Ερευνητική συνεργασία με το Tμήμα Μηχανολόγων του Ιmperial College (διπλωματική)

Machine Learning for Computational Combustion Applications
The simulation of turbulent combustion is essential for the effective design of combustion equipment such as IC engines and gas turbines, to attain improved energy efficiency and reduction in emissions of pollutants. However, the coupling of Computational Fluid Dynamics (CFD) and combustion modelling is very computationally demanding, due to the need for coupling the non-linear equations of chemical kinetics with fluid dynamics and transport phenomena. The bottleneck of the whole approach is the computation of chemical kinetics, which requires the numerical integration of a stiff system of ordinary differential equations (ODEs). Reaction mechanisms for commercial fuels, such as kerosene, involve hundreds of chemical species and thousands of reactions, which must be integrated over a time period at every spatial point of a grid comprising millions of cells.

The research group at Imperial College London (Dr S. Rigopoulos) has been working on an approach for circumventing this bottleneck via machine learning. The essence of the approach is the training of a machine learning tool, such as an Artificial Neural Network (ANN), so as to function as a regression model to replace the computationally expensive numerical integration of ODEs. The approach consists of the following steps:
1. The generation of a database of results of ODE integrations for training the machine learning tool. This is accomplished through the simulation of an abstract set of model combustion problems that anticipate the parts of the composition space expected to be encountered in real combusiton problems. Setting up these problems requires considerable insight into combustion science.
2. The clustering of the resulting data set. Combustion kinetics are multi-dimensional dynamical systems exhibiting complex non-linear behaviour. As a result, it can hardly be expected that a single machine learning tool could obtain satisfactory regression across the entire composition space. We employ clustering algorithms such as the Self-Organising Map (SOM) to generate sub-sets so that a grid of regression models can be ultimately obtained. This is a crucial step, and any improvement in the clustering is likely to have significant impact on the whole approach; new algorithms are currently under investigation.
3. The training of the machine learning tool. We currently employ Multi-Layer Perceptrons (MLPs), a form of ANN, for the regression at each cluster. Training the MLP is a difficult non-linear optimisation problem, and we have so far employed the steepest descent and conjugate gradient methods, but many other possibilities exist. Devising an optimal strategy for utilising the big data sets generated so far is another important problem.
4. The testing and application of the machine learning tool. This involves developing methods for evaluating its performance and estimating errors, and the real-time application to turbulent flames. The turbulent flame simulation is carried out with Large-Eddy Simulation (LES) and the method of Stochastic Fields, which is a stochastic solution method for the transported probability density function (PDF) approach for modelling turbulence-chemistry interaction.
Projects are available in all of the above areas, as there are many open problems. So far, we have demonstrated proof of concept by simulating a series of turbulent flames with a data set developed through an abstract combustion problem. The objective of our current reseasrch is to further develop and expand the methodology in terms of range of applicability, accuracy and efficiency.

Για περισσότερες πληροφορίες καθηγητής Αντώνης Κοκόσης akokossis@mail.ntua.gr