Modeling of Advanced Combustion Concepts in Engine with FGM

Diesel engines have higher efficiency than spark ignition engines due to the high compression ratio and the lack of throttling losses. However, high NOx and soot emissions are the most important challenges. Low-temperature combustion methods have been proposed to address these issues. These combustion approaches are promising strategies to achieve low NOx, soot pollutants, and heat losses. Homogeneous Charge Compression Ignition (HCCI) engines fall into this category. However, these methods have failed to see broad commercialized implementation because of difficulties in controlling the start of combustion and the pressure rise rate. In recent years, Reactivity Controlled Compression Ignition (RCCI) has been proposed which uses in-cylinder fuel blending with two fuels of different reactivity and two injectors (one in the intake port with high octane fuel and the other in the cylinder with high cetane fuel) to control in-cylinder heat release rate and to optimize combustion phasing and duration. In RCCI engines, lower NOx and soot emissions and higher gross indicated efficiency compared to conventional diesel engines have led to considerable interest in their application. However, combustion simulation in RCCI engines is very complex because two different fuels are involved and many intermediate species are formed during the combustion process, so the chemical kinetics must be examined in detail, but the simulation of combustion with detailed chemical kinetics is too difficult since it includes solving the transport equation for each chemical species and solving the set of differential equations that models the detailed chemical kinetic process. These differential equations form a stiff system, which represents a wide spectrum of timescales. The integration of these equations requires specialized advanced solvers and is time-consuming. Since most numerical methods for the calculation of fluid flow are, by themselves, very demanding in computer time and require many integrations of the system of equations, a direct integration of the flow and the Ordinary Differential Equation (ODE) solvers is, in general, impractical. Also, since detailed chemical kinetics mechanisms usually include more than a thousand species, it is very time-consuming to simulate these engines with this method. The tabulated chemistry approach has been, for a long time, the only affordable way to include detailed chemistry effects in practical simulations of turbulent combustion systems. The Flamelet-Generated Manifold (FGM) is one example of a tabulation method. In this approach, a multi-dimensional turbulent flame can be treated as a set of so-called one-dimensional (1D) laminar flamelets. The equations governing these flamelets are solved, and the solutions are mapped to a new coordinate system to build a so-called low-dimensional manifold. This thesis studies the applicability of the FGM for simulating RCCI engines. One of the challenges regarding the FGM usage in the RCCI engine application is the proper definition of a progress variable (PV). The PV definition highly affects the model's performance and accuracy in predicting combustion characteristics. To overcome this limitation, a generalized approach has been proposed based on Genetic Algorithms (GA) optimization, which automatically finds PV definition by maximizing its monotonicity across the mixture fraction domain. The results show that this approach significantly improves the model performance compared to traditional FGM implementation, while removing the manual selection process for PV definition. The new approach has been validated against detailed kinetic and ECN Spray A experimental results. The validation shows significant improvement for the prediction of the most important combustion characteristics, such as ignition delay, flame lift-off length, and species distribution over multiple operating conditions. In addition, this thesis addresses another critical limitation of the FGM, which is excessive usage of memory, storage, and interpolation errors, especially when the dimension of the manifold is increased in some applications, such as RCCI engine simulation. To address this issue, different machine learning techniques, namely Artificial Neural Networks (ANNs), Random Forests (RF), and Gradient Boosted Trees (GBT), are combined with FGM. These ML models are trained to predict the important combustion parameters directly from control variables and incur much smaller memory and runtime costs than those of the traditional FGM approach. Among the ML methods investigated, shallow neural network architectures exhibit the best predictive performance, especially for complex parameters like the source term of the progress variable. Despite the fact that it needs more training time than the other ML models, ANNs continue to offer accurate predictions that are crucial for credible combustion modeling. Benchmark calculations show that the ML-FGM model performs as accurately as conventional detailed kinetic calculations. Furthermore, the ML approach is successfully generalised to predict other transport properties such as thermal conductivity, viscosity, and diffusivity, ensuring the compactness and efficiency of the FGM model while maintaining accuracy. In general, the generalized approach for PV definition and ML-FGM highly improves the applicability of the traditional FGM in simulating advanced combustion concepts such as RCCI engines. These methods provide promising ways for computationally efficient, accurate simulations of complex dual-fuel combustion processes, which are critical for the development of next-generation, high-efficiency, low-emission combustion systems.