Machine-Learned Acceleration of Gyrokinetics
Learned models accelerate gyrokinetic turbulence simulation by replacing costly subcomponents or emulating full flux outputs.
Gyrokinetics and its cost
Gyrokinetic simulation resolves plasma turbulence by evolving the distribution function in a reduced five-dimensional phase space, averaging over the fast gyromotion. It is the most trusted tool for turbulent transport but is among the most expensive computations in the field, limiting how many cases can be run.
Acceleration strategies
- Emulate the input-output flux map so many cases skip full simulation
- Learn better initial conditions to shorten transient startup
- Replace expensive inner kernels, such as collision operators, with fast approximations
- Predict saturated states to reduce the time integrated
The most common route is emulation: train on a database of completed runs to predict fluxes directly, as in quasilinear neural models. This does not speed up any single simulation but removes the need to run most of them, which is the dominant cost in large scans.
In-the-loop acceleration
A subtler approach speeds up the simulation itself by learning components that are otherwise costly, or by providing warm starts that reduce the number of timesteps to reach a saturated turbulent state. These preserve fidelity while trimming runtime, at the cost of careful validation that the shortcut does not distort the physics.
Verification
Because gyrokinetics is the reference physics, any acceleration must be checked against unaccelerated runs on representative cases. Errors that look small in flux can matter for stiff transport, so validation focuses on the quantities that downstream models actually consume.
For building transport databases for concepts like the Hyperion breeder, acceleration makes broader parameter coverage feasible within a fixed compute budget. The accelerated results are computational estimates for a machine in design, anchored to full gyrokinetic runs at key points.