Fusion · Plasma stability · Trustworthy ML

Fast surrogates you can trust, or know not to.

Machine-learning surrogates of plasma stability answer in milliseconds, but they rarely say how wrong they might be, or when a new design lies outside what they learned. dataquanta builds the trust layer: calibrated error bars, out-of-distribution alarms that route doubtful cases to the physics code, and recalibration from a few simulations of a new device.

Any surrogate · model-agnostic Pre-registered · tests and failures reported arXiv:2611.06597 · published

Read the paper  → Run a pilot on your surrogate  →

01 · Approach

Wrap the surrogate you already have.

Instead of building yet another surrogate, the trust layer wraps whichever one you use: an in-house network, a published model such as KARHU or ELMO, or a neural operator. Every prediction comes back with a calibrated interval. Two alarms check whether the input resembles the calibration data; if it does not, the case goes to the physics code. A new device or plasma shape is onboarded with a few dozen of its own simulations rather than a new training campaign.

EquilibriumG-EQDSK or profiles
Any surrogateyour model
Conformal barcalibrated interval
OOD alarmsin range?
Decideor run the physics code

02 · Results

Tested with pre-registered hypotheses.

89.9% coverage at a 90 % target on 38,879 public equilibria with ideal-MHD growth rates (KARHU dataset).
~30 simulations of a new plasma shape restore 90 % coverage after a shape shift (90.3 %), with the bar close to its original width.
92.3% coverage at 90 % on 42 held-out resistive BOUT++ runs; a 42-point scenario scan in about 2 s instead of about 12 h, with no wrong confident decision.
0.44× the baseline error on unseen SPARC equilibria (96 SPARC and DIII-D equilibria, whole equilibria held out). A surrogate trained elsewhere ranks the same cases backwards, and its alarm declines them.

Limits, stated plainly: the tests are against simulation codes, not yet against measured ELM onset; the error bar holds for cases that resemble the calibration set, which is why the alarms exist; and not every pre-registered hypothesis held (letting the layer pick further simulations saved 10 %, short of the 30 % we had pre-registered). Details and all numbers are in the paper.

03 · Publications

Publications.

arXiv:2611.06597 · physics.plasm-ph · 2026

A trust layer for fast pedestal-stability surrogates: calibrated error bars, out-of-distribution alarms and few-shot recalibration

Amin Vakhshouri · dataquanta LLC

A model-agnostic trust layer for stability surrogates, tested with pre-registered hypotheses on a large public dataset of JET-like equilibria and on resistive BOUT++ surrogates across equilibria from the SPARC library and the DIII-D coil set. For every surrogate tested, the trust layer keeps coverage near target, with a bar width set by that surrogate's own error.

04 · Position in the literature

Complementary to the tools you already use.

ELITE, PB3D and BOUT++ compute stability from first principles, at seconds to hours per equilibrium. ML surrogates such as KARHU, EuroPED-NN and ELMO answer in milliseconds but mostly as point estimates. The trust layer sits on top of any of them: it does not replace the physics codes or the surrogates, it tells you when the fast answer can be used.

WorkOutputSpeedError bars / OOD
ELITEWilson et al., PoP 9, 1277, 2002 γ (ideal MHD) ~tens of s —
EPEDSnyder et al., NF 51, 103016, 2011 pedestal height / width ~min —
PB3DWeyens et al., JCP, 2017 γ (ideal MHD) ~min —
KARHUBruncrona et al., PoP 32, 092501, 2025 γmax ms-class —
EuroPED-NNPanera Alvarez et al., PPCF 66, 095012, 2024 pedestal scalars ms-class Bayesian NN
dataquanta trust layerarXiv:2611.06597, 2026 wraps any of the above ms-class conformal bars · OOD alarms · few-shot recalibration

05 · Work with us

For fusion companies and labs.

Pilot on your surrogate

A scoped 6–10 week project: we wrap one of your existing surrogates or simulation datasets with calibrated error bars and alarms, and report where it can and cannot be trusted.

Surrogate validation

Independent, pre-registered checks of a surrogate before it enters a design or control workflow, including behaviour on new devices and plasma shapes.

Research partnership

Collaboration with groups holding stability databases or experimental ELM-onset data, and joint research proposals.

Now building: the same approach for liquid-metal flow in fusion breeding blankets, where 3D magnetohydrodynamic pressure drops are expensive to simulate and hard to predict. In development.

06 · Contact

Open conversations.

info@dataquanta.io

For pilots, collaboration, or technical questions.

Paper: arXiv:2611.06597  →

  • Pilots trust layer on your surrogate
  • Validation independent surrogate checks
  • Research collaboration on data and methods