Computing Library › Ml For Fusion
Ml For Fusion

Surrogates for Neutronics

Learned models that approximate expensive neutron-transport calculations for shielding, activation, and breeding studies.

Why neutronics is costly

Fusion neutrons must be transported through complex geometry to compute shielding, material activation, heating, and, in a breeder, tritium production. Monte Carlo neutron transport is accurate but slow, and design studies need many evaluations across varying geometry and materials. Surrogates make repeated evaluation feasible.

What is approximated

Kronos motion — neutron transport

Building the surrogate

Training data come from Monte Carlo transport runs across a sampled design space. Because each run is expensive, active learning and space-filling designs are used to cover the space efficiently. The surrogate then predicts the quantities of interest fast enough for optimization and sensitivity studies.

Physical constraints and validation

Fluxes and breeding ratios are non-negative and vary smoothly with geometry, constraints that can be built in. Every surrogate is validated against held-out transport runs, with error reported per quantity, because a good average can hide a poor estimate exactly where a design margin is tight.

Design context

Neutronics surrogates support tritium-breeding and shielding studies for D-T concepts. For the Kronos breeder design, such studies address the tritium breeding ratio (design target 1.8) and magnet shielding within a simulation framework, always cross-checked against full transport calculations. The burner, a D-3He concept, has a much lower neutron fraction (5.44 percent), which changes but does not remove the need for neutronics analysis. All figures are design targets from simulation, not measurements.