TCV foundation model · October 16 results

The five findings

1. Long-data reconstruction: pretrained balanced leads the five fluctuation errors within the six matched long-data runs. The old/long overlay is a comparison across source records, not a controlled estimate of data scaling.
2. Fluctuations: matching power or pooled cross-phase can hide local amplitude and phase errors. We now show those separately.
3. Diffusion: forecast conditioning helps; history noise trades better particle transport against worse target heat. Local realization fidelity remains unresolved.
4. 85606: all fine-tuning fractions completed; more target data continues improving all five fluctuation errors.
5. XY → divertor: XZ-task pretraining finished, but the physical patch still does not match full-XZ input accuracy. Target-only errors are less severe than downstream-volume errors.

Arrows are weight inheritance except the chronological model → temporal preparation dependency. Scratch long-data models have independent initialization. Click a branch to see its comparison. Complete historical tree and configuration archive.

1. Longer data versus shorter data

Question: do more data and different reconstruction objectives preserve more of the plasma state?

Read overlays correctly: dashed short-data curves use short-data truth; long-data curves use long-data truth. Spectra show both truths. Different source records/splits are not matched test cases. Within the six long-data models, the validation data are identical. “Pretrained” means our short-data balanced codec, not historical LoLA.

Short / long × chronological / random

Split	Short legacy record · 624 frames	Long final record · 2,712 frames
Chronological	Fit 0–431; exclude 432–495; validate 496–623. Continuation plots use 500–623.	Fit 0–2168; exclude 2169–2232; validate 2233–2711. Final evaluation pending.
Random	432 fit / 128 validation / 64 unused, scattered across 624; seed 20261008.	2,169 fit / 543 validation, scattered across 2,712; seed 20261014.

Random split asks about unseen snapshots interspersed with training data; chronological split withholds a later interval. Both are reconstruction tasks with the complete target-time volume as input. Short and long models always use their own validation truth. See Section 2 for their error curves.

Physical errors and target heat · short reference overlaid with long models

Figure All five fields: full, mean and fluctuation errorsLocal and integrated radial transportElectron recovered target heatIon recovered target heatTotal recovered target heat

Spectra and complex errors · all five fields

Figure Density modesElectron temperature modesIon temperature modesPotential modesParallel velocity modes

Local versus pooled cross-phase · all ten field pairs

Figure Density × Electron temperatureDensity × Ion temperatureDensity × PotentialDensity × Parallel velocityElectron temperature × Ion temperatureElectron temperature × PotentialElectron temperature × Parallel velocityIon temperature × PotentialIon temperature × Parallel velocityPotential × Parallel velocity

Controlled six-way long-data comparison and earlier short-data controls

Matched objectives

Figure Long-data six-way field errorsLong-data six-way radial transportLoss tradeoffs relative to pretrained balancedEarlier short-data continuationsEarlier short-data transport

Balanced loss: equal normalized contributions from toroidal mean and fluctuations. Local covariance loss: auxiliary matching of local cross-field second moments. Joint-patch distribution loss: auxiliary distribution matching for multifield patch vectors; it does not guarantee that every reconstructed structure is in the correct place.

Shuffled versus chronological: separate experiments

Ablation	What changed	Finding
Training order	Individual frames versus four-frame bins within the same fitting interval	Nearly identical fluctuation errors; validation split unchanged.
Validation split	Chronological versus randomly assigned legacy frames	Random-split own-shot fluctuation errors were about 11–18% lower, but on different validation frames. On identical 85606 frames, differences were below 1.5%.
Short versus long, both random split	Different source records; random held-out frames in each	Descriptive comparison, not a pure data-size ablation.

Earlier split ablations

Figure Chronological versus random split · all five fieldsChronological versus random · transfer transportShort random versus long random · reconstruction

Earlier training-order ablation details. The long chronological comparison has no evaluated final result in this report snapshot.

2. What the fluctuation investigation resolved

Two different questions: how much fluctuation is present, and whether the model reconstructs the actual fluctuation at the right place. The first can be right while the second is wrong.

Split comparison · fluctuation-error curves through time

Figure Density · short and long split comparisonsElectron temperature · short and long split comparisonsIon temperature · short and long split comparisonsPotential · short and long split comparisonsParallel velocity · short and long split comparisons

See why matching spectra is insufficient

Figure Cross-phase explainedFluctuation magnitude versus actual reconstruction errorLocal versus pooled errors

Fluctuation reconstruction error · line plots

Figure Density · fluctuation error through time and by modeElectron temperature · fluctuation error through time and by modeIon temperature · fluctuation error through time and by modePotential · fluctuation error through time and by modeParallel velocity · fluctuation error through time and by mode

Resolution: the earlier spectrum plots were useful amplitude diagnostics, but we overinterpreted them as evidence of local reconstruction fidelity. Keep amplitude, complex error, local cross-phase and transport together. Better pooled phase alone is not a demonstrated fix.

3. Diffusion: what we changed, and what it helped

Four fresh seed-0 denoisers, 16,384 updates each, same frozen short-data codec; 32 DDIM sampling steps. Evaluation: 105 overlapping chronological validation starts, one sampled path per start, leads 1–16 (about 3.13–50.11 µs). These are not ensemble-calibration results. The physical-loss and history-noise arms are separate changes to the forecast-conditioned control.

Forecast performance through time

Figure Fluctuation error and amplitudeAll five physical fieldsParticle and energy transportElectron, ion and total target-heat errorsTarget heat: truth and forecastForecast toroidal spectraPooled cross-phase at lead 1Pooled cross-phase at lead 16

Forecast conditioning improved the first-step result and retained more power. History noise reduced integrated particle error but worsened target-heat error relative to clean forecast conditioning. No one recipe wins every observable. The new long-data and block/rolling tests have not yet produced results.

Earlier temporal ablations: what these builds on

Earlier families

Figure Deterministic training objectivesDiffusion history mixture and noise weightingHorizon and frame-weight curriculaFull-target versus residual diffusion

Latest evaluation provenance · Per-field values and RMSE skill · Earlier curriculum methods/results

4. How much 85606 data is needed?

Fractions 5/10/15/20/25/100% use 22/43/65/86/108/432 eligible fitting frames. 0% is the unadapted model. All use the same evaluation frames [496,624); 100% does not include guard or validation frames. Each fraction starts from the same parent rather than continuing the previous fraction. Balanced reconstruction objective; no old-shot replay in these runs.

Completed long-parent fraction sweep

Figure All five fluctuation errorsLocal and integrated radial transport

Earlier comparisons and retained source skill

Figure Short versus long parentRetained long-source reconstruction

There is no demonstrated fraction where all metrics saturate: 25% remains worse than 100% on all five fluctuation errors. The 85606 record has now been used for adaptation and repeated development evaluation; it is not a fresh blind shot test.

5. Can a tiny XY input match full-XZ input at the divertor?

This is the actual observation goal. “Full XZ” describes the OMP input slab. The output of interest is a different XZ plane at the outer divertor.

Open the 3D viewer full-size · Verified geometry definitions

Training: five fields on the downstream region x16–63, y20–31, including target y31; normalized mean/fluctuation GradNorm with clean reconstruction replay coefficient 0.5. AdamW 10⁻⁵ with 100-update warmup, 30,000 updates per adaptation stage. Fit [0,432); evaluate [500,624). Clean replay bypasses the observation adapter. These all use the short-data codec.

Evaluation: target-only XZ = x16–63, y31, all 81 toroidal positions. Broader downstream-volume scores are separate. We have not tested a target-only training objective in this batch.

Read the lines: one model key for every target plot

Black = simulation truth. Model colors and line styles below are identical in the heat, power-spectrum and local cross-phase figures. Electron / ion / total identify the physical quantity, not different models.

Full OMP XZ input
Reconstruction checkpoint → 30k full-OMP-XZ-to-divertor updates; accuracy reference

Direct XY patch
Reconstruction checkpoint → 30k physical-patch updates

XZ-pretrained XY patch
Full OMP XZ input checkpoint → 30k physical-patch updates

XZ-pretrained XY patch · longer
Same full-OMP-XZ parent → 30k + 30k physical-patch updates

XZ → radial line → XY patch
Same full-OMP-XZ parent → 30k native-radial-line updates → 30k physical-patch updates

Electron heat is the recovered electron contribution at the target; ion heat is the recovered ion contribution; total heat = electron + ion, summed before computing the error. Left panel: area-mean truth/model heat flux. Right panel: local error before area averaging; lower is better. Both use the exact outer-divertor target and common finite support.

Target heat · electron, ion and total

Figure Electron target heatIon target heatTotal target heat

Target field errors

Figure Outer-divertor fluctuation errors, all five fields

Downstream region versus the target: why the earlier curves looked better

The spectra in the page linked to Ben on October 14 averaged the downstream interior and excluded the boundary y31. The exact-target figures below score y31 alone. “Downstream region” and “outer-divertor target” are therefore separate choices, not interchangeable labels. The earlier model also used direct masked-volume input, whereas the current full-XZ control uses a trainable observation adapter.

Current models · downstream-region spectra

Figure Density · downstream interior spectraElectron temperature · downstream interior spectraIon temperature · downstream interior spectraPotential · downstream interior spectraParallel velocity · downstream interior spectra

Earlier Ben reference versus current full-XZ control

Figure Density · earlier view and matched region/target checkElectron temperature · earlier view and matched region/target checkIon temperature · earlier view and matched region/target checkPotential · earlier view and matched region/target checkParallel velocity · earlier view and matched region/target check

Same-target check over all 124 frames: current full-XZ fluctuation RMSE differs from the earlier adaptive+replay model by −1.14% density, −0.08% electron temperature, −1.06% ion temperature, −1.92% potential and −0.93% parallel velocity. This is near-parity, not a target collapse. Both can still have substantial target error. Checkpoint-result and support audit.

Same saved target-only evaluations as the earlier report. This is sparse-input inference at x16–63, y31, not full-volume reconstruction. Low-mode power is underestimated; sharp relative high-mode excess is amplified by very small truth power. The factor of two counts both ±k and changes truth and model together, not their ratio. The source of the narrow peaks is not yet established.

Outer-divertor power spectra · all five fields

Figure Density target power and complex errorElectron temperature target power and complex errorIon temperature target power and complex errorPotential target power and complex errorParallel velocity target power and complex error

Outer-divertor LOCAL cross-phase error · all ten field pairs

Figure Density × Electron temperature target local phaseDensity × Ion temperature target local phaseDensity × Potential target local phaseDensity × Parallel velocity target local phaseElectron temperature × Ion temperature target local phaseElectron temperature × Potential target local phaseElectron temperature × Parallel velocity target local phaseIon temperature × Potential target local phaseIon temperature × Parallel velocity target local phasePotential × Parallel velocity target local phase

Target spectral values · Target local-phase values and collapsed-signal coverage · Common-truth and frame checks

Broader downstream region · separate from target-only plots

Figure Downstream-volume full/mean/fluctuation errors

XZ-task pretraining did not make XY match XZ. On the exact target, the XZ-pretrained 30k patch model has about 20% higher density fluctuation error and 49% higher electron-temperature fluctuation error than full-XZ input. That gap is smaller than the 80%/123% increase in the broader downstream-volume scores. The direct-from-compressor patch model is better than the XZ-pretrained patch on all five exact-target fluctuation errors. The radial-line intermediate does not close the gap.

Definitions used throughout

What is running, and what remains a question?

Checked 2026-10-16T15:29:26.607786+00:00. Chronological long-data compressor: 21760 / 51200 updates, running.

Timestamped scheduler snapshot · Queued temporal protocol · Original detailed observation maps · Complete experiment archive