Computing Library › Ml For Fusion
Ml For Fusion

Machine-Learned Surrogates for Neutronics

Surrogate models approximate expensive neutron-transport simulations, giving fast estimates of tritium breeding, heating, and shielding for design scans.

Why neutronics is costly

Fusion neutrons must be tracked through the blanket and structure to compute tritium breeding, nuclear heating, radiation damage, and shielding effectiveness. The reference method is Monte Carlo neutron transport, which is accurate but slow, needing many particle histories per configuration. Design optimization requires thousands of evaluations, so a surrogate is valuable.

Building the surrogate

Kronos motion — neutron transport

A surrogate is trained on a database of Monte Carlo runs spanning blanket composition, thickness, coolant, and multiplier fractions. It learns the map from these design parameters to outputs such as the tritium breeding ratio and peak heating, returning them instantly for new configurations.

Tritium breeding context

The tritium breeding ratio measures tritium produced per tritium consumed and must exceed one for a self-sufficient fuel cycle. Surrogates let designers explore which blanket choices raise it. The Hyperion breeder concept targets a breeding ratio of 1.8 in its blanket design studies, and a surrogate accelerates the many-parameter search toward such a target.

Validation discipline

Surrogate outputs are only trustworthy inside the sampled design space and inherit the assumptions of the training simulations, including nuclear data uncertainties. Promising configurations found with a surrogate are always re-run with full Monte Carlo transport before being adopted, and the surrogate never replaces the reference calculation for final numbers.

These are computational design tools. The breeding target and heating estimates are simulation results for a concept under design, not measured performance, since the machine is not yet built and construction is scheduled to begin in Q2 2027.