Page 546 - AI for Good Innovate for Impact
P. 546
AI for Good Innovate for Impact
Use Case 13: AI Control System for Fusion Energy
Organization: Next Step Fusion
Country: Luxembourg
Contact Person(s):
Aleksei Zolotarev, info@ nextfusion .org
Georgy Subbotin, gs@ nextfusion .org
Olga Spiridonova, os@ nextfusion .org
1 Use Case Summary Table
Item Details
Category Manufacturing
Problem Developing robust and reliable plasma control systems is one of the chal-
Addressed lenges in the transition from experimental fusion devices to fusion power
plants. Plasma control is crucial for maintaining the high temperatures and
magnetic confinement necessary for sustained fusion reactions. An effec-
tive plasma control system must coordinate numerous magnetic coils and
adjust their voltages thousands of times per second to keep the plasma
stable - tasks current systems can't reliably manage, highlighting the need
for AI-powered solutions.
Key Aspects of • Next Step Fusion Simulator (NSFsim) - our in-house physics-based toka-
Solution mak simulator used to model plasma behavior and train RL agents safely
and efficiently. More information on our NSFsim can be found at [1]
• Machine Learning model that performs real-time control of plasma shape
and position by processing raw magnetic diagnostics data and output-
ting commands to adjust currents in the tokamak's magnetic coils.
• Reinforcement Learning algorithms to train ML models for real-time
plasma parameter control, as well as for training an ML model for plasma
shape reconstruction.
Technology Machine Learning, Reinforcement Learning, Plasma Control System, Plasma
Keywords Simulation
Data Availability Private (Data can be reached through the DIII-D national user facility
program. More information about joining the program can be found at [2])
Metadata (Type • Time-series experimental data from fusion devices, including magnetic
of Data) sensor measurements (e.g., magnetic probes, flux loops)
• Control command for power supply (converted to voltage commands by
power supply hardware)
• Additional synthetic data is generated using our NSFsim tokamak simu-
lator to supplement real-world data and accelerate model development.
510
