Does an ENSO ocean perturbation move the atmosphere more than emulator noise?
Atmosphere SFNO latents (384 channels, 8 Fourier blocks) from three single-member runs with matched batch layout: control (seed 0), perturbed (same seed 0, ENSO-perturbed ocean IC) and control-seed1 (unperturbed, new noise draw). Quantities are area-weighted RMS latent differences from the control, divided by the control's own global spread. One perturbed member, one start (1985-01-01), 20 days: a feasibility test, not a statistic.
perturbed − control (ocean IC change, same noise)
control-seed1 − control (new noise only)
1. Difference from control vs time
Hover for values. Log scale shows the early orders-of-magnitude gap.
2. Perturbation as a multiple of the noise effect
Dashed line = 1: the ocean perturbation matters as much as a new noise draw. Below 1, noise dominates.
3. Layer × time heatmap of the ratio (perturbed / noise)
Selected region. Click a row to select that layer.
4. Where the difference is (selected layer and time)
Top: perturbed − control. Bottom: seed1 − control. Same colour scale, per-node RMS over channels / global spread, 3° blocks.