Working notes from ongoing research — written as the work happens, not after. Exploratory and work-in-progress by design: single-trajectory experiments, numbers with stated caveats, dead-ends included. Read them as a record of what was tried and what came out, not as finished results.
An update to the July petch evidence pack: from a plasma tool's settings through the rf sheath to 3-D etched features on real chip layouts, a same-GPU head-to-head with ViennaPS scored against measured holes, predictions of held-out wafer data, and checks against textbook physics.
Short and long data, chronological and random splits, fluctuation errors, and downstream-region versus target spectra. Physical figures, model comparisons and an interactive 3D observation geometry.
Figures and movies with what each compares: Sironi's kinetic dynamo runs, learned closures scored on heating and force across ion mass, the first closure-in-MHD test, the TCV edge and its connection to the TORAX core, and how well the trust score ranks and calibrates.
Interactive plasma slices, anisotropy spectra and distributions, mirror/firehose diagnostics, and transfer across box size and mass ratio. Includes completed coarse-scale and raw-pretraining evaluations.
Physical movies and figures: five-field reconstruction, upstream-to-downstream inference, cross-shot adaptation, and autoregressive diffusion forecasts.
Two magnetizations, one 45° passage: accretion, magnetic flux, and the question of memory between encounters. Interactive simulation slices and measured histories.
September research review: models, inputs and prediction results →
Edge filaments follow the magnetic field, so a whole radial–toroidal plane at the divertor target should be a sheared image of the midplane plane. On Hermes-3 simulations of TCV it is: a rigid toroidal shift finds nothing, the equilibrium field-line map finds it at zero lag, structures wider than ~5 cm are inherited and blob-scale ones are not, mean density and momentum arrive later at the sound speed, and a toroidally equivariant linear map predicts the plate as well as any operator transformer. Three tabs: planes and correlations, a 3-D torus, and a prediction atlas with a pre-registered ridge-versus-GAOT pilot. Movies, captions, and open questions included.
A self-contained account of the two-shot dataset, LOLA training procedures, representation choices, rollout calibration, ETKF sensor conditioning, and the physics tests needed before diagnostic rankings can be interpreted as design guidance.
Measured plasma fluxes and a material stack in; coupled transport, surface chemistry, self-consistent charging, and a moving 3-D profile out. Reproduces a published SiO₂ trench experiment to 0.19% on the mask opening, measured head-to-head speed against the open-source state of the art, and a preregistered blind-transfer protocol with the held-out data mechanically sealed. Movies, scorecards, and claim boundaries included.
Four model families — a deterministic ViT operator, a geometry-aware graph transformer, predictive self-supervision, and latent diffusion — each tried on one Hermes-3/BOUT++ trajectory of tokamak edge turbulence, scored on physics gates rather than pointwise error. The recurring enemy is "the freeze": the collapse of MSE-trained models onto the conditional mean.
Given the generative emulator of notebook I, what can a realistic set of edge diagnostics (gas-puff imaging, divertor Langmuir probes, reflectometry) actually constrain? Anchored rollouts with synthetic sensors, two misleading signals caught along the way, and a first-pass observability estimate with its caveats stated up front.