We discuss our framework for using the adjoint technique to perform efficient uncertainty quantification in nuclear reactor simulations. This approach requires the solution of two systems of equations: a forward, engineering system that models the reactor, and a related but different system that provides sensitivity and error estimates. For reasons we discuss, algorithms that solve these systems tend to be memory intensive, posing unique scaling challenges as we push toward larger problems and new architectures. We discuss our approach for reducing the memory footprint of our algorithm and argue via scaling experiments that this approach is the only viable option for performing uncertainty quantification at scale on realistic reactor-analysis calculations.
