Adaptive spatial convolution weight window smoothing for global variance reduction in deep-penetration Monte Carlo simulations

In deep-penetration Monte Carlo simulations, the efficiency of the weight window global variance reduction (WWGVR) method critically depends on the accuracy of the lower weight window bounds. Statistical fluctuations during the initial iterations of self-coupled Monte Carlo method often lead to biased weight window estimates, triggering long-history problems. Conventional practice relies on empirical parameters as mandatory constraints. However, these parameters exert limited influence on Figure-of-Merit (FOM) within reasonable ranges, indicating that the efficiency bottleneck lies not in parameter fine-tuning but in the rationality of the weight window distribution. To address this, an Adaptive Spatial Convolution Weight Window Smoothing (ASCWWS) method is proposed, which employs an error-adaptive neighbourhood convolution kernel to achieve low-pass filtering and regularization of the weight window field, transforming point estimation into field estimation to suppress high-frequency noise and restore physically consistent importance gradients. Validation on the HBR-2 and EURACOS benchmarks shows that ASCWWS improves the FOM by factors of 1.86 and 4.06 compared to the conventional MAGIC method, respectively, with the incidence of importance inversion reduced by over 31%. By enhancing weight window distribution quality, ASCWWS reduces reliance on empirical parameters and can be directly integrated into existing self-coupled Monte Carlo method frameworks as a post-processing tool, offering a new pathway for efficient WWGVR simulation of deep-penetration problems.