Reducing Variance in Nuclear Monte Carlo Simulations Through Deep Learning Post-Processing Techniques in the Gaming Domain

Abstract In nuclear science and engineering, Monte Carlo simulation, a critical tool for solving complex physical problems, faces significant computational costs and challenges in achieving high precision due to the extensive time and hardware resources required. Drawing inspiration from the gaming industry, which has developed advanced deep learning techniques for denoising and super-sampling in raster graphics rendering, this study investigates the adaptation and integration of these techniques into Monte Carlo simulations to enhance precision and reduce computational demands. By incorporating deep learning super-sampling and denoising techniques, such as Intel OIDN, NVIDIA OptiX, SRCNN, and waifu2x, as post-processing steps in simulations of nuclear events, the study demonstrates a notable reduction in variance and improvement in accuracy. This novel approach not only presents a promising avenue for improving the efficiency of Monte Carlo simulations in nuclear science and engineering but also opens up possibilities for applying these advanced techniques to a broader range of complex issues in the field.