Improved modelling and experimental validation of atmospheric dispersion on the very local scale

In the event of the release of a radiological pollutant into the atmosphere, a fast and accurate estimation of the pollutant dispersion is vital for the initial emergency assessment. To this end, many dispersion models have been developed to evaluate the impact of the release on the environment. Unfortunately, the limited ability to account for vegetation, buildings and other large structures in most of these models, restricts their applicability at the near-range, i.e. within the first few hundred meters from the nuclear installation. In recent years however, Computational Fluid Dynamics (CFD) has proven to be a promising technique for atmospheric dispersion studies in the direct vicinity of the pollutant source, using either Reynold-Averaged Navier-Stokes (RANS) turbulence modeling or large eddy simulation (LES) turbulence modeling. Its main drawback however, is its computation speed which is by several orders of magnitude too slow for real-time emergency assessment. In this thesis it is investigated how CFD can be applied in the context of nuclear emergency preparedness and response. Focus is on both improving the accuracy of CFD as a base model, and on the formulation of fast reduced order models (ROMs) that retain the accuracy of CFD. We first focus on improving RANS modeling of atmospheric dispersion. The main advantage of performing RANS simulations is the lower computational cost compared to LES. A frequently recurring error corresponds to the observation of a large discrepancy in lateral plume spread between simulations and experimental data, combined with a significant overestimation of the concentrations in a vertical plane through the point of release and parallel with the wind direction. We argue that this is due to the fact that fluctuations in wind directions observed in experiments are only partly accounted for by the modeled turbulence. Therefore, a simple approach is introduced to estimate, based on experiments, the correct level of variability in wind direction that is required as additional boundary condition for the simulations. It is illustrated that including this unmodeled wind variability significantly improves predictions over traditional RANS models, and the Gaussian model. A second part of this work looks into LES. The added value of LES over RANS is its improved accuracy which results from including the motion of the large eddies in the simulation. Accordingly, also the dispersion due to turbulent eddies is better captured. We make use of LES to study the variability of radiological dose rate at ground level due to instantaneous turbulent mixing processes. For this, the CFD model is coupled with dose rate models for beta and gamma radiation. By performing a set of time-dependent simulations of a constant release into an open field, the dose rate at ground level is computed. A large variability of the dose rate is observed. We show that the dose rate from gamma radiation can be reduced effectively by performing time-averaging. However, it is illustrated that neglecting this variability can result in errors up to a factor of four on the dose estimation when long-term measurements are used to estimate the resulting dose from short-term exposures. However, even after averaging over a long time-period the variability in dose rate from beta radiation remains very high. These results indicate that the gamma dose measurements from nearby sensors cannot be used to accurately estimate the dose from beta radiation. Finally, the third part of the work aims at constructing fast ROMs that retain full CFD accuracy. To this end, we focus on deriving a ROM by projection of the CFD model onto a Krylov-subspace that is produced by the Arnoldi algorithm. The method results in a stable ROM, and the algorithm is formulated in such a way that it can be used with any choice of CFD solver. First, the model reduction is applied to the forward simulation of the pollutant dispersion at the Doel Nuclear Power Station. It is illustrated that after initialization, the ROM runs 25 times faster than real-time, including a possible dose assessment. Next, the model reduction methodology is used for the development of a method for the fast reconstruction of transient, multi-source emissions. The problem of the source reconstruction is formulated as a regularized least squares problem comparing the measurements to model predictions. By limiting the possible source locations to a finite number of possibilities, the Arnoldi algorithm can be applied to reduce the order of the problem. The size of the resulting system is reduced to such extent that the required time to solve the system is in the order of the simulated physical time span. Using a number of case studies, it is demonstrated that the method is both effective and robust for the source estimation of one or multiple possible sources from near-range measurements.