← Lab notebookOctober 14, 2026 · Stan DeLaurentiis

TCV foundation model
State of the work

Reconstruct five plasma fields. Transfer to a new equilibrium. Generate their evolution.

Truth → models → error, through time

Free forecasts from four observed volumes. Fixed physical color scales; predicted states feed back.

Field Component

R–Z native geometry. Fluctuations use native toroidal index 0; this is not a certified constant-azimuth cut. One sampled trajectory per diffusion model.

Electron, ion and total target heat

Quantity

Recovered sheath estimate, not measured tile loading. Five saved leads; no temporal interpolation. Radial internal-energy transport is a different diagnostic.

Choose a field, start and forecast lead

Open full-screen interactive viewer ↗ ·

Reconstruct the same volume

Mean/fluctuation weighting improved the reference compressor; spectra alone hid local errors.

Toroidal mean on the R–Z mesh. Five fields; truth versus four reconstruction losses. Full volume supplied independently at every time. 50:50 and 25:75 denote mean:fluctuation loss weights. MP4

Fluctuations on native radial–toroidal cells. Five fields; truth versus four reconstruction losses. Full volume supplied independently at every time. 50:50 and 25:75 denote mean:fluctuation loss weights. MP4

Balanced compressor: fluctuation amplitude and reconstruction error. PNG

Reconstruction objectives: electron, ion and total recovered target heat. PNG

Same-time reconstruction: local and integrated particle / energy transport. PNG

More figures from the earlier reports (32)

Reconstruction error across toroidal modes PNG

All physical fields: spatial profiles PNG

Local and integrated radial transport PNG

Upstream → downstream

Partial observations in; same-time downstream fields out. This is distinct from compression and forecasting.

Downstream adaptation: with versus without clean reconstruction replay. PNG

Downstream field relationships: convolution and attention comparisons. PNG

More figures from the earlier reports (28)

Transfer to another equilibrium

Small target-data fractions help. At 25%, fluctuation and transport scores have not all reached the 100% reference.

All five fields across 85606 fine-tuning fractions. PNG

Local and integrated particle / internal-energy transport versus target-data fraction. PNG

Chronological versus shuffled-split compressors on common 85606 validation. PNG

More figures from the earlier reports (5)

Which components remain worse at 5–25%? PNG

Fluctuation power recovered across fractions PNG

Does shuffling fix target-shot spectral content? PNG

Local relationships between fields

Pooled cross-phase can look good while local errors remain. Joint losses produce metric-specific gains.

Local versus pooled cross-phase. PNG · SVG

Learned field–mode weighting and independent codecs: local phase in lower modes. PNG

Joint-loss continuations: source-based target heat estimates. PNG

More figures from the earlier reports (51)

Density × Electron temperature — n=5…50: pooled error, average local error, retained signal PNG

Density × Electron temperature — n=5…200: pooled error, average local error, retained signal PNG

Density × Ion temperature — n=5…50: pooled error, average local error, retained signal PNG

Density × Ion temperature — n=5…200: pooled error, average local error, retained signal PNG

Density × Potential — n=5…50: pooled error, average local error, retained signal PNG

Density × Potential — n=5…200: pooled error, average local error, retained signal PNG

Density × Parallel velocity — n=5…50: pooled error, average local error, retained signal PNG

Density × Parallel velocity — n=5…200: pooled error, average local error, retained signal PNG

Electron temperature × Ion temperature — n=5…50: pooled error, average local error, retained signal PNG

Electron temperature × Ion temperature — n=5…200: pooled error, average local error, retained signal PNG

Electron temperature × Potential — n=5…50: pooled error, average local error, retained signal PNG

Electron temperature × Potential — n=5…200: pooled error, average local error, retained signal PNG

Electron temperature × Parallel velocity — n=5…50: pooled error, average local error, retained signal PNG

Electron temperature × Parallel velocity — n=5…200: pooled error, average local error, retained signal PNG

Ion temperature × Potential — n=5…50: pooled error, average local error, retained signal PNG

Ion temperature × Potential — n=5…200: pooled error, average local error, retained signal PNG

Ion temperature × Parallel velocity — n=5…50: pooled error, average local error, retained signal PNG

Ion temperature × Parallel velocity — n=5…200: pooled error, average local error, retained signal PNG

Potential × Parallel velocity — n=5…50: pooled error, average local error, retained signal PNG

Potential × Parallel velocity — n=5…200: pooled error, average local error, retained signal PNG

Density: mode RMS, RMSE and field-to-truth phase error PNG

Electron temperature: mode RMS, RMSE and field-to-truth phase error PNG

Ion temperature: mode RMS, RMSE and field-to-truth phase error PNG

Potential: mode RMS, RMSE and field-to-truth phase error PNG

Parallel velocity: mode RMS, RMSE and field-to-truth phase error PNG

Density × Electron temperature: all three snapshots at n=5,10,20, with local and pooled scores PNG

Density × Ion temperature: all three snapshots at n=5,10,20, with local and pooled scores PNG

Density × Potential: all three snapshots at n=5,10,20, with local and pooled scores PNG

Density × Parallel velocity: all three snapshots at n=5,10,20, with local and pooled scores PNG

Electron temperature × Ion temperature: all three snapshots at n=5,10,20, with local and pooled scores PNG

Electron temperature × Potential: all three snapshots at n=5,10,20, with local and pooled scores PNG

Electron temperature × Parallel velocity: all three snapshots at n=5,10,20, with local and pooled scores PNG

Ion temperature × Potential: all three snapshots at n=5,10,20, with local and pooled scores PNG

Ion temperature × Parallel velocity: all three snapshots at n=5,10,20, with local and pooled scores PNG

Potential × Parallel velocity: all three snapshots at n=5,10,20, with local and pooled scores PNG

Predict future volumes

The later diffusion recipes retained more turbulent power. That did not guarantee correct phase, trajectories or every transport metric.

Early diffusion versus later full-target and residual diffusion. The early-to-later recipe changes are not a single-variable ablation. PNG

Full-volume density spectrum across forecast leads. PNG · PDF

Full-target versus residual diffusion: recovered target heat through time. PNG

More figures from the earlier reports (25)

Truth autocorrelation by field and region PNG

Field errors: earlier deterministic and diffusion recipes PNG

Earlier diffusion: noise-level and conditioning probe PNG

Phase and full complex cross-spectrum error PNG

Which intervention improves which error? PNG

The density power budget, mode by mode PNG

Where and when target-heat errors occur PNG

Do gains persist through validation? PNG

Skill against the fitting-mean baseline PNG

All physical figures and movies

Earlier reports folded into one searchable collection. Historical model comparisons retain their original labels.

For slides

Download the compact movie & figure selection (about 7 MB). Individual movies, 300-dpi PNGs and PDFs are linked above.

Simulation data: Ben Zhu and collaborators. Models and diagnostics: Stan DeLaurentiis. Results are development comparisons; adapted 85606 validation is not an untouched test. Current longer-data reconstruction and controlled diffusion experiments are still being evaluated.

TCV foundation modelState of the work