TCV · selected physical figures and movies
October 14, 2026. Download the original files for slides. This compact selection contains seven panels; the complete gallery adds all fields, all experiments, and interactive views. No new training results are claimed here.
1 · Density fluctuations: truth, forecasts and error
Actual autoregressive forecasting from four observed volumes. Top: truth and three models. Bottom: signed model − truth. R and Z in metres; shared fixed physical scales; lead and microseconds labeled in every frame. Fixed native toroidal index 0, not a certified constant-azimuth cut.
MP4 · PNG · PDF
Original context: docs/physical_truth_review_20261009/model_guide.html
2 · Recovered target heat through time
Total outer-target sheath estimate: truth and three forecasts, with toroidally averaged profiles below. Native toroidal coordinate in degrees, target arclength in mm, heat in MW/m². Five saved leads only; no interpolation. This is not a measured target load.
MP4 · PNG · PDF
Original context: docs/physical_truth_review_20261009/rollout_physics_verification.json
4 · The mean evolves differently from fluctuations
Instantaneous toroidal n=0 density on the native R–Z geometry. Compare this with the first fluctuation movie at the same start and times. Both are forecasts, with truth and signed errors.
MP4 · PNG · PDF
Original context: docs/physical_truth_review_20261009/rollout_verification.json
5 · Same-time full-volume reconstruction
All five fields, toroidal n=0, truth and four reconstruction objectives on the rendered R–Z mesh. Every frame is independently reconstructed from its complete volume; this is not a forecast. Historical comparison: ordinary MSE, two mean/fluctuation weightings and spectral loss.
MP4
Original context: docs/physics_movies_20260925/README.md
6 · The spectrum behind the movie
Full-volume density fluctuation variance by toroidal mode n=5k at four forecast leads. Positive/negative Fourier pairs are combined. Same start as the movies. Matching this power does not establish correct local phase.
PNG · PDF
Original context: docs/physical_truth_review_20261009/rollout_physics_verification.json
7 · Local versus pooled cross-phase
Compare phase relationships locally before averaging. Opposite signed local errors can cancel in pooled complex spectra; amplitude and valid-signal coverage also matter.
PNG · SVG
Original context: docs/phase_followup_20261013/all_plots.html
8 · Physics gains from joint reconstruction losses
Matched 6,400-update continuations: joint patch distribution reduced total target-heat RMSE by 6.7%; local covariance improved integrated particle error by 9.9%. Neither uniformly improved local phase. Heat is a recovered sheath estimate, not a measured target load.
PNG
Original context: docs/joint_losses_20261013/results.html