Utilization of artificial neural networks to resolve chemical kinetics in turbulent fine structures of an advanced CFD combustion model

Date
2017-03
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: This work investigates an alternative chemistry integration approach to be used with the eddy dissipation concept (EDC) advanced combustion model for large-scale industrial applications where detailed or reduced mechanisms are utilised. The goal of the study was to reduce the computational resources required to solve reduced or detailed mechanisms using the EDC model. The unique approach uses artificial neural networks (ANNs) as a chemistry integrator for the reactions that take place in the fine structure regions created by the turbulence field. The ANNs are therefore utilised to predict the incremental species changes that occur in these fine structure regions as a function of the initial species composition, temperature and the residence time of the mixture in the fine structure regions. The ANN’s weights- and bias matrices were changed to minimise the network’s prediction error using the back-propagation algorithm and datasets generated using the results of separate ideal plug-flow reactor simulations. The effect that the ANN’s architecture has on its ability to accurately predict the temporal evolution of the species was also investigated and the best-performing configuration was selected. The novel chemistry integration approach for the EDC model was implemented to model a piloted methane/air turbulent jet diffusion flame (Sandia Flame D) at a Reynolds number of 22400. To prove the concept, a 5-step methane combustion mechanism was used to model the chemical reactions of the experimental flame. The results of the new approach were benchmarked against experimental data and the simulation results using the standard integration approaches in Fluent. It was shown that once the ANN is well-trained, it can predict the species mass fractions with relative accuracy in both a time and computer memory efficient manner compared to using traditional integration procedures.
AFRIKAANSE OPSOMMING: Hierdie werk ondersoek ’n alternatiewe chemie-integrasie metode wat gebruik maak van die EDC (eddy dissipation concept) gevorderde verbrandingsmodel vir grootskaalse industriële toepassings waar gedetailleerde of verminderde chemiese meganismes gebruik word. Die doel van die studie is om die rekenaar hulpbronne wat benodig is om verminderde of gedetailleerde meganismes met die EDC-model op te los, te verlaag. Dié unieke benadering gebruik kunsmatige neurale netwerke (KNN) as ’n chemie-integreerder vir die reaksies wat plaasvind in die fyn struktuur streke wat deur die vloeiveld se turbulensie geskep is. Die KNN’e is dus aangewend om die inkrementele spesiesveranderinge wat in hierdie fyn struktuur streke plaasvind as ’n funksie van die aanvanklike spesiesamestelling, temperatuur en die reaksietyd van die mengsel in die fyn struktuur streke te voorspel. Die geweegde veranderlikes matrikse van die KNN is verander om die voorspellingsfout van die netwerk te minimeer deur die terug-voortplanting algoritme te gebruik deur middel van datastelle wat gegenereer is met behulp van die resultate van ideale propvloei-reaktor simulasies. Die effek van die KNN se argitektuur op sy vermoë om die tydelike evolusie van die spesies akkuraat te voorspel is ook ondersoek en die beste presterende opset is gekies. Die chemie-integrasie benadering vir die EDC-model is geïmplementeer om ’n geloodsde metaan / lug onstuimige straler diffusie vlam (Sandia Flame D) met ’n Reynoldsgetal van 22400 te modelleer. Ten einde die konsep te bewys is ’n 5-stap metaan-verbranding meganisme gebruik om die reaksies van die eksperimentele vlam te modelleer. Die resultate van die nuwe benadering is vergelyk met eksperimentele data en die simulasie-resultate wat met behulp van Fluent se standaard integrasie benaderings verkry is. Daar is bewys dat wanneer die KNN goed opgelei is, dit met redelike akkuraatheid die spesies massa-fraksies kan voorspel op ‘n wyse wat doeltreffend is beide ten opsigte van tyd en rekenaar geheue.
Description
Thesis (PhD)--Stellenbosch University, 2017.
Keywords
Neural network computers, Combustion -- Mathematical models, Computational fluid dynamics, Turbulence, Chemical kinetics, UCTD
Citation