Machine Learned Compact Kinetic Models for Combustion

Abstract

Chemical kinetic models are an essential component in the development and optimisation of combustion devices through their coupling to multi-dimensional simulations such as computational fluid dynamics (CFD). Due to the significant level of detail contained within, detailed chemical kinetic models are computationally prohibitive for use in CFD. Therefore, minimally-sized kinetic models which retain good fidelity to the reality are needed, the production of which requires considerable human-time cost and expert knowledge. In this thesis the concept of Compact Models is introduced. This concept overcomes the kinetic model detail-fidelity coupling as it describes the reaction mechanism virtually rather than retaining a physical authenticity in the reaction mechanism description. This allows for the numerical optimisation of model parameters to accurately replicate a set of combustion calculations without the constraint of maintaining a physical authenticity in the reaction mechanism description. A novel automated compute intensification algorithm called Machine Learned Optimisation of Chemical Kinetics (MLOCK) is constructed to produce compact models, removing the constraint of expert knowledge. The first step in the production of a compact model (?compaction?) is the construction of a virtual reaction network. This is a network of nodes (virtual species) connected to each other by connections (virtual reactions). The strength of the connections (virtual reaction rate constants) are numerically optimised using the MLOCK algorithm. MLOCK simultaneously perturbs all three virtual Arrhenius reaction rate constant parameters of important connections and assesses the suitability of the new parameters through objective error functions, which quantify the error in the calculations of each model candidate to a set of optimisation targets. In this thesis, compact models are developed for three combustion systems in response to an Industry brief. The aim is to provide a model that is efficient and accurate to a set of calculations pertinent to the gas turbine design process. The MLOCK methodology is first applied to the archetypal methane combustion system. Virtual reaction networks are constructed through application of conventional model reduction algorithms, followed by optimisation of sensitive virtual Arrhenius parameters using MLOCK1.0. It is shown that the NUGMECH1.0 detailed model, comprised of 2,789 species, is reliably compacted to fifteen nodes, whilst retaining an overall fidelity of 79-90% to the detailed model at Industry-defined performance target calculations, outperforming the prior state-of-art models. Additionally, the 118-species NUIG1817C3 detailed model is compacted to produce a fifteen and nineteen-node compact model with an overall fidelity of 75.0 ? 83.9% and 93.6 ? 95.0% respectively. The MLOCK methodology is then applied to the task of NOX formation. A virtual reaction network comprised of three nodes (N, NO and NO2) is constructed using a combinatorial method, and is appended to the fifteen-node model for methane/air combustion. MLOCK1.1 is applied to find the optimal set of virtual Arrhenius parameters for these NOX connections. A state-of-art eighteen-node compact model is produced with a fidelity of 89.5% to NUIGMECH1.0 perfectly stirred reactor (PSR) calculations of NOX. Although, for flame calculations, it was found that it is not possible to find one model of over 57% fidelity across the entire industry-defined performance envelope using a three-node NOX network. Finally, the MLOCK methodology is applied to the task of liquid fuels. A step-by-step generalised methodology that allows for the creation of a compact model for a liquid fuel was developed and demonstrated for the case of Arabian Extra Light (AXL) crude oil, at the behest of Siemens Energy. This led to the creation of a thirty-five-node compact model with fidelities of 79.6 ? 89.1% to the 6,476 species detailed model at the Industry-defined set of performance conditions for high temperature combustion. The method was then extended to the task of low temperature chemistry. A virtual reaction network for low temperatures was created and the virtual bin scheme was simplified. MLOCK1.2 was used to find the optimal set of virtual Arrhenius parameters for sensitive connections. This led to the creation of a thirty-two-node compact model with a high Su fidelity at lean conditions and a high fidelity to detailed model 0-D calculations across the complete performance envelope.