TANAGER
Follow on X
Tanager-M · 2026
A compact, high-field stellarator sized to what today's superconductors, steels and shields can really do. Then put through ten simulated tests, including two days nonstop, with every result shown, good or bad.
The reactor, reduced
The first Tanager was a 12 m power-plant concept. Tanager-M keeps its ideas and shrinks to the smallest machine that can still burn deuterium-tritium fuel.
Baseline uses the world-average stellarator confinement law (ISS04). Best case uses the best confinement Wendelstein 7-X has measured, 30 % above that law. Everything on this page is a simulation result, not a measurement.
Why this size
REBCO superconductor has been proven at 20 T in a large coil. The optimised coils see 2.32× the plasma's field at their hottest spot, so the plasma gets 8.2 T.
60 cm of tungsten carbide is what keeps neutrons out of the superconductor. Shrink the machine and that layer stays the same, so it eats the space the plasma needs.
A stellarator holds heat roughly as its radius to the power 2.3. Below about 5.1 m the plasma can't pay back its heating, even with the best confinement ever measured.
01 / 06
Deuterium and tritium at 80 million °C in the core. The shape is a published, precisely quasi-symmetric plasma, solved here with DESC at finite pressure with its own bootstrap current.
02 / 06
Tungsten armour on water-cooled reduced-activation steel, the same recipe planned for Europe's DEMO. No liquid metal: this version sticks to what has been built.
03 / 06
A double steel wall filled with borated steel plates and water, as in ITER's vessel. It holds the vacuum and does the first round of neutron shielding.
04 / 06
60 cm of tungsten carbide with water cooling. The optimised coils sit 1.1 m from the plasma, so the space in between became shielding. OpenMC traced 1,000,000 neutrons through it: the fast flux drops about 1.6 × 10⁵× before the superconductor.
05 / 06
Every coil shape on screen came out of a SIMSOPT optimisation: five unique coils, repeated four times, that recreate the plasma boundary to 0.12 % while leaving room for the shield.
06 / 06
Holds the coils near 20 K. The whole machine is a little wider than Wendelstein 7-X's 16 m cryostat and would fit in a large sports hall.
The test campaign
Each test uses open research code that fusion scientists use: DESC for the plasma, SIMSOPT for the coils, OpenMC with ENDF/B-VIII.0 nuclear data for the neutrons. The verdicts are mine.
DESC · 3-D MHD equilibrium at operating pressure
Re-solved at Tanager-M's real pressure (β 0.81 %) with its own bootstrap current. It keeps a magnetic well everywhere, but the Mercier stability number sits just below zero across the radius. Marginal: the next optimisation needs a stability target.
Biot–Savart, finite winding pack, regularised self-field
Peak field on the superconductor is 19.0 T, inside the 20 T that REBCO has shown in a large coil. Dropping the winding-pack current density to 100 A/mm² cut the peak by 12 % and bought the plasma 0.9 T. Forces reach 82 MN/m; a 10 cm steel case carries them.
SIMSOPT · coil optimisation + field-line tracing
Twenty coils recreate the plasma boundary to 0.12 % mean field error while staying at least 1.11 m from the plasma, which leaves room for the shield. Traced field lines form nested surfaces out to 90 % of the target radius; the last 10 % is still open.
0-D power balance, calibrated to Stellaris
Size scanned at fixed shape, heating capped at 40 MW, field capped by the magnets. With world-average confinement nothing under 7 m breaks even. With W7-X's best confinement, breakeven starts near R ≈ 5.1 m. Tanager-M sits right at that line.
Monte Carlo · 3,000 machines
Every uncertain input varied at once: confinement, impurities, radiation, profile shapes, density and pressure limits, fuel mix. The answer is a probability, not a promise, and it's low.
1-D time-dependent transport
Start-up to a steady burn in about 6 s. A 30 % density kick settles back by itself. Cut the heating and fusion falls below 1 % in 0.7 s; stop the fuel and it takes 1.9 s. The 1-D model lands 20–25 % above the 0-D numbers.
OpenMC · FLiBe blanket on the 12 m Tanager
A check on the original concept. Even with perfect coverage, 50 cm of FLiBe breeds 0.98 tritium per burned; 80 cm at 90 % lithium-6 reaches 1.03. A real plant needs about 1.1, so the blanket needs a neutron multiplier.
OpenMC · ENDF/B-VIII.0 · 1,000,000 histories
Neutrons and gammas traced through every layer. The coils sit 1.11 m out, so the space became 60 cm of tungsten carbide: it cuts the fast flux about 1.6 × 10⁵×. The superconductor would take ≈800 full-power years to reach its damage limit; the coils absorb about 152 W.
48-hour plant simulation · 400 runs with random faults
Every 20 ms for 48 hours: plasma, divertor, fuelling, pumps, tritium, cooling and magnets, with faults drawn at rates taken from today's hardware. On average it burns 96 % of the time and stops 1.6 times; 19 % of runs never stop. Tungsten flakes are the main plasma threat. The full 48 hours →
OpenMC depletion · ENDF/B-VIII.0 chain · shutdown dose
Two days of neutrons activate the steel and tungsten. At shutdown the structure gives off 243 kW of decay heat (1.9 % of fusion power) and 51 kW a day later, mostly from tungsten-187. Inside the vessel the dose is 351 Sv/h after a day and 657 mSv/h after a year, so maintenance has to be remote.
Code, inputs and outputs for every test are in the Tanager-M test kit. Re-run it and you get the same numbers.
Test 09 · endurance run
Switch it on and leave it for two days. Large fusion machines today hold plasma for minutes: France's WEST set the tokamak record of 22 minutes in 2025, on 2 MW of heating, and Japan's LHD stellarator ran about an hour in 2005. This run asks for 48 hours at 39 MW, with the hardware failing the way real hardware does.
Heating holds the stored energy, pellets hold the density, neon holds the divertor at 689 °C. About 6 pellets a second: 1.05 million over two days.
First wall and vessel settle in minutes. The 60 cm tungsten-carbide shield is so heavy it needs about 4 hours to reach its working temperature of 100 °C.
Six cryopumps take turns regenerating. About 1.0 kg of tritium passes through the plant, but only 4.1 g is burned. Pumps and processing hold up to 39 g at once; 0.4 g sticks in the walls.
Grid or power-supply trips (0.42 per run), control faults (0.47) and tungsten flakes (0.43). Anything heavier than 0.25 mg of tungsten, a speck 0.29 mm across, radiates the plasma away in under a second. Most restarts take 20 to 60 minutes. In 2 % of runs a long cryoplant outage forces the magnets down, which ends the run.
The plant draws about 4.5 GWh from the grid and makes 636 MWh of fusion energy: it is still a net consumer. At shutdown the structure gives off 243 kW of decay heat, and the vessel reads 351 Sv/h a day later.
0-D plasma around the design point, stepped every 20 ms, with OpenADAS radiation for tungsten, neon and helium. Fault rates are assumptions from today's hardware: halve them all and 51 % of runs never stop; double them and 4 % do. With W7-X-best confinement it makes 1,824 MWh if nothing breaks, but flakes stop it more often (0.8 per run): its denser plasma radiates more tungsten power and its heating is already at the 40 MW limit.
Operating points
0-D power balance. Both cases keep density under the stellarator limit, pressure under 3 % of magnetic pressure and heating at 40 MW or less. The 1-D burn simulation gives Q 0.43–1.17.
Against standard reactors
Swipe the table →
| Tanager-M | AP1000 PWR | NuScale module | eVinci micro | SPARC | ITER | |
|---|---|---|---|---|---|---|
| Type | fusion · stellarator | fission | fission · SMR | fission · micro | fusion · tokamak | fusion · tokamak |
| Status | design, simulated | operating (Vogtle 3 & 4) | NRC standard design approval (2025) | zero-power criticality tests | assembly, first plasma 2027 | construction, D-T 2039 |
| Power | 13–38 MW fusion (steady) | 3,415 MWt / ~1,117 MWe | 250 MWt / 77 MWe | 15 MWt / 5 MWe | 140 MW fusion, 10 s pulses | 500 MW fusion, 400–600 s |
| Gain / net power | Q 0.34–0.97 · net electric: no | — · net electric: yes | — · net electric: yes | — · net electric: yes | Q ≈ 11 predicted (>2 minimum) · net electric: no | Q = 10 target · net electric: no |
| Fuel in the machine | ≈37 mg D-T in the plasma | tens of tonnes of uranium | LEU core | TRISO, 19.75 % HALEU, 8+ yr core | — | — |
| Runaway risk | impossible: burn stops in ~1 s without fuel or heating | prevented by design; needs decay-heat cooling | passive safety | passive heat pipes | disruptions possible (tokamak) | disruptions possible (tokamak) |
| After shutdown | ≲1 % of fusion power (activation) | ≈6.5 % at shutdown (~220 MW) | ≈6.5 % at shutdown (~16 MW) | ≈6.5 % at shutdown (~1 MW) | small | small |
| Waste | activated steel & tungsten, no spent fuel | spent fuel, long-lived actinides | spent fuel | spent TRISO fuel | activated structure | activated structure |
| Weapons material | no fissile material | plutonium in spent fuel | plutonium in spent fuel | HALEU | — | — |
Feasibility
Mostly. Every system below has been built somewhere at a scale close to what Tanager-M needs. The twisted high-field coils are the one real gap.
| System | Tanager-M needs | Exists today | Status |
|---|---|---|---|
| HTS magnetsPartly | 19.0 T peak, 20 twisted coils, ≈14,600 km of REBCO tape | 20 T large-bore REBCO coil (MIT/CFS, 2021); SPARC's 18 flat coils use ≈10,000 km of tape — Twisted stellarator HTS coils not yet built; Proxima's model coil targets 2027 | Partly |
| Plasma heatingExists | 39 MW ion-cyclotron heating | JET ran >30 MW ICRH; SPARC plans 25 MW at 12 T | Exists |
| FuellingExists | Frozen D-T pellets, steady | W7-X: about 90 pellets over a 43 s record plasma (2025) | Exists |
| First wall and divertorPartly | ≈ 6 MW/m² with 70 % radiated in the divertor | W7-X island divertor, tungsten components rated ≈10 MW/m² — Needs detached operation; attached it would be 20 MW/m² | Partly |
| Neutron shieldExists | 60 cm tungsten carbide + water | WC shielding is industrial; studied for compact fusion — Magnets outlast the machine (≈800 full-power years) | Exists |
| TritiumExists | 2.0 g per full-power day, bought in | JET's D-T campaigns (1997, 2021, 2023); Canada's CANDU reactors supply tritium — No breeding blanket on this machine | Exists |
| Remote maintenancePartly | In-vessel handling after D-T | JET remote handling, ITER designs — Stellarator access through ports is harder than a tokamak | Partly |
A machine like this is a national-lab or well-funded-company project, in the same class as Wendelstein 7-X (about €1.1 billion). It is not a garage build.
Lineage
Wendelstein 7-X proved optimised stellarators hold heat. SPARC shows what 20 T superconductors buy. Tanager-M is the stellarator version of that step, and the 12 m Tanager power plant is what it would lead to.
Swipe the table →
| Wendelstein 7-X | SPARC | Tanager-M | Tanager (plant) | |
|---|---|---|---|---|
| Type | Stellarator | Tokamak | Stellarator | Stellarator |
| Status | Operating since 2015 | First plasma 2027 | Simulated, 2026 | Concept |
| Major radius | 5.5 m | 1.85 m | 4.8 m | 12 m |
| Plasma volume | 30 m³ | 20 m³ | 61 m³ | 433 m³ |
| Field on axis | up to 3 T | 12.2 T | 8.2 T | 10 T |
| Fuel | H, D | D-T, 10 s pulses | D-T, steady | D-T, bred |
| Fusion gain | — | Q ≈ 11 predicted | Q 0.34–0.97 | ignition in 0-D; blanket short of breeding |
What I'm exploring next
No lab, no company, no deadlines. Just the next questions I want answered, in the order I'm taking them.
A 12 m power-plant idea, checked against Proxima Fusion's published Stellaris numbers.
DESC, SIMSOPT and OpenMC on a machine sized to today's materials. This page.
Optimise both at once, with the plasma's own current included, to close the last 10 % of the edge.
Let my self-improving research system, the RSI Runtime Module, search for a smaller machine with the same tests.
Publish every input and ask people who work in fusion what I'm getting wrong.
About
I'm Junai P. Felix, an independent AI researcher. I'm not a nuclear engineer, and Tanager isn't a company. It's a personal exploration of new power sources and the efficiency gains that could make them practical.
My main work is self-improving AI systems. Tanager is where I point that work at a hard physical problem: start from the best published fusion research, test it with the tools the field uses, and report what comes out.
Every figure here is labelled for what it is. If you work in fusion and see a mistake, I want to hear it.
Renders
