Power spectrum · Target heat flux
Four observed volumes → frozen encoder → predict next latent → frozen decoder.
The predicted latent goes back into that model’s history. No future truth is supplied after initialization. Full-target diffusion generates the next latent; Diffusion Plus adds a generated correction to its deterministic forecast.
Loading saved trajectory…
Top row: truth and each model on the same physical color scale, fixed over all 16 leads. Bottom row: model − truth on a shared signed-error scale, also fixed over time. Display uses 256 color levels. R–Z polygons follow native mesh connectivity. Fluctuations are at fixed native toroidal index 0; this is not a certified constant-physical-azimuth cut. Mean is instantaneous toroidal n=0.
These curves use native-volume weights on the displayed mean map or single toroidal fluctuation slice. They are not full-volume fluctuation power or RMSE. Each point belongs to this one continuous rollout, not a new forecast initialized from truth. Playback is slowed down for viewing. Each diffusion curve is one sample; it does not establish ensemble calibration.
Axes stay fixed through time for the selected field and start. At each lead: native-volume-weighted power in the full volume, with the positive/negative Fourier pair counted together. Summing n > 0 gives fluctuation variance; its square root is fluctuation RMS. Equal power does not establish correct structure locations or phase. The field selector controls these plots; the mean/fluctuation map selector does not change their definition.
Saved heat steps (use “Saved heat steps” playback to watch only these):
Recovered outer-target normal sheath heat, not radial internal-energy transport or measured tile loading. Computed from each predicted joint state. Ion heat uses the imposed Bohm boundary speed, so it does not directly test the predicted ion velocity. Profile curves use identical valid cells across truth and all models. No averaging over time or starts. Each diffusion model is one sampled rollout.