Monte Carlo Methods and ML
How random sampling underpins neutron transport and uncertainty studies, and where learning accelerates it.
The Monte Carlo idea
Monte Carlo methods estimate quantities by random sampling. In fusion, they track many simulated particles through geometry and physics, averaging outcomes to estimate fluxes, breeding, and dose. The estimate's error shrinks as the square root of the number of samples, which is why high accuracy is expensive.
Where the cost lives
- Deep-penetration shielding, where few particles reach the detector
- Rare interactions requiring many histories for a stable estimate
- High-dimensional design or uncertainty spaces
Variance reduction
Classic variance-reduction techniques (importance sampling, weight windows) guide computational effort toward the samples that matter, cutting the number needed for a given accuracy. Machine learning is being used to learn these importance maps automatically, improving on hand-tuned schemes.
ML acceleration
Beyond guiding sampling, surrogates trained on Monte Carlo output can replace repeated full runs across a design space, and learned models can propagate uncertainty faster than brute-force sampling. In each case the Monte Carlo result remains the ground truth against which the accelerated result is validated.
Honest accounting
A Monte Carlo estimate always carries a statistical error, which must be reported alongside the value; an under-converged run can look precise and be wrong. When ML accelerates or replaces sampling, its approximation error adds to, and must be bounded against, the statistical error. Used carefully, learning stretches a compute budget; used loosely, it hides error behind a fast, confident number.