–p4_mT	float32	|B·nᵢ|, field projection onto each NV axis
D_shift_MHz	float32	zero-field-splitting shift (−74 kHz/K)
E_strain_MHz	float32	transverse strain term
linewidth_MHz	float32	Lorentzian FWHM
B_magnitude_mT	float32	|B|, equals √(0.75·Σpᵢ²)
contrast	float32	ODMR dip depth
photons_per_point	float32	readout photon budget, sets shot-noise floor

The frequency grid is in freq_grid_MHz.npy (512 points, uniform).

Generation

Spectra come from exact diagonalisation of the NV ground-state Hamiltonian,

H = D(S_z² − 2/3) + E(S_x² − S_y²) + γ(B_x S_x + B_y S_y + B_z S_z)

per NV axis, with γ = 28.024 MHz/mT and D ≈ 2870 MHz. Four ⟨111⟩ orientations give eight lines; ¹⁴N hyperfine coupling (2.16 MHz) splits each into a triplet, for 24 components. Each spectrum carries Poisson readout noise plus a slow instrumental baseline (laser power and collection-efficiency drift).

The Hamiltonian is diagonalised exactly rather than expanded to first order, because the expansion fails precisely in the transverse-field and overlapping regimes this dataset is built to cover.

generate.py and odmr.py reproduce every split from seed.

Why the labels are the projections

An ODMR spectrum fixes the magnitude of the field projection onto each NV axis. The sign is unobservable — the crystal's point-group symmetry maps sign flips onto identical spectra. Regressing the Cartesian vector would mean training on an ill-posed target. The four projection magnitudes are uniquely identified, and |B| follows exactly from the tetrahedral identity Σᵢ(B·nᵢ)² = (4/3)|B|².

Limitations

Fully synthetic. For ODMR this is normal — the lineshape is known physics, so labels are exact and noise-free ground truth is available, which is not true of most sensing datasets. But real spectra contain microwave power broadening, laser noise correlated across the sweep, frequency calibration error, and diamond-specific strain distributions that are not modelled here. A model trained only on this data should be validated on measured spectra before deployment.