Swiss Plasma Center at EPFL and Google DeepMind
Verified evidenceAutonomous + backstopOperators set plasma objectives while a simulator-trained policy coordinates all coils on physical hardware.
Engineering design · Lausanne, Switzerland
Collections: Embodied work · Negative results
Executive brief
AI can move humans up from hand-crafting every control law to specifying objectives, but only when simulation, constraints, and machine protection form a hard envelope.
AI value · Plasma current and shape tracking error
VerifiedAcross the full 0.1-1.0 second interval, RMSE was 0.62 kA for plasma current and 0.75 cm for shape.
Research-machine demonstration, not commercial fusion-power operation.
Before
Designs a controller for a predefined plasma state and configuration. → Runs the engineered controller during an experimental discharge.
After
Specifies plasma shape, current, and location objectives. → Commands all magnetic coils at high frequency to track the requested plasma configuration.
Human boundary
The policy controls coils only within the approved discharge; researchers authorize experiments, objectives, constraints, and shutdowns.
Why it matters
Operators set plasma objectives while a simulator-trained policy coordinates all coils on physical hardware.
Before
How the work ran before the change.
Control engineer
Designs a controller for a predefined plasma state and configuration.
ControlPhysics models, engineering constraints, and machine protection.
Tokamak operator
Runs the engineered controller during an experimental discharge.
ControlOperational safety limits.
What changed
Operators set plasma objectives while a simulator-trained policy coordinates all coils on physical hardware.
Decision rightAI acts within a human backstop
After
How the same work runs now.
Research operator
Specifies plasma shape, current, and location objectives.
ControlSimulator training plus physical and operational constraints.
RL controller
Commands all magnetic coils at high frequency to track the requested plasma configuration.
ControlTCV machine protection and operator-defined objectives.
Exception path
Machine-protection systems and operators terminate or override a discharge when limits are breached.
Work removed
Decision authority
The policy controls coils only within the approved discharge; researchers authorize experiments, objectives, constraints, and shutdowns.
Before
Control engineer
Designs a controller for a predefined plasma state and configuration.
Control: Physics models, engineering constraints, and machine protection.
Tokamak operator
Runs the engineered controller during an experimental discharge.
Control: Operational safety limits.
After
Research operator
Specifies plasma shape, current, and location objectives.
Control: Simulator training plus physical and operational constraints.
RL controller
Commands all magnetic coils at high frequency to track the requested plasma configuration.
Control: TCV machine protection and operator-defined objectives.
Work that left the path
Human role before
Engineers designed configuration-specific feedback controllers and operators executed them.
Human role after
Humans define objectives, validate policies in simulation and hardware experiments, and retain experimental and safety authority.
AI roleDecision mode: closed-loop control within a protected experiment. The learned policy sets coil commands to shape and stabilize plasma.
A novel learned controller had not previously demonstrated direct full-coil magnetic control on tokamak hardware.→Across the full 0.1-1.0 second interval, RMSE was 0.62 kA for plasma current and 0.75 cm for shape.
Experimental discharges reported in 2022. · Physical TCV tokamak controlling 19 magnetic coils.
Research-machine demonstration, not commercial fusion-power operation.
Engineering-driven control of pre-designed states.→The controller produced conventional, elongated, negative-triangularity, snowflake, and sustained two-droplet configurations.
Reported experimental campaign. · Multiple real-world TCV configurations.
Demonstrated configurations do not establish generalization to other tokamaks.
Anti-pattern
Treating a research demonstration as evidence of autonomous commercial fusion operation.
Questions
Portability conditions
Reputation risk
medium
Evidence and authority
Current · updated
1 peer reviewed; publication outcomes are verified.
Bundle 1.0.0 · reviewed 2026-08-23 · stable ID 72fcb19d466f7070
Google · Autonomous + backstop
That AI should only recommend cooling set-point changes for operators to implement manually.
Lawrence Berkeley National Laboratory and University of California, Berkeley · Autonomous + backstop
That scientists must manually choose every recipe, execute every synthesis step, interpret each XRD result, and plan the next experiment.
National Grid Electricity Transmission · Autonomous + backstop
Central BVLOS drones/AI automate routine pylon imaging and corrosion grading; engineers decide maintenance.
MIT randomized professional-writing study · Autonomous + backstop
The redesign is a shorter drafting loop with human revision and accountability, not autonomous publication.