AI summary by TwinMind
First Nuclear Plasma Control with Digital Twin
Sabine Hossenfelder
7m 11s
Nuclear Fusion Background
- Nuclear fusion great idea in principle - could solve world energy worries
- Problem: takes more energy than it creates
- Japan/US team achieved progress using digital twin to control nuclear plasma in stellarator
Fusion Reactor Types
- Tokamaks: Traditional approach
- Donut-shaped device with magnetic fields
- Hot nuclear fuel held/heated until nuclei fuse
- Existing tokamaks huge (ITER: 70m high building, ~20 floors)
- Stellarators: Alternative approach
- Still donut-shaped but more complicated magnetic field
- Allows smaller device, harder for plasma to escape
- Increases efficiency but more difficult to build/model
- Largest existing: Wendelstein 7-X in Greifswald, Germany
- Inertial Confinement: Different approach
- Shoot at fuel pellet hoping it explodes
- Methods: lasers (National Ignition Facility, US) or bullets (First Light Fusion's "Big Friendly Gun")
- Other startup ideas: Z-pinch and various approaches
Plasma Control Challenges
- Major issue: difficult to control plasma in tokamaks/stellarators
- Motion is chaotic, hard to steer
- Uncontrolled plasma can damage reactor vessel
- Must quickly abort heating/fusing when plasma becomes unstable
- Biggest problem for energy efficiency
Previous AI Control Work
- 2019 Harvard/Princeton research:
- Trained AI on Joint European Torus (largest tokamak) data
- AI recognized patterns signaling plasma instability
- Hindsight analysis: 80%+ accuracy 1 second ahead, almost all instabilities detected 30 milliseconds ahead
- DeepMind 2023 work:
- Real-time plasma control in TCV tokamak (Swiss Plasma Center, Lausanne)
- Small reactor: ~2m each direction
- AI trained on tokamak simulator first, then controlled real device
- Successfully coaxed plasma into various shapes (triangular, two separate droplets)
New Research - Digital Twin Method
- Team: Japan/US researchers at Large Helical Device (Japan)
- Device: Stellarator, second largest in world after Wendelstein 7-X
- Method: Digital twin (virtual replica of plasma on computer)
- Uses real-time observations to update twin
- Predicts what happens next to control plasma
- Process called "data assimilation"
- More physics built-in than AI-based approach
- Ensemble forecasting: Assumes uncertainty in current state, makes predictions for set of initial values, calculates most likely outcome (same method used for weather forecasting)
- Results: Successfully controlled temperature in center of plasma, reached target
Current Limitations
- Neither this experiment nor DeepMind's controlled actual ongoing fusion process
- Only steered plasma without fusion
- Next step: actively controlling fusion process
