Machine Learning for Divertor Heat-Flux Prediction
Models predict and monitor the intense, narrow heat-flux loads on the divertor, informing detachment control and component protection.
The exhaust challenge
Power leaving the plasma is funneled to the divertor, where it deposits over a narrow footprint. The peak heat flux there can approach the limits of any known material, so predicting and controlling it is one of the hardest engineering problems for a power-producing device.
Prediction targets
- Peak heat flux and the width of the deposition footprint
- Degree of detachment, where edge radiation spreads the load
- Location and motion of the strike point
- Onset of conditions that threaten component limits
Models learn these from upstream plasma conditions, divertor diagnostics, and magnetic geometry. Because the footprint width depends on turbulent and neutral processes that are hard to compute, data-driven predictors complement first-principles edge codes that are too slow for control.
Detachment control
The main mitigation is detachment, in which impurity radiation and neutral interactions cool the exhaust before it reaches the surface, spreading and reducing the peak load. Learned models estimate proximity to and stability of detachment so a controller can hold the plasma in this protective regime with the right impurity seeding.
Monitoring
Infrared thermography combined with convolutional models tracks the actual surface heat load in real time, flagging hot spots that exceed limits. This closes the loop between prediction and observation, letting protective action trigger when models and measurements disagree with expectations.
Exhaust handling is central to any breeder concept, including Hyperion. In Kronos design work these predictors are developed against simulated edge and divertor conditions ahead of construction; they act on modeled loads because the hardware is not yet built, and their estimates are computational.