Orientador: José Antônio Perrella Balestieri / Abstract: Direct injection spark ignition (DISI) engines aim at reducing specific fuel consumption and achieving the strict emission standards in state of the art internal combustion engines. Therefore, in this work the goal is to develop code for simulations of the internal flow in DISI engines, as well as the phenomenon of fuel spray injection into the combustion chamber using a Lagrangian-Eulerian approach for representing the multiphase flow, and Large-eddy Simulations (LES) for modeling the turbulence of the continuum medium by means of the open-source CFD library OpenFOAM. In order to validate the obtained results and the developed models, experimental data from the Darmstadt optical engine, and the non-reactive “Spray G” gasoline injection case, along with the reactive “Spray A” case from the Engine Combustion Network (ECN) will be employed. Finally, a novel open-source solver will be proposed to simulate the Darmstadt optical engine in motored and fired operation under stratified mixture condition, using data compiled by the Darmstadt Engine Workshop (DEW) for validation. Moreover, a deep learning framework is presented to train an artificial neural network (ANN) with the engine LES data generated in this work, in order to make predictions of the small scale turbulence behavior. / Resumo: Motores de ignição a centelha com injeção direta (direct injection spark ignition engines, DISI engines) visam reduzir o consumo específico de combustível e respeitar os restritos níveis de emissão em motores de combustão interna de última geração. Assim, pretende-se com este trabalho desenvolver código para simulação do escoamento interno em motores DISI, assim como os fenômenos de injeção de combustível no interior da câmara de combustão utilizando uma abordagem Lagrangeana-Euleriana para representação do escoamento multifásico e Simulação de Grandes Escalas (Large-eddy simulation, LES) para a modelagem da turbulência no meio contínuo, por intermédio da biblioteca CFD de código aberto OpenFOAM. De modo a validar os resultados e os modelos desenvolvidos, dados experimentais serão utilizados, obtidos do motor óptico de Darmstadt, e do caso de teste de injeção de gasolina não-reativo “Spray G”, juntamente com o caso reativo “Spray A” da Rede de Combustão em Motores (Engine Combustion Network, ECN). Enfim, um novo código aberto será proposto para simular o motor óptico de Darmstadt em condições de escoamento a frio (sem combustão) e com combustão em condição de mistura estratificada, usando dados compilados pelo Workshop do Motor de Darmstadt (Darmstadt Engine Workshop, DEW) para validação. Além disso, uma abordagem de aprendizado profundo (deep learning) será apresentada para treinar uma rede neural artificial (artificial neural network, ANN) com dados de simulação LES de moto... (Resumo completo, clicar acesso eletrônico abaixo) / Doutor

This thesis presents the development of a numerical water quality model using a general paradigm of reaction-based approaches. In a reaction-based approach, all conceptualized biogeochemical processes are transformed into a reaction network. Through the decomposition of species governing equations via Gauss-Jordan column reduction of the reaction network, (1) redundant fast reactions and irrelevant kinetic reactions are removed from the system, which alleviates the problem of unnecessary and erroneous formulation and parameterization of these reactions, and (2) fast reactions and slow reactions are decoupled, which enables robust numerical integrations. The system of species transport equations is transformed to reaction-extent transport equations, which is then approximated with two subsets: algebraic equations and kinetic-variables transport equations. As a result, the model alleviates the needs of using simple partitions for fast reactions. With the diagonalization strategy, it makes the inclusion of arbitrary number of fast and kinetic reactions relatively easy, and, more importantly, it enables the formulation and parameterization of kinetic reactions one by one. To demonstrate the general paradigm, QAUL2E was recasted in the mode of a reaction network. The model then was applied to the Loxahatchee estuary to study its response to a hypothetical biogeochemical loading from its surrounding drainage. Preliminary results indicated that the model can simulate four interacting biogeochemical processes: algae kinetics, nitrogen cycle, phosphorus cycle, and dissolved oxygen balance.