Sampling with Log ΔB²

Cutout centers are sampled with probability weighted by the surface buoyancy gradient magnitude squared (ΔB²). This intends to concentrate cutouts on dynamically active regions. Sampling runs on log10(ΔB²) — see Weighted Sampling.

Log ΔB² calculation

ΔB² is computed in generate-global (the frontal_structure subset) and stored as the gradb2 channel; the cutout job reads it and takes log10 at sample time (processing.sample_cutout_centers_with_loggradb). Reference impl: calculate_additional_fields.grad_b2 / log_grad_b.

Buoyancy

Surface buoyancy b is derived from Theta and Salt (physical_calculations.buoyancy_of_field).

ΔB²

Horizontal gradient of b on the native grid, then squared magnitude: ΔB² = (∂b/∂x)² + (∂b/∂y)² (units s⁻⁴).

Log scale

ΔB² spans many orders of magnitude, so we sample on log10(ΔB²). Combined with the exponential weight exp(bias·(v − min)), bias then acts as a power-law knob on ΔB².

ΔB² probability map

Sampling uses the weighted coordinate sampling module on log10(ΔB²), restricted to the ocean mask (land + ice halos). Higher bias → more concentrated on fronts.

Sampling points

Controlled by sampling.sample_points_per_snapshot and sampling.bias_to_high_gradients in the run config.

River outflow issue

Freshwater river outflows produce very large ΔB², which can dominate the weighting and pull a disproportionate share of cutouts toward river mouths. This is controlled by the bias weight. If the weight is too high, all samples will be drawn from near freshwater or ice melt. The tuned value is 1.3.