Advancing nuclear reactor simulations using ray tracing, machine learning and kinematic models of energy deposition

Abstract

Neutronic simulations are essential in the design of nuclear reactors as well as in their operation. In addition to their use in justification of safety and operating capability to regulation authorities, neutronic simulations provide real-time data and critical indicators for operators to optimize fuel consumption and equipment lifetime. Among the many benefits in advancing the performance of neutronic simulation tools is the potential to substantially accelerate the design of the next generation of advanced reactors. Neutronic simulations rely on physics models and numerical methods which typically require significant amounts of nuclear data. Despite considerable improvements in computing hardware in past decades, neutronic codes still rely on tabulated data and simplified physics models for performing simulations. While using pre-tabulated nuclear data can reduce simulation run times, this can also place limits on accuracy and drastically increase the demand on memory. Although high throughput Graphics Processing Units (GPU) and GPU-accelerated High Performance Computing (HPC) platforms can reduce this dependence, optimizing neutronic codes and their algorithms is key in fully leveraging advanced computing capabilities and maximizing fidelity.

Three computational projects which aim at enhancing neutron transport codes' performance via modular approach are presented. These projects share the common objective of enabling neutronic codes to make optimal use of massively parallel acceleration offered by GPUs. First, an extension of escape probability calculation to complex geometries using ray tracing is proposed. Ray tracing algorithms benefit from years of refinement and optimization for graphics applications allowing them to run efficiently on GPUs. The second project aims at reducing the memory burden associated with nuclear data by performing data compression with auto-encoders. Data transfer through system memory can exceed operation time in parallel applications. Reducing the size of data in memory is a logical way to limit memory overhead. Lastly, a method to calculate more accurate energy deposition from Monte Carlo methods with less pre-tabulated data is proposed.