WEST’s interior, which broke a fusion record for a tungsten tokamak. CEA-IRFM.

WEST Fusion Reactor Longest Run: A Case Study Analysis

A 1,337-second plasma on the WEST tokamak is the strongest test case yet for the Lawson Stability Manifold. Read with computation first, it shows that duration is itself a control resource, and that is a lever fusion can use to compete.

On 12 February 2025, the WEST tokamak at CEA Cadarache in France held a hydrogen plasma for 1,337 seconds. That is 22 minutes, with 2.6 gigajoules injected and extracted through actively cooled tungsten walls, at zero loop voltage the whole way. It is the longest tokamak plasma on record, and it sits inside the 400 to 3,000 second pulse window ITER is designed for.

Duration is the property every fusion power plant needs, and the one the classic scorecard does not measure. The Lawson criterion asks how hot, how dense and how well confined a plasma is. It does not ask whether you can hold it there for an operating shift.

This article reads WEST’s record through the Lawson Stability Manifold (LSM), a framework I developed in three open preprints, and speaks to two audiences. For engineering physicists, the numbers: what a pre-registered computation says about holding a plasma for 22 minutes. For energy-sector participants, the meaning: what a computation-first approach offers fusion as it moves toward the grid.

The headline result fits in one line. On the model host, settling onto the controlled state costs a fixed product of gain and time, about 5.5, so a 22-minute pulse lets a controller work with 30 to 1,000 times less gain than the design values, depending on which plasma timescale the control acts on.

Why duration is the question a power plant asks

A power plant earns its keep in megawatt-hours, so its value is set by how long and how reliably it runs, not by its best second. Every restart costs time, and every thermal cycle costs wall and blanket lifetime.

Early 2025 brought two tokamak pulses past 1,000 seconds. On 20 January, China’s EAST held a high-confinement plasma for 1,066 seconds. Three weeks later WEST held its plasma for 1,337 seconds and removed 2.6 GJ through tungsten components that a power plant would also need. CEA’s stated next step is to raise WEST’s heating from 2 MW to 10 MW, still over pulses of about 1,000 seconds.

That moves the central question for the sector. It is less “can we reach fusion conditions?” and more “can we hold them, cheaply and reliably, for hours?” Holding is a control problem, and control is where computation pays back fastest.

The lens: the Lawson Stability Manifold

The Lawson criterion scores an operating point. The Lawson Stability Manifold is the design space for holding one, with the Lawson criterion kept as the exit condition. Three preprints build it:

• Host invariance (Zenodo 10.5281/zenodo.23204696) sets the acceptance rule: a control claim counts only if it holds across a library of model hosts, tested against a pre-registered null model.
• The Nosé–Hoover thermostat and the Hala quench operator (Zenodo 10.5281/zenodo.23282632) maps the design space and shows that the control gain sets the rate of approach to a controlled cycle.
• Covariant electron-cyclotron deposition (Zenodo 10.5281/zenodo.23285857) carries the loop into a tokamak: a log-frequency coordinate for where microwave heating lands, and a single-probe diagnostic for the controlled oscillation.

WEST’s record maps onto this loop part for part. WEST held its current at 0.22 MA by adjusting lower-hybrid power, while the central solenoid kept the loop voltage at zero. That is an integral regulator whose actuator efficiency moved with density, confinement and MHD regime. The model host has the same anatomy:

• The thermostat integrates an energy error, as WEST’s current loop integrates its current error.
• The friction term is actuator effort, the lower-hybrid power request.
• The Hala term is a state-dependent leak that steers the loop onto a controlled cycle.
• Actuator efficiency g, seven klystrons on WEST, enters the host only as an effective strength κ_eff = g κ.

For readers who want the equations, the host is three lines, with the Hala leak L = γ̄ R² (z − c):

dx/dt = y
dy/dt = −x − g z y
dz/dt = κ (y² − 1) − L

Computation first: what the numbers say about a 22-minute plasma

The computation pushed the model host into WEST’s regime, with every claim, comparison and pass mark written down and hash-locked before the first run. Six results stand out. Each is tagged pre-registered, or exploratory when it was found after the registered results were in.

1. Duration buys control: gain × settling time ≈ 5.5

Pre-registered. The time to settle onto the controlled cycle scales as one over the gain. The product γ̄·Ts stays at 5.5 for κ = 1 and 4.8 for κ = 0.5, flat from a gain of 0.0005 to 0.05, with log–log slopes of −1.001 and −1.010.

The gain buys speed, not a different endpoint. Settling has a fixed price in gain × time, and a long pulse pays it with time.

Gain × median settling time against gain, 32 initial conditions per point. Filled: Hala leak. Open: equal-power linear leak. Orange: exploratory κ = 2 scan.

The same law explains the onset reported at a gain of 0.05 in earlier short runs. The 90th percentile of γ̄·Ts is 7.04, so an integration of about 141 time units shows exactly that apparent threshold.

At κ = 2 an exploratory scan found something new: a mirror-image pair of cycles that merges near γ̄ ≈ 0.028, with critical slowing at the merger. In that regime the gain selects which state the loop settles into, as well as how fast.

2. Standard actuators can implement it

Pre-registered. A plain linear leak, matched to the Hala leak in actuator power, lands on the same curve (γ̄·Ts of 5.6 to 5.7 at κ = 1) and reaches the same cycle (amplitude 1.254 in both). The settling law belongs to the whole class of leaks on an integrator, so it needs no exotic hardware.

3. Losing one klystron in seven is absorbed by the actuator

Pre-registered. Cutting actuator efficiency from 1 to 6/7, WEST’s klystron fraction, moves the controlled oscillation by only 2.1%. The delivered actuation drops 14.7%: the actuator side carries the loss, as κ_eff = g κ requires, and re-settling scales as one over the gain (slope −1.08).

WEST did the same thing twice, after klystron losses near 75 s and 1,040 s. Each time the plasma moved into a new regime and the pulse carried on.

4. Slow drift is followed without extra authority

Exploratory scan. With the thermostat strength drifting as κ(t) = 1 + 0.3 sin(2πt/Td), the tracking error falls below 10% once the drift is slower than about 4.5 oscillation periods, and below 1% beyond about 35. That holds at two gains sixteen-fold apart: the loop rides along its family of cycles. The registered runs agree, with error at or below 0.02 at every point.

Tracking error against drift period. Filled: return-map value (exploratory). Open: registered runs. Dashed: registered first-order-lag prediction.

5. Everything that shaped the record ran three decades slower than the plasma

From the WEST record. The plasma’s own timescales, the 0.8 kHz mode, the 32–50 ms confinement time and about 1.5 s of current diffusion, all sit below two seconds. Every event that shaped the pulse sits between 75 and 1,337 seconds: the klystron losses, the MHD regime changes, the radiated-power jump and the end of the pulse.

Timescales and events of WEST #61299 and its sister pulses on one log axis. Sources: CEA-IRFM; Ekedahl et al., EPS 2025; Dumont et al., IAEA FEC 2025.

That separation is good news for control. Mapping host time onto WEST gives the gain needed to settle 90% of starting points within the first 10% of a 1,337 s pulse:

• Host oscillation = the 0.8 kHz MHD mode: gain 1.0×10⁻⁵
• Host oscillation = the confinement time (about 40 ms): gain 3.4×10⁻⁴
• Host oscillation = current diffusion (about 1.5 s): gain 1.3×10⁻²

On the first two maps, a 22-minute pulse needs gains 30 to 1,000 times below the design gains. The mapping is illustrative until it is calibrated to a measured WEST signal, which is one of the next tests.

6. WEST’s place on the Lawson map is the duration axis

Pre-registered. The record spans 27,000 to 42,000 energy confinement times and 670 to 1,340 current-diffusion times. Its triple product, 1.5–5.1×1⁰¹⁸ keV s m⁻³ in hydrogen at 0.22 MA, sits about three orders of magnitude below the ignition benchmark. WEST was built to master duration and heat exhaust; burning-plasma machines such as ITER carry the triple product.

The LSM keeps those two axes apart: Lawson is the exit condition, and holding is the design space. WEST’s record is the strongest data yet on the holding axis, and it argues for scoring designs on pulse length per confinement time, or on injected energy, beside the triple product.

A forecast WEST can check: where its 105 GHz microwaves will land

The wave layer gives WEST a number to test. At the record’s central parameters, 105 GHz O-mode power lands 7 to 11 mm on the high-field side of the cold resonance for central temperatures from 2.0 to 4.3 keV.

WEST commissioned a 105 GHz electron-cyclotron system at 1 MW in 2025, with 3 MW expected in 2026. The record pulse used none, so this is a forecast at the record’s own parameters: a line density of 2.3×1⁰¹⁹ m⁻³ and a field of 3.7 T.

Get Ahmed M. Hala’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

The forecast uses the log-frequency coordinate u = ln(ω/ωce) = ln(R/Rres). In it, harmonics sit at fixed offsets, relativistic and Doppler shifts simply add, and the machine enters the optical depth through a single number, B0R0, which is 9.25 T m for WEST.

Pre-registered forward model. It reproduces the preprint’s table at 2 keV and extends it across the temperature range:

• Te = 2.0 keV: optical depth 5.1; absorption line 14.4 mm, deposition peak 6.8 mm on the high-field side
• Te = 3.4 keV: optical depth 8.7; line 24.4 mm, deposition 9.4 mm
• Te = 4.3 keV: optical depth 11.0; line 30.7 mm, deposition 10.8 mm

Deposited power (solid) and absorption line (dashed) at 105 GHz, O-mode fundamental, central parameters of #61299.

The plasma is optically thick, so the deposited power stays close to the resonance while the relativistic line moves 14 to 31 mm. The deposition shift scales as Te to the power 0.61, inside the 0.605–0.680 range the preprint predicted.

With the beam steered onto off-midplane flux surfaces, the temperature-driven shift is 0.9 to 2.9 mm in flux-surface distance, an order of magnitude inside the loop’s alignment window. Alignment is therefore set by whether the target q ≈ 4 surface can be reached, which field, axis position and launcher steering decide. Inside the control loop, aligned deposition enters only as κ_eff = g κ.

The experimental test is direct: modulated O-mode heating across discharges, with the pass mark fixed in advance: a deposition-shift slope of 0.63 (0.60 to 0.68).

What the record says in its own words

Read as a story, the record has one core theme: bounded authority over a drifting plant. A loop kept re-balancing as the plasma and wall drifted and as the actuator gained or lost efficiency, all inside limits set by hardware.

That theme comes from a grounded-theory reading, built by constant comparison across the CEA announcement, two 2025 conference papers by the WEST team and the three preprints. Eight categories emerged:

1. Integral regulation through a drifting actuator. Current held by lower-hybrid power at zero loop voltage. Matches the thermostat-plus-leak loop.
2. Coherent oscillation. A mild 0.8 kHz mode with periodic temperature relaxations throughout. Matches the controlled-cycle endpoint.
3. Regime switching on actuator events. New regimes after each klystron loss. Partly captured by coexisting cycles.
4. Slow open-system drift. Outgassing and wall conditioning. The framework’s next layer.
5. Technology bounds on authority. RF cooling loops, wall protection, seven klystrons. The next layer.
6. Termination outside the plasma. #61299 ended when the power request went to zero. The next layer.
7. Ensembles from plateaus and sister pulses. Matches host invariance in spirit.
8. Distance from Lawson. Reframed as the duration axis.

The three categories that describe the loop match the preprints. The three that describe the plant’s surroundings, drift, hardware bounds and termination, set the agenda for the next layer of the model.

WEST’s own team already works this way. Before the campaign, they scanned current, density and lower-hybrid power with integrated modelling, mapped the predicted loop voltage across that box and trimmed it by technology limits. That is the design-space idea of the LSM, carried by a model calibrated to the machine: computation first, then the shot.

What this offers the energy sector

Fusion will compete on availability and cost per megawatt-hour. The WEST case study points to six ways a computation-first approach moves both.

1. Time is a resource the plant already has. On the model host, settling costs a fixed gain × time, so a plant that runs for hours can regulate gently. Gentler control leaves more margin in the same actuators.
2. Redundancy keeps the plant running. Losing one of seven heating sources moved the controlled state by 2.1% on the host, and WEST rode through two such losses. Actuator banks designed to absorb a single failure are one route to the availability utilities expect.
3. Ordinary hardware is enough. The settling law holds for any equal-power leak on an integrator, so standard actuators and standard control engineering can carry it.
4. Forecast before you fire. The 105 GHz deposition forecast was computed at the record’s parameters with its pass mark already fixed, ready for the first modulated-heating shots. Questions settled on a workstation free experimental time for the questions only a machine can answer.
5. Evidence an investor or regulator can audit. Pre-registration locks the claims, comparisons and pass marks with a cryptographic hash before any run. Anyone can verify that the results were not tuned after the fact, much as an audit trail does in financial reporting.
6. A progress metric anyone can follow. Beside the triple product, track pulse length in confinement times, or injected energy. WEST’s 27,000 to 42,000 confinement times and 2.6 GJ are numbers a non-specialist can follow from one year to the next.

Computation first does not replace experiments; it makes each one count. Computation chose what to look for in a 22-minute pulse, and the next pulse can be designed to test it.

How to read the results

Every result above carries one of three tags, and each tag answers the question “when was the pass mark set?”

• Pre-registered: the claim, the comparison and the pass mark were written down and hash-locked with SHA-256 before any run. Nine items were registered this way, and the numbers reported are what the frozen code produced.
• Exploratory: found after the registered results were in, and labelled as such. The κ = 2 cycle merger and the drift-tracking threshold are the next items to pre-register.
• From the record: published WEST facts, cited below.

The control results come from model hosts, and the next tests tie them to WEST signals. The registration, freeze record, scripts, results, figures and a SHA-256 manifest accompany the case-study paper, so every number here can be rerun with open Python tools.

The derivations, code, runs and figures were prepared with Claude (Anthropic) under my direction. These results are part of my APS-DPP 2026 talk, “Integrated Active Control of Fusion Plasmas via the Lawson Stability Manifold and Coordinate-Free Symplectic Phase Space.”

What comes next

Three steps turn this design case into control on WEST:

1. Modulated O-mode heating on WEST tests the forecast deposition slope of 0.63 across discharges.
2. Actuator saturation, rate limits and command loss in the host find the authority margin a long pulse needs, and confirm that a one-in-seven loss is survived above it.
3. A calibrated time map ties host time to a measured WEST signal, such as the relaxation period or the power-request oscillation.

Twenty-two minutes is a milestone worth celebrating. WEST showed that a tokamak can be held at ITER pulse lengths through tungsten walls a power plant would also use. Read through the Lawson Stability Manifold, the record also shows why duration matters for control: on long pulses, time does work that would otherwise take actuator authority. Fusion will compete when holding a plasma is routine, and computation is the fastest way to get there.

Sources

• CEA-IRFM, “WEST sets a new plasma duration record”, February 2025.
• A. Ekedahl et al., “Long pulse achievements in the WEST tokamak,” 51st EPS Conference on Plasma Physics (2025), O2.035.
• R. J. Dumont et al., “WEST long-pulse achievements in support of next-step fusion devices”, 30th IAEA Fusion Energy Conference (2025); Nuclear Fusion (2026).
• Chinese Academy of Sciences, EAST’s 1,066-second high-confinement plasma tops China’s 2025 science advances, February 2026.
• A. M. Hala, “Host-Invariance as the Acceptance Criterion for Plasma Chaos Control”, Zenodo (2026).
• A. M. Hala, “Action of a Nosé–Hoover Thermostat and the Hala Quench Operator on the Lawson Stability Manifold”, Zenodo (2026).
• A. M. Hala, “Covariant Electron-Cyclotron Deposition and Single-Probe Lyapunov Diagnostics for an Alignment-Coupled Nosé–Hoover and Hala Control Loop in the WEST Tokamak”, Zenodo (2026).
• J. D. Lawson, “Some criteria for a power producing thermonuclear reactor”, Proc. Phys. Soc. B 70, 6–10 (1957).
• B. G. Glaser and A. L. Strauss, The Discovery of Grounded Theory (1967); K. Charmaz, Constructing Grounded Theory, 2nd ed. (2014).