Large eddy simulation of under-ventilated fires
The present thesis constitutes an important contribution to the understanding of a partially premixed combustion system associated with the hazardous backdraft phenomena. Backdraft may occur when fresh air is suddenly introduced into a vitiated environment where fire has already died out due to lack of oxygen but there are still unbumt fuel and products of incomplete combustion left. In the context of backdraft or deflagration, a complex flame structure is expected. Both, non-premixed and diffusion combustion might, in principle, be present. The present study focuses on the development of sub-grid scale (SaS) models to facilitate the study of such complex flame structure using the large eddy simulation (LES) technique. Before applying the model to the backdraft simulation, the individual sas models were firstly validated using simple configurations where detailed experimental data is available. A flamelet-like model for premixed combustion was introduced and thereafter coupled with the non-premixed combustion model through a "flame index" parameter. This concept makes use of the gradient signs of oxygen and fuel mass concentrations to distinguish between premixed and diffusion combustion regimes. In the present implementation in LES, an improved version of the flame index concept developed by Domingo et al.  was adopted. The model takes into account the fluctuations of both gradients at sub-grid level, which subsequently, might affect the filtered flame index value. In order to track the premixed flame front we used an approach which filters the progress variable balance equation using a filter larger than the actual LES grid. This approach has the advantage that it represents a physical meaningful variable and is stable from the numerical point of view because of the smooth gradients of the progress variable at the flame front and the species concentrations related to it. The flame front tracking technique was tested with an unstrained planar hydrogen flame front. Reasonably good results in burning speed and density ratio were obtained. The non-premixed or diffusion combustion regime was modelled using a flamelet-like model which considers the flame to be located at the stoichiometric value of the mixture fraction and it is related to the strain rate imposed by the counter flow of oxygen and fuel mass concentration feeding streams. This approach has the advantage that it has been tested for different scenarios and it is relatively fast as the variables can be pre-stored in a table. The flame index approach was tested using a laminar triple flame configuration. It was observed that the model could capture the different combustion regimes and predicted the lift-off height with reasonable accuracy. The location of the triple point was well predicted and the three branches further downstream could also be easily discerned from the predictions. Subsequently, a partially premixed turbulent lifted flame was simulated. In this case, it was ne,cessary to introduce the augment of burning velocity induced by the wrinkling of the flame front at sub-grid level. The sas flame front wrinkling factor is defined as the sub-grid scale flame surface divided by its projection in the resolved propagating direction. This can be regarded as the ratio of the sub-grid turbulent flame speed at grid scale (SrLl) and the laminar flame speed (s?). Reasonably good agreement was found on the lift-off prediction, the flame structure, and the mixture fraction profiles. A stabilization mechanism was discussed based on concepts previously exposed where the flame base faces a high velocity flow and a flammable mixture. Thereafter, the flame attempts to find its way upstream through low-speed flammable sections of the flow. It was found that in this process the stabilization point, herein identified as the maximum premixed heat release, plays an important role driving the flame base upstream the flow. Finally, two real scenarios of backdraft in a full scale fire test were simulated. These include the full scale backdraft experiment of Gojkovic  and the reduced scale experiment of Weng and Fan  . Unfortunately, there exist neither extensive nor accurate measurements før the former one and hence, the comparison against the numerical simulation was largely carried out on qualitative grounds. Five different stages were identified: 1) initial phase, 2) spherical propagation, 3) planar propagation, 4) flame front stretching and 5) fire ball. Qualitatively, the simulation agreed well with the experiment. The ignition delay time (the time from the opening of the hatch until the time when the ignition occurs) was well predicted by the simulation. It was also observed that the flame structure in the backdraft was predominantly premixed. More detailed measurements were available in the tests of Weng and Fan . These included the upper layer temperatures, mass concentrations and pressures at the openings. Different opening geometries were used and the total mass flow rates in and out of the container were also measured. Overall, the predictions were in good agreement with the measurements and the model predicted the correct trend for pressure and mass flow rates in the tests with different openings. Furthermore, the predicted occurrence and non-occurrence of backdraft in different geometrical configurations was in line with the experimental observations in which backdraft did not always happen. During the earlier stages of the study, some effort was devoted to improving the sas turbulence models and to implement a CMC type SGS combustion model into the code. Unfortunately both models were later found to be unsuitable for the backdraft simulation. The first one suffered numerical instabilities caused by the under prediction of the Smagorinsky constant when applied to the backdraft case. The second one was deemed inappropriate due to its requirement of an homogeneous plane of conditional values. Nevertheless, some reasonably good results have been obtained with both models during the validation using simple geometrical configurations, namely a buoyant plume, a backward facing step and the Sandia-D non-premixed turbulent flame. The effort in this direction is therefore still included in the thesis as summarised below. A Lagrangian SGS turbulence model was implemented. Good results were found for classical benchmarking flow configurations such as the buoyant turbulent plume and the backward facing step. It was, however, found that the SGS turbulence viscosity became negative in a larger percentage than originally stated by Maneveour. Because of this, the model is prone to numerical instability. When applied to the backdraft simulation, the dynamically calculated Smagorinsky constant using this model was found to be consistently lower than the conventional range (0.1-0.23). This caused stability problem and made it very difficult to achieve converged solutions. The conditional source estimation (CSE) approach, which is a variation of the Conditional Moment Closure (CMC) approach, was also implemented. This model produced good results for the classical turbulent diffusion flame (SANDIA flame D). Even though the present implementation is not capable of predicting extinction/re-ignition events it was showed that it is very economic from computational point .of view. However, as explained above, this model was also considered as unsuitable for the backdraft simulation.