Accelerating Simulation with ML
Using learned models to speed up, replace steps in, or guide expensive plasma and reactor simulations.
Why simulations are slow
Fusion simulation spans turbulence, transport, MHD stability, heating, and neutronics, often coupled and resolved over fine grids and short time steps. Full-fidelity runs can take hours to days. ML offers several ways to make these tractable without abandoning physics.
Modes of acceleration
- Surrogates that replace an expensive submodel with a fast learned map
- Emulators that predict the full simulation output from inputs
- Learned closures or corrections inside a coarse solver
- Neural preconditioners and solver accelerators
Replacing submodels
The most common and reliable pattern replaces the slowest component, such as the turbulence model, with a surrogate, leaving the rest of the physics solver intact. This preserves the trusted structure while removing the bottleneck, and the surrogate can be validated against the component it replaced.
Coarse-plus-correction
Another pattern runs a cheap, low-resolution solver and adds a learned correction toward high-resolution results. This can recover much of the accuracy at a fraction of the cost, provided the correction is trained on representative cases and its error is monitored.
The trust boundary
Every accelerated simulation inherits the limits of its learned parts: they are reliable only where trained. Acceleration is legitimate when the fast result is checked against full-fidelity runs on a validation set and used within the validated envelope. It is misleading when speed is claimed without stating that trust boundary. Speedups are means to more exploration, not license to trust extrapolation.