Fusion · Plasma stability · Neural operator
Fast machine-learning surrogates for fusion plasma stability.
71 ms peeling-ballooning growth-rate inference from a magnetic equilibrium, on CPU. Roughly two to three orders of magnitude faster than the ideal-MHD codes commonly used as the inner stability solver inside EPED-class workflows.
01 · Approach
A neural operator on the equilibrium, not a fit on profiles.
plasmafast is a neural-operator surrogate — a Flux-Surface Attention / Vector-Potential Neural Operator — that maps a G-EQDSK magnetic equilibrium, plus resistivity η and perpendicular viscosity ν, to the linear peeling-ballooning growth-rate field γ(n, ρpol). End-to-end inference is 71 ms on CPU (36 ms equilibrium parse, 35 ms model forward), against tens of seconds for ELITE and minutes for PB3D on the same problem class. The model is trained on a synthetic BOUT++ cbm18-family dataset and is designed for drop-in integration with real-time tokamak control loops.
843,992 parameters · canonical checkpoint 2026-05-22 (exp_059_v2)
02 · Validation snapshot
What we can show today.
All three acceptance gates pass on the current canonical model (cbm18 γ within 30 %, closed-loop rollout ≤ 1.10, SPARC stability classifier 14 / 15 — unchanged from the pre-Gate-1-closure baseline). The closed-loop pass is asymmetric: the surrogate slightly under-predicts γ in rollout, a conservative direction for control triggers but not yet symmetrically calibrated; tightening to a symmetric 0.9–1.1 window is on the near-term roadmap. Cross-machine inference is now N = 5 real-machine equilibria (DIII-D 145419, JET 79692, NSTX 132588, ITER 15MA, KSTAR 018451) from the SCOREC `Fusion_Public` archive; all five are correctly flagged out-of-distribution by the k-NN detector and all five predict a growing (positive γ) ideal-MHD mode — sign correctness 5 / 5. The two cases with BOUT++ ground truth (DIII-D, JET) are off by 2–14× in opposite directions. With N = 3 different supervised equilibria (DIII-D 145419, SPARC 1514, SPARC 1519; γ_BOUT spanning 8×) the γ-head learns to discriminate per-equilibrium — each supervised case converges within ~10 % of its own γ_BOUT, and Gate 3 actually improves to 10.9 % off PB3D. The binding constraint for cross-machine magnitude calibration is therefore the size and diversity of the supervised equilibrium set, not γ-head capacity. Full validation, OOD audit, and stated limitations live in the technical fact sheet.
03 · Position in the literature
Where plasmafast sits among existing tools.
ELITE and PB3D compute γ from first-principles ideal MHD, and the EPED workflow wraps ELITE for pedestal prediction — all at seconds-to-minutes timescales built for design-cycle iteration. KARHU is the closest prior ML surrogate: it is fast, but it regresses a single scalar — the maximum growth rate — from Europed-format profiles, with no out-of-distribution detection. EuroPED-NN predicts pedestal scalars rather than γ. plasmafast's position is the combination none of these hold at once: a full γ(n, ρpol) field, computed directly from a raw G-EQDSK equilibrium, in 71 ms, with a calibrated OOD detector. It is intended to be used alongside existing pedestal design oracles and RMP-ELM controllers, not in place of them.
| Work | Output | Inference | UQ / OOD |
|---|---|---|---|
| ELITEWilson et al., PoP 9, 1277, 2002 | γ scalar | ~tens of s | — |
| EPEDSnyder et al., NF 51, 103016, 2011 | pedestal height / width | ~min | — |
| PB3DWeyens et al., JCP, 2017 | γ scalar | ~min | — |
| KARHUBruncrona et al., PoP 32, 092501, 2025 | γmax scalar | ms-class | — |
| EuroPED-NNPanera Alvarez et al., PPCF 66, 095012, 2024 | pedestal scalars (not γ) | ms-class | Bayesian NN |
| Kim, Kolemen et al.Nat. Commun. 15, 2024 | RMP control scores (not γ) | ms-class | not stated |
| plasmafastthis work · 2026-05-22 | γ(n, ρpol) field | 71 ms | split-conformal CI · k-NN OOD |
04 · About
A solo research effort, building toward a real product.
dataquanta is a Delaware LLC building fast machine-learning surrogates for fusion plasma physics. Founded and led by Amin Vakhshouri.
05 · Contact
Open conversations.
For collaboration, technical questions, or program-officer enquiries.
- Funding INFUSE co-PI proposals
- SBIR DOE FES FY27 · C59-21a · C59-22c
- Milestone capability-enhancement subawards