No description available.

The combustion of petroleum-based fuels leads to the formation of several pollutants. Among them, soot particles are particularly harmful due to their severe consequences on human health. Over the past decades, strict regulations have been placed on automotive and aircraft engines to limit these particulate matter emissions. This work is primarily focused on understanding the fundamental behaviour of soot particles and their formation. Though the focus of this work is on soot formation and growth pathways, the study of the gas-phase combustion process was also an integral part to validate the mechanism. A reduced mechanism is developed with retaining the larger PAH species till coronene from KAUST-ARAMCO mechanism. Counterflow diffusion flames had emphasized the simulation of canonical configuration where the reduced mechanism is validated and the soot growth pathways are evaluated. The importance of the significant contribution of larger PAH species on the soot growth pathways in both SF and SFO flames is evident in this analysis. The sensitivity of these flames with respect to strain rates, dilution, and at higher pressures are analysed. Direct Numerical Simulation (DNS) of two-dimensional counterflow diffusion flames is conducted to understand the impact of vortex interactions on soot characteristics. The results indicate that the larger PAH species contributes to the soot formation in the air-side perturbation regimes, whereas the soot formation is dominated by the soot transport in fuel-side perturbation. The study is extended to simulate and compare coflow laminar flame using different statistical moment methods MOMIC, HMOM and CQMOM.

Novel combustion technologies ensuring low emissions, high efficiency and fuel flexibility are essential to meet the future challenges associated to air pollution, climate change and energy source shortage, as well as to cope with the increasingly stricter environmental regulation. Among them, Moderate or Intense Low oxygen Dilution (MILD) combustion has recently drawn increasing attention. MILD combustion is achieved through the recirculation of flue gases within the reaction region, with the effect of diluting the reactant streams. As a result, the reactivity of the system is reduced, a more uniform reaction zone is obtained, thus leading to decreased NOx and soot emissions. As a consequence of the dilution and enhanced mixing, the ratio between the mixing and chemical time scale is strongly reduced in MILD combustion, indicating the existence of very strong interactions between chemistry and fluid dynamics. In such a context, the use of combustion models that can accurately account for turbulent mixing and detailed chemical kinetics becomes mandatory.Combustion models for conventional flames usually rely on the assumption of time-scale separation (i.e. flamelets and related models), which constrain the thermochemical space accessible in the numerical simulation. Whilst the use of transported PDF methods appears still computationally prohibitive, especially for practical combustion systems, there are a number of closures showing promise for the inclusion of detailed kinetic mechanisms with affordable computational cost. They include the Partially Stirred Reactor (PaSR) approach and the Eddy Dissipation Concept (EDC) model.In order to assess these models under non-conventional MILD combustion conditions, several prototype burners were selected. They include the Adelaide and Delft jet-in-hot coflow (JHC) burners, and the Cabra lifted flames in vitiated coflow. Both Reynolds Averaged Navier Stokes (RANS) and Large Eddy Simulations (LES) were carried out on these burners under various operating conditions and with different fuels. The results indicate the need to explicitly account for both the mixing and chemical time scales in the combustion model formulation. The generalised models developed currently show excellent predictive capabilities when compared with the available, high-fidelity experimental data, especially in their LES formulations. The advanced approaches for the evaluation of the mixing and chemical time scale were compared to several conventional estimation methods, showing their superior performances and wider range of applications. Moreover, the PaSR approach was compared with the steady Flamelet Progress Variable (FPV) model on predicting the lifted Cabra flame, proving that the unsteady behaviours associated to flame extinction and re-ignition should be appropriately considered for such kind of flame.Because of the distributed reaction area, the reacting structures in MILD combustion can be potentially resolved on a Large Eddy Simulation (LES) grid. To investigate that, a comparative study benchmarking the LES predictions for the JHC burner obtained with the PaSR closure and two implicit combustion models was carried out, with the implicit models having filtered source terms coming directly from the Arrhenius expression. Theresults showed that the implicit models are very similar with the conventional PaSR model on predicting the flame properties, for what concerns the mean and root-mean-square of the temperature and species mass fraction fields.To alleviate the cost associated to the use of large kinetic mechanisms, chemistry reduction and tabulation methods to dynamically reduce their size were tested and benchmarked, allowing to allocate the computational resources only where needed. Finally, advanced post-processing tools based on the theory of Computational Singular Perturbation (CSP) were employed to improve the current understanding of flame-turbulence interactions under MILD conditions, confirming the important role of both autoignition and self propagation in these flames. / Doctorat en Sciences de l'ingénieur et technologie / info:eu-repo/semantics/nonPublished

An experimental study was performed to study soot formation and evolution in atmospheric, laminar, coflow, diffusion flames of Jet-A1, Synthetic Paraffinic Kerosene and their surrogates. Light extinction, rapid thermocouple insertion and thermophoretic sampling followed by transmission electron microscopy and atomic forced microscopy were used to obtain soot volume fraction profiles, temperature profiles and soot morphologies, respectively. Different soot evolution processes were observed on the flame centerline and on a streamline with a significantly different temperature history. Formation and agglomeration of the first soot particles are different on the two streamlines. Transparent liquid-like particles are produced in large volumes in the early regions of the flame centerline where T < 1500 K; these particles are undetectable by the extinction method with the wavelength of 632.8 nm. Most of the currently used computational soot models do not predict the liquid-like nature of nascent soot particles which has major effects on the modeling.

An experimental study was performed to study soot formation and evolution in atmospheric, laminar, coflow, diffusion flames of Jet-A1, Synthetic Paraffinic Kerosene and their surrogates. Light extinction, rapid thermocouple insertion and thermophoretic sampling followed by transmission electron microscopy and atomic forced microscopy were used to obtain soot volume fraction profiles, temperature profiles and soot morphologies, respectively. Different soot evolution processes were observed on the flame centerline and on a streamline with a significantly different temperature history. Formation and agglomeration of the first soot particles are different on the two streamlines. Transparent liquid-like particles are produced in large volumes in the early regions of the flame centerline where T < 1500 K; these particles are undetectable by the extinction method with the wavelength of 632.8 nm. Most of the currently used computational soot models do not predict the liquid-like nature of nascent soot particles which has major effects on the modeling.

Deep Neural Networks (DNNs) find widespread applications across various domains, including computer vision, recommendation systems, and natural language processing. Despite their versatility, training DNNs can be a time-consuming process, and accommodating large models and datasets on a single machine is often impractical. To tackle these challenges, distributed deep learning (DDL) training workloads have gained increasing significance. However, DDL training introduces synchronization requirements among nodes, and the mini-batch stochastic gradient descent algorithm heavily burdens network connections. This dissertation proposes, analyzes, and evaluates three solutions addressing the communication bottleneck in DDL learning workloads. The first solution, SwitchML, introduces an in-network aggregation (INA) primitive that accelerates DDL workloads. By aggregating model updates from multiple workers within the network, SwitchML reduces the volume of exchanged data. This approach, which incorporates switch processing, end-host protocols, and Deep Learning frameworks, enhances training speed by up to 5.5 times for real-world benchmark models. The second solution, OmniReduce, is an efficient streaming aggregation system designed for sparse collective communication. It optimizes performance for parallel computing applications, such as distributed training of large-scale recommendation systems and natural language processing models. OmniReduce achieves maximum effective bandwidth utilization by transmitting only nonzero data blocks and leveraging fine-grained parallelization and pipelining. Compared to state-of-the-art TCP/IP and RDMA network solutions, OmniReduce outperforms them by 3.5 to 16 times, delivering significantly better performance for network-bottlenecked DNNs, even at 100 Gbps. The third solution, CoInNetFlow, addresses congestion in shared data centers, where multiple DNN training jobs compete for bandwidth on the same node. The study explores the feasibility of coflow scheduling methods in hierarchical and multi-tenant in-network aggregation communication patterns. CoInNetFlow presents an innovative utilization of the Sincronia priority assignment algorithm. Through packet-level DDL job simulation, the research demonstrates that appropriate weighting functions, transport layer priority scheduling, and gradient compression on low-priority tensors can significantly improve the median Job Completion Time Inflation by over $70\%$. Collectively, this dissertation contributes to mitigating the network communication bottleneck in distributed deep learning. The proposed solutions can enhance the efficiency and speed of distributed deep learning systems, ultimately improving the performance of DNN training across various domains.

Autoignition characteristics of pre-vaporized iso-octane, primary reference fuels, gasolines, and dimethyl ether (DME) have been investigated experimentally in a coflow with elevated temperature of air. With the coflow air at relatively low initial temperatures below autoignition temperature Tauto, an external ignition source was required to stabilize the flame. Non-autoignited lifted flames had tribrachial edge structures and their liftoff heights correlated well with the jet velocity scaled by the stoichiometric laminar burning velocity, indicating the importance of the edge propagation speed on flame stabilization balanced with local flow velocity. At high initial temperatures over Tauto, the autoignited flames were stabilized without requiring an external ignition source. The autoignited lifted flames exhibited either tribrachial edge structures or Mild combustion behaviors depending on the level of fuel dilution. For the iso-octane and n-heptane fuels, two distinct transition behaviors were observed in the autoignition regime from a nozzle-attached flame to a lifted tribrachial-edge flame and then a sudden transition to lifted Mild combustion as the jet velocity increased at a certain fuel dilution level. The liftoff data of the autoignited flames with tribrachial edges were analyzed based on calculated ignition delay times for the pre-vaporized fuels. Analysis of the experimental data suggested that ignition delay time may be much less sensitive to initial temperature under atmospheric pressure conditions as compared with predictions. For the gasoline fuels for advanced combustion engines (FACEs), and primary reference fuels (PRFs), autoignited liftoff data were correlated with Research Octane Number and Cetane Number. For the DME fuel, planar laser-induced fluorescence (PLIF) of formaldehyde (CH2O) and CH* chemiluminescence were visualized qualitatively. In the autoignition regime for both tribrachial structure and mild combustion, formaldehyde were found mainly between the fuel nozzle and the lifted flame edge. On the other hand, they were formed just prior to the flame edge for the non-autoignited lifted flames. The effect of fuel pyrolysis and partial oxidation were found to be important in explaining autoignited liftoff heights, especially in the Mild combustion regime. Flame structures of autoignited flames were investigated numerically for syngas (CO/H2) and methane fuels. The simulations of syngas fuel accounting for the differential diffusion have been performed by adopting several kinetic mechanisms to test the models ability in predicting the flame behaviors observed previously. The results agreed well with the observed nozzle-attached flame characteristics in case of non-autoignited flames. For autoignited lifted flames in high temperature regime, a unique autoignition behavior can be predicted having HO2 and H2O2 radicals in a broad region between the nozzle and stabilized lifted flame edge. Autoignition characteristics of laminar nonpremixed methane jet flames in high- temperature coflow air were studied numerically. Several flame configurations were investigated by varying the initial temperature and fuel mole fraction. Characteristics of chemical kinetics structures for autoignited lifted flames were discussed based on the kinetic structures of homogeneous autoignition and flame propagation of premixed mixtures. Results showed that for autoignited lifted flame with tribrachial structure, a transition from autoignition to flame propagation modes occurs for reasonably stoichiometric mixtures. Characteristics of Mild combustion can be treated as an autoignited lean premixed lifted flame. Transition behavior from Mild combustion to a nozzle-attached flame was also investigated by increasing the fuel mole fraction.