Multi-scenario validation and parameter sensitivity of the GPU-QES dispersion model under local-scale wind tunnel experiments at a Nuclear Power plant site

Rapid and accurate prediction of atmospheric radionuclide dispersion during nuclear power plant (NPP) accidents is essential for effective emergency response planning and radiation exposure minimization, but remains challenging for complex local-scale scenarios. Accelerated by graphics processing unit (GPU), the Quick Environment Simulation (GPU-QES) is a three-dimensional Lagrangian particle dispersion model with a building-effect module tailored for high-resolution atmospheric dispersion modeling, making it suitable for radiological emergency response. However, its performance in local-scale scenarios involving both complex terrain and high-density buildings has not yet been systematically validated. In this study, GPU-QES was comprehensively evaluated against four local-scale wind tunnel experiments. The simulation results were compared with measurements of two-dimensional ground-level wind and concentration fields, as well as vertical profiles of wind speed and pollutant concentration. Sensitivity analyses were performed with respect to particle number, horizontal and vertical grid resolutions. The results demonstrate that GPU-QES consistently produces accurate three-dimensional wind fields across all four wind directions, reliably capturing both ground-level wind and vertical profiles. In addition, with GPU acceleration, it completes the computation of 90 million grid points within just 12 seconds under each wind condition, highlighting its remarkable efficiency. The model accurately reproduced the high concentration values observations in the building area, achieving a minimum absolute fractional bias of 0.04 and a normalized mean square error of 0.12, respectively. Vertical concentration profiles also agreed well with the measurements. For axial concentration distributions, over 82% and 90% of simulated values fall within a factor of 2 and 5 of measurements, respectively. An optimal parameter configuration was also identified to balance accuracy and computational efficiency, providing valuable guidance for modeling complex NPP scenarios.