Machine-Learned Radiation-Transport Surrogates
Learned surrogates approximate particle- and radiation-transport codes, speeding shielding design and dose estimation across many configurations.
The transport bottleneck
Estimating how neutrons, gamma rays, and other radiation propagate through materials underlies shielding design and dose assessment. High-fidelity transport codes solve the Boltzmann transport equation by deterministic or Monte Carlo methods, both expensive when many geometries or source terms must be evaluated.
Surrogate strategy
A surrogate learns the mapping from source and geometry parameters to transport outputs such as flux spectra, heating, and dose at points of interest. Trained on a suite of reference solutions, it evaluates new cases in milliseconds, enabling optimization loops that would be infeasible with the full solver.
- Predict integrated quantities (dose, heating) or full spatial fields
- Enforce non-negativity and physically bounded outputs
- Use uncertainty estimates to trigger fallback to the reference code
- Retrain as the design space or nuclear data updates
Field prediction
Beyond scalar outputs, convolutional and operator-learning architectures can predict entire spatial flux maps, learning the smooth spatial structure of transport solutions. This is useful for spotting hot spots in shielding without a full solve, though sharp gradients near sources remain difficult.
Discipline
Radiation-transport surrogates support screening, not certification. Because errors in dose or damage estimates carry safety weight, final shielding numbers come from validated reference codes. The surrogate narrows the design space quickly; the reference solver confirms the chosen configuration.
In Kronos design studies these surrogates support shielding and activation analysis for concepts still on paper. Every output is a computational estimate for an unbuilt machine, cross-checked against first-principles transport before informing any safety-relevant decision.