Predicting Weight Windows for Duct-Streaming Shielding Geometries with a Convolutional Neural Network

Forward-flux-based weight windows give global variance reduction for duct-streaming shielding problems, but each new geometry needs an iterative Monte Carlo calculation to build them. We report preliminary results on a convolutional neural network (CNN) that learns the map from a two-dimensional duct geometry and source position to the lower window bound of such an iteration, here the MAGIC generator in OpenMC, in milliseconds. A set of 250 random concrete ducts of three curved and rectangular families on the Protvino labyrinth bounding box was run in OpenMC with a Cf-252 point source and MAGIC generation, and a 1.9-million-parameter U-Net was trained on the logarithm of the lower bound. On held-out ducts the error in the cells that carry the flux is 0.14 to 0.51 decades; replicate MAGIC runs show this part of the map to be reproducible to a few hundredths of a decade, so the residual is model error, while below $10^{-6}$ the reference itself is noise. Applied without retraining to the rectangular Protvino labyrinth, the prediction correlates with MAGIC at 0.96. Used as weight windows there, with its deepest cells clamped at deployment, it gives an exit-detector figure of merit 54 times the analog value, 1.7 times that of the MAGIC generation run for the same geometry, and 0.80 of that of reused MAGIC windows whose construction it does not pay for. The run cost of any window set is shown to be set by its deepest cells and predictable from a population proxy.